The Basic Principles of People Analytics Learn how to use HR data to drive better outcomes for your business and employees

1. 1Name the judgment you want to de-bias and write down the assumption behind it. Google's was that its managers hired world-class performers, which nobody had tested. State your version in one sentence, e.g. 'Our interviewers pick people who perform well.'
2. 2Collect the two data sets the test needs: the predictions or ratings decision-makers made at selection time, and the later actual performance of the people hired. Google's task force compared interviewers' predictions with new hires' real performance.
3. 3Run the comparison before changing anything. Check how accurately gut-based predictions track outcomes, and whether extra input (more interviews) improves accuracy. Google found that estimating performance after four interviews was about as good as after ten, so the extra rounds added time but not accuracy.
4. 4Cap the subjective input where the data shows no gain. Google forced a top-down decision of no more than four interviews per candidate, over the wish of everyone who wanted to meet the replacement for Jimmy.
5. 5Take question-writing away from individual interviewers. qDroid generates the questions: the interviewer enters the candidate's function and asks what the app prompts. Everyone gets a structured interview, and no interviewer's own preferences shape the questions.
6. 6Test the standardized questions against performance before relying on them. Google's qDroid questions had been extensively tested and shown to predict job performance accurately.
7. 7Capture answers and ratings in the tool on specific, pre-defined scales instead of free-form impressions, and convert them into a single number. Google's single number proved highly predictive of future performance.
8. 8Keep the loop running. Re-compare scores with later performance and treat the number as a prediction to validate. Gut feeling is replaced or supplemented, and the data decides how much weight it keeps.

Google's HR head tested the assumption that interviewers hired world-class performers. A task force compared interviewers' predictions with the actual performance of new hires, and the findings were surprising. Over ten years, managers and employees had spent more than a hundred thousand hours interviewing, yet the system was time-consuming without producing better hires. The data also showed that an estimate after four interviews was about as good as one after the tenth. Google then capped candidates at four interviews, a top-down decision. It replaced free-form interviewing with qDroid. The interviewer enters the candidate's function and asks the app's pre-tested questions. The interviewer records answers in the app and rates the candidate on very specific scales, and the app converts everything into a single number. That number proved highly predictive of future performance. Because interviewers no longer write their own questions, the interviews are far less biased.

• If no one has checked whether the current judgment predicts later performance, then test that first. Compare past predictions with actual outcomes before redesigning anything.
• If extra rounds of subjective input add no accuracy, then cap them top-down, even against resistance. Google did this at four interviews.
• If interviewers write their own questions, then replace them with a structured, pre-tested question set. Interviewer-authored questions are where personal bias enters.
• If you cannot show that a scored question predicts performance, then do not treat its score as objective. Validate it against outcomes first.
• If you want to keep gut feeling in the process, then use it as a supplement to the data. The definition allows 'instead of (or in addition to)' gut feeling, but the data score carries the decision.

• Assuming the existing process works. The tell: no one can produce a comparison of predictions against later performance, as at Google before the task force.
• Letting everyone keep interviewing 'just to be sure'. The tell: interview counts climb and hours pile up (Google: over 100,000 hours) while hire quality does not improve.
• Structuring the interview but leaving question choice to each interviewer. The tell: different candidates for the same role are asked different questions, and ratings rest on impressions instead of specific scales.
• Trusting the final number without validation. The tell: the score has never been compared with later performance, so it is not shown to be predictive.
• Doing the analysis but presenting it poorly. The tell: managers ignore the output. The passages note that how data is presented can affect what people do with it more than the data itself.

• ▸Decision under review: name the people decision (e.g., candidate selection for one role family).
• ▸Assumption being tested: one sentence stating what the organization currently believes about decision quality.
• ▸Subjective inputs in use: list the interviews, impressions, and self-written questions that feed the decision.
• ▸Prediction record: where past interviewer predictions or ratings are stored, and for how many hires.
• ▸Outcome measure: the actual performance data for those hires that you will compare against.
• ▸Accuracy finding: how closely predictions tracked outcomes, and whether more rounds improved accuracy.
• ▸Cap or change decided: the limit on subjective input (e.g., max number of interviews) and who decides it.
• ▸Standardization step: who writes the questions, which scales are used, and how answers become a single score.
• ▸Validation date: when the score will next be compared against later performance.

1. 1Name the assumption you are about to act on and write it as a testable claim, e.g. 'managers can predict new-hire performance from interviews.' Google's HR head did this: the belief that managers hired world-class performers had never been tested.
2. 2Turn the claim into a question the data can answer. The passages model five: Will leadership training make better managers? Does sales training change sales performance? Is the appraisal system effective? Do people policies have the intended effect? Are we hiring the right people?
3. 3Pair a prediction with an outcome. Collect what decision-makers predicted (interviewer ratings) and the actual result (new-hire performance), then compare them. This is how Google's task force measured interviewer accuracy.
4. 4Check the context before you trust the result. Ask whether environmental factors, such as the national labor market in an attrition analysis, explain the change. Ask whether your own expectations shaped what you looked for.
5. 5Re-analyze for confirmation bias. Ask: did I find what I was looking for, or an answer to a different question? Did I weigh all relevant factors? If not, rerun the analytics cycle and revise.
6. 6Interpret against the original question. If new questions surface, loop through the cycle again. Take a second look at any interesting finding before you act on it.
7. 7Convert the finding into a policy change and tie it to a business metric: time, cost, ROI, or performance. Google capped interviews at four after the data showed no accuracy gain beyond the fourth.
8. 8Replace the biased element with structure where the data points to it. Google used qDroid, which supplies tested questions, specific rating scales and a single predictive number. Then measure again.

Google's managers spent roughly five to ten hours interviewing each new hire, and some were in hiring almost full-time. Over ten years the company spent over a hundred thousand hours interviewing candidates. Nobody had tested whether this worked. A small task force of data scientists compared interviewers' predictions of future performance with the actual performance of the people hired. The interviews did not reliably predict performance; the only thing they accurately measured was whether the interviewer liked the candidate. Interviewers formed an impression at the handshake and spent the hour looking for confirming cues. Estimates after the fourth interview were as good as after the tenth. Google capped interviews at four through a top-down decision. It then moved to structured interviews run through qDroid, with tested questions, specific rating scales and one predictive score, which reduced bias and wasted time.

• If a policy rests on a manager's confidence ('I can tell a good hire'), then test it against outcomes before scaling it. Conviction is not evidence, and managers rarely believe their hunches are wrong.
• If more of an activity (more interviews) is claimed to improve accuracy, then check whether accuracy plateaus. Cut the activity at the plateau (Google: four interviews).
• If a result matches what you expected, then re-analyze it harder. Expectations steer which evidence you include and which you ignore.
• If a finding surprises you or raises new questions, then run the analytics cycle again before acting.
• If the data shows human judgment is biased, then replace it with structure (standard questions, specific scales) rather than asking people to try harder. An algorithm predicting even 30% of future performance beats human judgment.

• Measuring efficiency only: you report cost or hours saved and never ask whether the policy works. The tell is a dashboard of cost per hire with no outcome metric.
• Confirmation-driven analysis: you collect and keep only data that supports the favored hypothesis. The tell is that the analysis never contradicts the sponsor's expectation.
• Ignoring context: you explain attrition changes without the national labor market. The tell is that year-to-year swings stay unexplained.
• Stopping at the first interesting finding: you act without a second look or re-run. The tell is a conclusion that answers a different question from the one you started with.
• Untested tradition: a long-standing practice (e.g., lengthy interview loops) continues because it feels thorough and everyone wants a say. The tell is that no one can show data that the practice improves outcomes.

• ▸Current assumption or practice: state the belief as a testable claim.
• ▸Business outcome it should affect: name the metric (performance, attrition, cost, time, ROI).
• ▸Question the data must answer: one sentence, phrased so the answer could be 'no'.
• ▸Prediction vs. actual data: list the decision-maker predictions and the outcome measures you will compare.
• ▸Context factors checked: labor market, timing, population changes, and anything else that could explain the result.
• ▸My expectation and bias check: write what you expected and what disconfirming evidence you looked for.
• ▸Finding and second-look notes: what the data shows, and any new questions that require another cycle.
• ▸Decision and policy change: the action, including any structure that replaces judgment.
• ▸Follow-up measure and date: how and when you will re-test the change.

1. 1Audit the result before you interpret it. Ask whether the finding is causal or correlational. Most are correlations, so list the variables you did not include that could explain the pattern.
2. 2Adjust for context. Ask whether superior performance comes from the individual or from the team around them. Re-cut the data by team, unit or manager before you credit or blame individuals.
3. 3Check your own expectations. Ask whether you found what you were looking for or an answer to a different question. If you only looked for confirming evidence, go through the analytics cycle again and revise the analysis.
4. 4Compare the result with the original question. Write down whether it answers the question the business asked. Log each new question that appears and run another loop of the cycle to answer it before you finalize the interpretation.
5. 5Take a second look at every interesting finding. Test whether a structural feature of the organization (hierarchy, seniority, role mix) explains the pattern before you report it as a trait of a group.
6. 6Turn the finding into an action. Prefer variables you can change (engagement) over fixed ones (sex, age). Keep fixed variables as controls, and ask WHY the pattern exists and what can influence those reasons. Draft the intervention before you present.
7. 7Choose the audience and channel before you design the visual. A monthly email report works better than a dashboard managers forget. A short, well-timed message in a tool like Slack can beat a whole dashboard. Show no complicated statistical models. Keep the message 'stupidly simple'.
8. 8Decide who sees what. For individual-level predictions such as attrition probability, restrict access so a manager cannot confront an employee with 'there's an 80% chance you will leave'.

The author studied innovative behavior at a professional service firm with a strongly hierarchical structure, in which many employees were managed by a small group of partners. The first analysis showed gender as an important predictor: men reported more innovative behavior and joined more innovation projects than women. That matched the previous literature, so it was tempting to report as is. A second look at the data showed that most of the firm's partners were men, which pointed to hierarchy as a competing explanation that the first model did not account for. The lesson is to re-cut any interesting finding by the organization's structure before communicating it. Reporting 'gender drives innovation' would have given the business a fixed attribute it cannot act on. Reporting the role of position or hierarchy points to levers such as how innovation projects are staffed and who gets access to them.

• If a finding is a correlation, then name the missing variables that could explain it and do not state causation in the message.
• If the variable is fixed (sex, age), then use it as a control and shift the question to the underlying WHY and to what can be influenced.
• If the result confirms what you or the manager expected, then re-analyze it for confirmation bias before you present.
• If the result contradicts a manager's gut feeling, then expect resistance and rely on a simple visual and a concrete intervention rather than on more statistics.
• If the output is a person-level prediction (such as attrition probability), then limit distribution and do not give it to line managers to use in conversations with the employee.

• Presenting findings without an intervention plan. The tell is that stakeholders say 'interesting' and nothing changes.
• Reporting a statistically sound result on an unchangeable attribute (sex, age) as if it were a lever. The tell is that the recommended action cannot be named.
• Treating the first hit as the answer, for example crediting gender when hierarchy was the driver. The tell is that no second cut by structure or context was run.
• Defaulting to a dashboard for managers. The tell is low usage, because managers forget it. A reminder-style email or a message in their existing tool works better.
• Showing complicated models to time-pressed executives, so the message gets lost. The tell is that the discussion turns to method instead of decisions.

• ▸Original business question: write the question the stakeholder asked, in their words.
• ▸Key finding: state it in one plain sentence with no model jargon.
• ▸Causal or correlational: mark which, and list the omitted variables that could explain it.
• ▸Context adjustment: record the team, unit, hierarchy or role cuts you checked.
• ▸Expectation check: state what you expected and whether the data answered a different question.
• ▸New questions raised: list them and mark which need another loop of the cycle.
• ▸Actionable lever: name the changeable factor and the intervention that targets its root cause.
• ▸Audience and channel: name who receives it and by what medium (meeting, email, dashboard, chat app).
• ▸Access restrictions: state who must not see individual-level results and why.

1. 1Name the people decision before you touch data. Pick one HR process the passages list (recruitment and selection, compensation, learning and development, firing/retention) and write the decision as a question, such as 'Which employees should I hire?' or 'Which of my employees are most likely to leave, and why?'
2. 2Write down the assumption currently driving that decision. Google assumed its managers hired world-class performers and had never tested it. State your equivalent in one sentence so you can validate or reject it.
3. 3Tie the question to a business outcome or priority. The source says good analytics should focus on and help solve business priorities (right people in the right place at the right time, product quality, sales performance, fewer workplace accidents). If you cannot name the outcome, drop the question.
4. 4Pair the decision-maker's prediction with the actual result. Google's task force compared interviewers' predictions of a candidate's future performance with new hires' actual performance to measure how accurate the predictions were. Build the same prediction-versus-outcome comparison for your decision.
5. 5Check the external and market data, not only internal data. Alexander compared workforce plans (number and skills needed, plus three-year growth projections) with external talent supply and demand data for the provisional China site, and found too few skilled people and heavy competition for them.
6. 6Remove interviewer discretion where bias enters. Google's qDroid approach has interviewers ask app-prompted, pre-tested questions, store answers in the app, rate on very specific scales, and have everything converted into one predictive number. Structure reduces bias because interviewers do not write their own questions.
7. 7Translate the result into numbers and a dollar amount for the audience. The source says analytics lets HR convert a people problem into a numeric rating and dollar figure and calculate the ROI of people policies. Present the finding in the language of the business.
8. 8Take the decision to the board or manager with an alternative. Alexander did not stop at 'this site is a mistake'; he worked with analytics counterparts to produce a new, better set of recommendations for the board to act on.

Alexander, Head of People Analytics at a large technology company, worked late after the company had provisionally settled on a multi-million dollar location in China for a key new operation. He set the workforce plans, which gave the number and skills of employees needed at opening plus three-year growth projections, against external talent supply and demand data his team had collected for that city. The comparison showed too few skilled potential employees, and the few available were in high demand from established companies there. He saw that the company was about to make an expensive mistake that could cost millions and block the projected growth. He then examined the same supply and demand data for the other Chinese cities under consideration and found more favourable locations. Working with analytics counterparts across the organisation, he produced a new set of recommendations for the board to act on. The decision changed because a data comparison replaced the provisional assumption.

• If a hiring or selection step consumes large amounts of time and rests on untested judgment, then compare predictions with actual performance before changing anything, as Google did with interviews.
• If interviewers write their own questions, then move to structured, pre-tested questions with specific rating scales, because the source ties fixed questions to less biased interviews.
• If a decision commits major money (a site location, a policy rollout), then check external supply and demand data alongside internal plans before the board acts.
• If you cannot connect a people question to a business priority or outcome, then do not analyze it; the source says analytics should solve business priorities.
• If you present a finding to leadership, then express it as a numeric rating, dollar amount or ROI, and bring a better alternative, not only the problem.

• Trusting gut feeling and untested assumptions: the tell is that no one can say how accurate past hiring predictions were, which was Google's position before it checked its interviews.
• Using only internal plans: the tell is a workforce plan stating the people needed with no external talent supply and demand data behind it, which almost led to the China site mistake.
• Leaving interviews unstructured: the tell is interviewers inventing their own questions, so evaluations vary and bias creeps in.
• Reporting static numbers that never reach a decision: the tell is a dashboard with no recommendation or action, when the foreword says to move beyond static reporting toward improved business outcomes.
• Starting from theory or tools instead of the business question: the tell is analysis with no named decision or business priority, a trap the foreword says many HR professionals fall into because the literature is too theoretical.

• ▸HR process and decision: name the process (hiring, compensation, development, retention) and phrase the decision as a question.
• ▸Current assumption: state in one sentence what is currently believed or relied on, such as gut feeling.
• ▸Business priority served: name the business outcome this decision affects.
• ▸Prediction versus outcome data: list what predictions or plans exist and which actual results you will compare them to.
• ▸External data check: list the external benchmarks or market data (e.g., talent supply and demand) needed.
• ▸Bias controls: describe how discretion is reduced (structured questions, specific rating scales, a single combined score).
• ▸Finding in numbers: state the result as a numeric rating, dollar amount or ROI.
• ▸Recommended action and alternative: say what the decision-maker should do and the better option if the current plan fails.
• ▸Decision owner and date: name who decides and by when.

1. 1Start from a business decision, not a dataset. Write the decision in one sentence (e.g., where to locate a new site, how to select candidates, whether a program pays back) and name the executive who owns it. Analytics that answer no owned decision become the 'fad' the source warns about: analyses that add no value to the business.
2. 2Size the money at stake before you analyze. Personnel expenses run around 70% of the budget in a typical company, and higher in service firms and firms with many highly educated employees. Pull the people-spend line connected to your question so you can say what share of spend the decision touches.
3. 3Pick the outcome metric the business already values and express it in dollars. The source argues analytics lets HR convert people policies into a dollar amount and compute ROI. Define the unit (cost avoided, revenue protected, time saved, performance gained) and the comparison baseline before you run anything.
4. 4Assemble the three skillsets the work needs: business/marketing (to frame the question and communicate), data (to obtain and aggregate the right data), and analytics (to select data, interpret outcomes, build valid models). If one is missing, borrow it from a department that has it in abundance, or hire for it.
5. 5Test the current practice against actual outcomes. Compare what the practice predicted with what happened, as Google's task force compared interviewers' predictions of candidate performance with later performance. Ask whether the practice measures what you assume it measures.
6. 6Bring in external data when the decision involves a market. In the foreword, Alexander compared internal workforce plans (number and skills needed, plus three-year growth projections) with external talent supply and demand data for the candidate city. Do the same: put your demand figures beside the external supply for every option under consideration.
7. 7Translate the finding into an action the owner can take, in the format the owner uses: a meeting presentation, a dashboard, or an email each need a different visualization. State the recommendation, the dollar effect, and the alternative you rejected. Expect numbers that contradict gut feeling and plan for that conversation.
8. 8Record the result as a competitive asset. Note what organization-specific data and capability produced the insight (for example, your own interview data scored against performance), because the advantage lies in what rivals cannot copy from a generic report.
9. 9Report outcomes in the business's terms afterward: the cost avoided or the performance difference achieved. This is how HR moves from business assistant to business partner, which the source says happens only when HR quantifies its own impact and uses data to influence business decisions.

The foreword describes Alexander, Head of People Analytics at one of the world's largest technology companies. The company had provisionally chosen a location in China for a multi-million dollar new operation. Alexander held workforce plans stating the number and skills of employees needed at opening, plus projections of additional hires over the first three years. He had his team collect external talent supply and demand data for the provisional city and compared the two. The city lacked enough skilled people, and the few who existed were in demand from companies already established there. He saw that the decision was a very expensive mistake, one that could cost millions and block the projected growth. He then ran the same supply and demand data for the other cities under consideration, found more favorable locations on talent, and worked with analytics counterparts across the organization to build a new set of recommendations for the board. The business impact is a multi-million dollar mistake avoided, and the method is reusable: internal demand plan against external supply, applied to every option.

• If the question cannot be tied to a decision a named executive will make, then do not start the analysis; reframe the question with the business first.
• If the people spend the decision touches is a small part of the roughly 70% of budget that goes to personnel, then rank the work below questions that touch larger spend.
• If you cannot state the result in dollars or a comparable business measure, then you cannot claim an ROI; define the dollar conversion before analyzing, not after.
• If a long-standing practice consumes large amounts of manager time (Google managers spent half their week interviewing), then test it against outcomes before refining it.
• If the decision depends on a labor market, then insist on external supply and demand data next to internal plans before recommending.
• If the finding contradicts managers' gut feeling, then present it in the owner's format with the dollar effect, because poor communication is a main reason analytics fail to be adopted.

• Running analyses that add no business value because they chase a trend. Tell: nobody can name the decision the output will change.
• Having the data skills but no business and communication skill, so managers do not act on the data. Tell: the report is accurate and then ignored.
• Having only simple-analysis skills, so the team never finds cause and effect or builds predictive models. Tell: outputs are descriptive counts and never answer why.
• Claiming benefits without a dollar translation, so HR stays a business assistant. Tell: results are in HR terms (satisfaction, activity) rather than cost, revenue or performance.
• Trusting an established practice because of the effort invested in it. Tell: nobody has compared what the practice predicts against later outcomes; at Google, interviews turned out mainly to measure whether the interviewer liked the candidate.

• ▸Decision and owner: one sentence naming the decision and the executive who makes it.
• ▸Spend at stake: the people-cost line the decision touches, and its share of total personnel spend.
• ▸Outcome metric: the one business measure that will move (cost avoided, revenue, performance, time).
• ▸Dollar conversion: how a unit change in the metric becomes dollars, and the baseline you compare against.
• ▸Internal data: the organization-specific data you will use and who holds it.
• ▸External data: market or benchmark data needed, if the decision depends on one.
• ▸Skills check: who covers business/communication, data, and analytics, and which gap you are filling and how.
• ▸Expected vs. actual: the prediction or assumption being tested and the outcome it is compared with.
• ▸Communication format: meeting, dashboard or email, and the single recommendation with its dollar effect.
• ▸Competitive asset: what rivals cannot copy about this insight, and where it will be reused.

1. 1Name the gap first. The passages say HR remains a support department with low impact on business decisions, mostly doing operational and tactical work. Write down which of your current deliverables are descriptive (sick days, scorecards, benchmarks between departments) and which have changed a business decision in the last year. The second list is usually short.
2. 2Pick one question the business already cares about, not one HR cares about. Use the passage's examples as templates: Does the sales training affect sales performance? Do managers get better after leadership training? Are we hiring the right people? Business relevance is what keeps the analysis from becoming a fad with no value.
3. 3Convert the problem into numbers and dollars. State the people problem as a numeric rating and a dollar amount, then calculate the ROI of the policy. Put the people cost in context: human capital is nearly 70% of operating expenses, so even small effects are large in money terms.
4. 4Check that you hold the three skillsets: business/marketing, data (extraction and aggregation), and analytics (statistics). If one is missing, borrow it from a department that has it or hire it. The failure signatures are managers who don't act, analyses that are simple but never reach cause and effect, or data you cannot obtain or combine.
5. 5Move past descriptive reporting. Combine data from multiple systems and run at least a correlation or regression. Test the model on fresh data, not the data that built it. Treat the result as a basis for action even when the mechanism is unknown, as in Jill's sick-day model.
6. 6Translate results for the audience. Decide whether the finding goes in a meeting, a dashboard, or an email, and build the visualization to fit that channel. Expect numbers that contradict a manager's gut feeling, and prepare to interpret them rather than just present them.
7. 7Put the finding in front of the decision-maker before the decision. Bring the recommendation, the dollar effect, and the proposed action to the line manager or board. Influence is measured by decisions changed, not reports delivered.
8. 8Specify the result for your own firm. Because context changes how people react, say explicitly that the finding applies to your organization, and use that as a competitive advantage rather than citing generic research.

The passage's Jill case shows the pattern at small scale. Jill has absence data and daily temperature. She fits a regression line: people calling in sick = 3.48 + .568 × temperature (°C). A 10-degree rise means about six more people calling in sick. She does not know why the relationship holds (beach days or poor sleep are both possible), but she can act: when next week's forecast is very hot, she calls in extra employees to cover the absences. That is a people question turned into a number, then into a staffing decision. The passage also flags the limit: the line was tested on the same data that built it, like a student marking their own paper, so she needs fresh data and more of it before the model is reliable. For strategic influence, the next step is to present the avoided coverage cost to operations leadership in dollars, not as a correlation coefficient.

• If your reports are only sick days, scorecards and benchmarks, then you are below the wall of Boudreau; invest in data aggregation and statistical skill before claiming a strategic role.
• If the analytics question matters only to HR, then drop it or reframe it around a business outcome such as sales performance or hiring quality, because HR-only questions produce analyses that add no value to the business.
• If the result contradicts a manager's gut feeling, then expect resistance and lead with the test design and the money, not the statistic. The Google example shows managers could not predict performance, yet none believed it.
• If a model has only been tested on its own training data, then do not present it as proven; get fresh data and present it as directional.
• If a finding comes from general literature, then check it against your own data before acting, because context differs between organizations.

• Gathering analytics skills only inside HR. The tell: the team gets stuck on simple analyses, or builds valid models the business never adopts. The cause is a missing business or data skillset.
• Reporting efficiency instead of effectiveness. The tell: your metrics show cost and effort reduction but cannot answer whether training, appraisal or hiring works.
• Treating a board-level title as the strategy. The tell: a CHRO or Chief People Officer role exists, yet the passage notes such roles have had very little effect so far, and decisions still bypass HR.
• Presenting analytical numbers without visualization or interpretation. The tell: the right results are sent by email or shown in a meeting in a format that does not fit the channel, and managers do not act.
• Assuming research findings transfer. The tell: you cite generic engagement or attrition drivers, but workload and pressure are perceived differently at a Big 4 firm than at a local municipality.

• ▸Business question: state the decision the business must make, in the business's language, not HR's.
• ▸Current HR output level: list your present reports and mark each descriptive, predictive or prescriptive.
• ▸Dollar frame: state the people cost involved (share of operating expenses, headcount cost) and the expected ROI of the policy.
• ▸Data sources to combine: list each system needed and the owner who can release the data.
• ▸Skills check: name who covers business, data and analytics skills, and where the gap is.
• ▸Method and validation: describe the analysis (e.g., regression) and the fresh data used to test it.
• ▸Communication format: choose meeting, dashboard or email, and describe the matching visualization.
• ▸Decision owner and date: name the person who decides and when you present before the decision.
• ▸Outcome recorded: note which decision changed and the measured result.

1. 1Start from a decision the business is about to make, not from a dataset. Write down the decision, who owns it, and what it costs if wrong. In the source example, the decision is where to site a multi-million dollar new operation in China, and the exposure is a very expensive mistake plus lost projected growth.
2. 2Phrase the question as an effectiveness or impact question, not a count. The source models this: 'Will our managers become better managers when they take leadership training?' and 'Does the sales training we offer have an effect on our people's...'. If the draft only asks 'how many' or 'how often', it is a report, not a business question.
3. 3Convert the problem into a number the CEO/CFO already uses: a rating, a dollar amount, or an ROI. People analytics lets HR turn a people problem into a numeric rating and a dollar amount, which is what makes HR a credible business partner. State the unit before touching data (hours, dollars, hires, performance).
4. 4Confirm with the business owner that the question matters to them. Talk to them before the analysis and again when you interpret results. Missing business-and-marketing skill is what produces analyses that add no value (a fad). Get the owner to say what they would do differently depending on the answer.
5. 5List the data the question requires, system by system, before committing. Google needed interviewer ratings, the Applicant Tracking System, and the Performance Review System. If the data needs aggregation across systems, check that you have the programming and IT knowledge to do it, or you will stall at the 'wall of Boudreau'.
6. 6Pick the analysis depth the decision needs. Descriptive outputs (sick days, benchmarks, scorecards) are easy but rarely settle a strategic question. If the decision is 'does X cause or predict Y', plan for predictive or causal work and confirm you have the skills for it.
7. 7Expect the result to contradict someone's gut feeling, and plan for it. Decide in advance how you will turn 'analytical numbers' into an actionable visual and which channel carries it: a meeting, a dashboard, or an email. Each channel needs a different visualization.
8. 8Close the loop with a decision. Bring a recommendation the owner can act on, as Alexander brought his board a new and better set of recommendations. Then track whether the action changed the number you named in step 3.

Google was hiring about 200 people a week at its 2007 peak. Every candidate was interviewed by the hiring manager and future colleagues, and some managers spent half their week interviewing. The business question was whether this very large time investment actually predicted performance. A small task force of data scientists compared interviewers' predictions of candidate performance with later performance. The inputs were interviewer ratings, the Applicant Tracking System, and the Performance Review System. They found the process was very time consuming without leading to better hires, after more than a hundred thousand hours of interviewing in ten years. Two decisions followed. Google imposed a top-down cap of no more than four interviews per candidate. It also moved to structured interviews directed by the qDroid app, with prompted questions, specific rating scales, and one summary number shown to predict future performance. The analysis started from a cost and quality question leadership cared about, not from available data.

• If you cannot name the decision owner and the decision the result will inform, then stop and do not start the analysis; you are exploring data, not answering a business question.
• If the question can be answered with a count, scorecard, or benchmark, then ask whether the owner needs cause or prediction. If so, reframe it as an effectiveness question such as 'does this training change performance?'
• If the question cannot be expressed in a number or dollar value the CEO/CFO uses, then rework it until it can; otherwise you cannot show ROI or earn business-partner standing.
• If the data sits in several systems and you lack aggregation or advanced-analysis skills, then narrow the question or add those skills before promising results; this is where HR hits the wall of Boudreau.
• If the likely answer will challenge managers' gut feeling, then agree the question and success measure with them before analysing, and prepare the visualization and channel in advance.

• Data-first exploration: the team picks an available dataset and looks for patterns. The tell: the project brief names data sources but no decision or owner.
• Analysis that adds no value to the business (the fad). The business and marketing skill is missing, so nobody tested relevance. The tell: the results are interesting but no executive acts on them.
• Staying in descriptive reporting (sick days, scorecards, benchmarks) because it is easy. The tell: every deliverable describes what happened, and none says what to change or what it is worth.
• Question agreed once, then silence. HR does not talk to the business at the start or while interpreting results. The tell: managers ignore the findings or dismiss them because the numbers contradict their gut feeling.
• Right analysis, wrong delivery. The numbers are never turned into actionable visuals suited to the channel. The tell: a dense output goes by email or sits in a dashboard when a decision meeting was needed, and no decision follows.

• ▸Decision at stake: State the specific business decision this analysis will inform (e.g., site selection, interview process design).
• ▸Decision owner: Name the executive or manager who owns the decision and has agreed the question matters.
• ▸Cost of being wrong: Describe the financial or growth exposure if the decision is made without data.
• ▸Analytics question: Write one effectiveness-style sentence (e.g., 'Does X affect Y?').
• ▸Business metric: Give the number or dollar/ROI unit the CEO/CFO will recognise.
• ▸Data sources required: List each system and who supplies the data (e.g., ATS, performance review system).
• ▸Analysis depth needed: Mark descriptive, predictive, or causal, and note any skill gap that could stall you.
• ▸Gut-feeling conflict: Note which beliefs the result might contradict and who holds them.
• ▸Delivery format and action: State whether results go in a meeting, dashboard, or email, and what action the owner will take for each possible outcome.

1. 1List the analytics questions your team will answer in the next 6–12 months. Judge each skillset against those questions, not against an abstract ideal team.
2. 2Score each of the five areas (business, marketing, HR, data analytics, IT) for named people: none, partial, or strong. Name the person who holds each skill. 'HR has it somewhere' does not count.
3. 3Define each area by the job it does. Business: pick questions that matter to business priorities. HR: know recruitment, hiring, firing and compensation well enough to choose the right data and interpret the outcomes. Data analytics: statistics and technique selection (correlation, regression, structural equation modeling). IT: pull and aggregate data from separate systems. Marketing: turn numbers into actionable visualizations and choose the channel (meeting, dashboard, email).
4. 4Test each area with a tell. Business: are your analyses tied to a strategic goal? Marketing: do managers act on what you present? HR: can you explain what the data means? Data analytics: can you run more than descriptive reports? IT: can you assemble the data without heroic manual effort?
5. 5Expect gaps, because the skills are rarely found together in HR. Close each gap by hiring someone with a different skillset or by borrowing from a department that has the skill in abundance. The source names Finance and IT as examples.
6. 6Assign a named owner to every borrowed skill, with a defined interface. Example: IT delivers aggregated data from the different systems, and the analyst states in advance which fields are needed.
7. 7Run the process in order: business question, data selection, aggregation, analysis, then communication. Check at each handoff that the responsible skillset is present before moving on.
8. 8Re-audit when your ambition changes. Moving from descriptive reporting to predictive work (the 'wall of Boudreau') demands stronger data-processing and statistical skills, so re-score then.

Take Google's question: do interviews predict future performance? The source notes that interviewing consumes so much employee time that small improvements make a big difference. That is the business area: the question is tied to a real cost. The HR area is needed to understand how interviews and ratings work and to read the result. IT is needed because the data sits in different systems. Google had to get interviewers to rate candidates and then collect data from its Applicant Tracking System. The data analytics area is needed to choose how to measure the link, and the source lists correlation, regression and structural equation modeling, noting that some techniques work considerably better than others. The marketing area comes last. Findings can contradict a manager's gut feeling, so they must be turned into actionable visuals suited to the setting (meeting, dashboard or email). The passages give no result for the Google analysis, so none is claimed. The point is that a single missing skillset breaks one step of this chain.

• If analyses are finished but managers do not act on them, then fix the marketing/communication skill first: change the visualization and the channel, not the statistics.
• If the team produces only simple analyses and cannot show cause and effect or build predictive models, then add data analytics capability before collecting more data.
• If you cannot obtain or aggregate the relevant data from several systems, then bring in IT (or a department with strong data skills) before starting the analysis.
• If a proposed analysis cannot be tied to a business priority or strategic goal, then drop it, because it risks being a fad that adds no value.
• If the team struggles to select the right data, interpret outcomes or build valid models, then strengthen HR knowledge of recruitment, hiring, firing and compensation.

• Staffing with HR generalists only. The tell: reports stay descriptive and no regression, root-cause or forecasting work happens.
• Building technically sound models with no business anchor. The tell: analyses get done, but nobody can name the business priority they serve.
• Treating the output as finished at the number. The tell: a correct analysis that managers ignore because it contradicts their gut feeling and was never visualized for their setting.
• Assuming the data will assemble itself. The tell: months spent extracting and merging files from separate systems, with no IT owner.
• Running analyses without HR context. The tell: the wrong variables are chosen or the results are misread, and the models are not valid.

• ▸Priority question: write the business question the team must answer next, and the strategic goal it supports.
• ▸Business skill owner / rating: name the person who ties questions to business priorities, rated none, partial or strong.
• ▸HR skill owner / rating: name who understands recruitment, hiring, firing and compensation well enough to select data and interpret results.
• ▸Data analytics owner / rating: name who can choose and run the technique (correlation, regression, SEM) and move beyond descriptive reporting.
• ▸IT owner / rating: name who can obtain and aggregate data from the different source systems.
• ▸Marketing/communication owner / rating: name who turns results into actionable visuals and picks the channel (meeting, dashboard, email).
• ▸Gap symptom observed: record the matching tell, such as no adoption, simple analyses only, no-value analyses, shaky data or models, or data not aggregated.
• ▸Closing route: for each gap, choose hire or borrow from another department (for example Finance or IT).
• ▸Review date: set when to re-score, especially before moving to predictive work.

1. 1Fix the source before you clean: adopt a dedicated HRIS so employee records are accurate and consistent. Level 1 organizations often argue about the data's validity instead of what it shows, and an HRIS ends that argument.
2. 2Run check 1 of the six-step checklist: confirm the data is up to date. Outdated data produces irrelevant results and can undo your work, so record the extract date of every source.
3. 3Run check 2: look for recurring unique identifiers. Some people hold multiple positions, and separate records may exist for each. Decide how to consolidate them before any merge.
4. 4Test validity with four questions: Does the data represent what we want to measure? Are there significant biases in how it was collected? Was it collected clearly and consistently? Are there outliers? Example: does the appraisal score measure individual performance, or who the manager likes best?
5. 5Check coverage for skew. Is data collected evenly across the organization? Find who is missing, because missing records narrow your population and the missing people often share traits. If one department uses an outdated performance system that omits questions, you lose that whole department and skew results toward the others.
6. 6Test reliability with two questions: Did the same thing measured several times give the same results? Were documented collection methods and instructions followed each time? Compare scores across raters and across time (morning vs. afternoon, different days of the week).
7. 7Standardize procedures in writing: the rating window, the scale anchors and the instructions. Vague scales cause rater bias, and so does one manager looking back six months while another looks back two weeks.
8. 8When combining systems, fix mislabeled function and section names first, then merge the datasets. Store the cleaned data separately or feed it back into the source system so quality improves for later analyses.

The city of Boston built a smartphone app that drivers installed to measure road bumps and report their GPS location, so the road service could fix them. A spokesperson said the data gave the city real-time information for fixing problems and planning long-term investments. It failed the validity questions. The app was used by younger people in more affluent communities, while poorer communities lacked equal access to smartphones and mobile data. The data therefore did not represent road quality citywide, which is a significant collection bias. Ask 'does the data represent what we want to measure?' and 'is the data evenly collected?' before acting, and the skew shows up early. The same logic applies to HR. If one department's performance system omits questions and its employees drop out of the dataset, your results skew toward the other departments.

• If the team debates whether the numbers are right instead of what they show, then move records into an HRIS and fix data quality before building any analysis.
• If records are missing, then check whether the missing people share a trait (same department, same system) before choosing a method. Some methods tolerate missing data and some algorithms struggle with it.
• If different raters score the same person differently, then treat the scale and instructions as the problem. Rewrite the scale, fix the time window and document the procedure before blaming the raters.
• If data comes from two systems, then reconcile labels (functions, section names) first, merge second, and write the cleaned result back or store it separately.
• If a metric measures something adjacent to your target (e.g., manager liking instead of performance), then do not use it as a proxy until you can show it measures what you intend.

• Treating cleaning as a one-off fix: cleaned data stays in an analyst's file and never returns to the source system. The tell is the same mislabeled functions reappearing in every new extract.
• Ignoring who is absent: you analyze only the records you have. The tell is a whole department or demographic missing, with results that look suspiciously favorable to the rest, as with Boston's affluent-skewed app data.
• Mistaking consistency for validity: raters agree, but the scale measures the wrong thing, such as being liked by the manager. The tell is that scores track manager relationships rather than output.
• Leaving rating procedures undocumented: managers use different time windows (six months vs. two weeks), moods and instructions. The tell is that the same person's scores swing between raters or between morning and afternoon.
• Duplicate records from people holding multiple positions inflate headcount or split one person's history. The tell is a repeated identifier across rows.

• ▸Data source and extract date: name the system and the date, then mark it up to date or outdated.
• ▸Unique identifiers: note whether any repeat (people with multiple positions or separate records) and how you will consolidate them.
• ▸Missing data: list the missing records or fields and whether the people affected share a department, system or trait.
• ▸Validity: does it measure what we intend? State the target construct and the actual measure, and flag any proxy such as a manager's liking.
• ▸Bias and coverage: is collection even across groups, locations and departments? Name who is under-represented.
• ▸Outliers: list any found and your decision to keep, correct or remove each.
• ▸Reliability: for repeat measures or multiple raters, compare scores across people and time and record the gap.
• ▸Procedure documentation: state the written scale, rating window and instructions, and confirm they were followed each time.
• ▸Label reconciliation and write-back: record mislabeled names fixed, the merge done, and where the cleaned data is stored or fed back.

1. 1Classify your current output before anything else. List every recurring HR deliverable (headcount, attrition, cost of labor, absenteeism) and mark each as reporting or as analysis. If the list is mostly metrics that are produced and then not acted on, you sit at level 1 or 2. The source says most companies were there in late 2016.
2. 2Separate reporting from advanced reporting. Level 1 is operational reporting: routine metrics, kept up to date at high manual cost. Level 2 is advanced reporting. Record how much effort the upkeep takes, because the source calls this a high hygiene factor: nobody notices current data, but everyone notices stale data.
3. 3Audit where the data lives. Name every system holding people data and mark whether it can be combined with the others. The source locates the wall of Boudreau here: data sits in separate systems, so combining it for analysis is hard and data-driven decisions stall.
4. 4Check whether the data is being consolidated. Look for cloud storage or business intelligence systems that already combine company data sources. The source treats this accumulation of aggregated data as a driver of company-wide analytics and as what makes extraction easier for HR analysts. Use it if it exists. If not, make it your first dependency.
5. 5Inventory the three skill sets: HRM knowledge, finance, and data analytics. Name who covers each. The source says puncturing the wall requires skill to measure data effectively and a more advanced statistical skill set to run analyses. If these are missing, hire for them or partner with departments that have them in abundance, such as Finance and IT.
6. 6Pick one people question tied to a business outcome and test it as an assumption. Use the type of question the source lists, such as which employees are most likely to leave and why, or the return on investment in learning and development. Choose an analysis method (correlation, regression, structural equation modeling) deliberately, since the source says some methods work considerably better than others.
7. 7Place your organization on the four levels from evidence, not belief. Level 4 means you actively predict outcomes such as performance. For context, 8% of organizations had done this in 2016 (IBM). Write down the one blocker (data, skills, or question) that keeps you at your current level, and make removing it the next-level goal.
8. 8Re-assess on a fixed cadence. The source shows capability moving quickly (organizations predicting performance grew 125% in a year), so a one-time self-rating goes stale.

Google was growing past 20,000 employees, doubling its population every year, and hired about 200 people a week at the 2007 peak. Managers spent up to half their week interviewing, yet nobody had tested whether interviews worked. Reporting would have counted hires and interview hours, which is level 1 to 2 output. Instead, a small task force of data scientists collected the interviewers' predictions of candidate performance and compared them with the new hires' actual performance. That moves into analytics: it tests an assumption and connects a people policy to an outcome. It also required combining data from different places, because interviewers had to rate candidates and performance data had to be gathered alongside. Methods such as correlation or regression could answer the question. The passages do not state the result, only that it was surprising. The maturity lesson is that the jump from reporting to analysis came from a specific question, joined data, and analytical skill, not from more dashboards.

• If your HR output is routine metrics that nobody acts on, then rate yourself level 1 or 2 and do not claim analytics, however polished the reporting is.
• If people data sits in separate systems that cannot be joined, then fix aggregation (shared storage or BI access) before attempting statistical modeling, because this is where the wall of Boudreau stops you.
• If the team cannot choose an analysis method or run it, then bring in statistical skill by hiring or by partnering with Finance and IT, rather than stretching traditional HR skills.
• If you are not predicting an outcome such as performance or attrition from data, then do not rate yourself level 4, which only a minority (8% in 2016) had reached.
• If an analysis cannot be tied to a business outcome, then drop it. The source warns that teams commit to analyses that add no value, a fad.

• Mistaking reporting upkeep for progress. The tell: the team spends most of its time refreshing headcount and attrition reports and gets no recognition for it, because hygiene work is only noticed when it fails.
• Reaching for prediction while data remains fragmented. The tell: analysts spend their time manually stitching systems together, and models rest on partial data.
• Staffing analytics with HR knowledge alone. The tell: the team struggles to select the right data, interpret outcomes, or build valid models, and finds correlations without cause and effect.
• Self-rating by optimism. The tell: you say you are advanced, yet nothing you produce tests an assumption or predicts anything. The source notes 92% of companies believe they are not optimally organized, so an unexamined self-score is unreliable either way.
• Running analyses disconnected from the business. The tell: results get presented, but managers do not act on the data or use it in decisions.

• ▸Recurring HR outputs: list each regular report or metric (headcount, attrition, cost of labor, absenteeism, etc.).
• ▸Action taken: for each output, record the decision or action it triggered, or write 'none'.
• ▸Maintenance effort: estimate the time spent keeping each output current (the hygiene cost).
• ▸Data systems and join status: list each system holding people data and whether it can be combined with the others (yes/no/partial).
• ▸Skills coverage: name who covers HRM knowledge, finance, and statistical/data analytics, and mark any gap.
• ▸Gap-closing plan: for each skill gap, record whether you will hire or partner (e.g. Finance, IT), and who owns it.
• ▸Outcome-linked question tested: state one people question tied to a business outcome that you have analyzed, and the method used (or 'none').
• ▸Predictive use: state whether you predict an outcome such as performance or attrition today (yes/no, with evidence).
• ▸Assigned level (1-4) and next blocker: record your level and the single constraint (data, skills, or question) you will remove first.

1. 1State the claim as a relationship at a named level: 'X is associated with Y for [individuals / teams / sites / the whole company].' Jill's claim is about daily absence counts across the workforce, not about any one employee's reason for calling in sick. Write the level down before you run anything.
2. 2List every system that produced each variable and who or what generated the number. Google pulled interviewer ratings, Applicant Tracking System data and Performance Review System data. Ask of each score whether it describes the person rated, the rater, or the process.
3. 3Name at least two rival explanations before looking at results. Include a hidden variable (what else changed at the same time?), a rater or observer effect, and a measurement effect. For Jill's data the rivals are beach trips and poor sleep in the heat. For Hawthorne the rival is being asked to participate.
4. 4Test whether the outcome moves when the supposed cause is held still. Mayo changed the lighting and productivity rose 25% anyway, which showed lighting was not the driver. If the outcome moves with no change in your named cause, a hidden variable is operating.
5. 5Check the pattern on repeated exposure. Google found accuracy after the fourth interview matched accuracy after the tenth. Flat returns signal that the measure is capturing something other than the candidate's future performance, such as the interviewer's first impression, which the interviewer spent the next hour trying to confirm.
6. 6Validate on fresh data, not the data that built the model. Jill's regression (people calling in sick = 3.48 + .568 × temperature in C) was checked against its own data, which the source likens to a student marking his own paper. Hold out new observations before generalising.
7. 7Separate 'act on it' from 'explain it'. A forecast is enough to act on: with a hot week ahead, Jill calls in extra cover. A mechanism claim (beach versus sleep) needs more research. Label each conclusion as one or the other.
8. 8Check scope before generalising. Alexander's China location recommendation changed once demand data for the other cities under consideration was added. Ask which units, places and periods your data leaves out, and do not extend the finding to them.

Elton Mayo's Hawthorne study set out to measure how lighting conditions affected workers, giving some workers more illumination than others. Post-intervention productivity was 25% higher than when the experiments began. It stayed higher no matter how the physical conditions were altered. A lighting-level reading would credit the illumination. Run the context steps and the picture changes. The outcome rose when the named cause varied in every direction, so lighting cannot be the driver. The hidden variable was the setting itself: by asking workers to participate, Mayo's team made them feel like an important group whose help and advice the company sought. The correct unit of analysis was the social conditions of the work group, not the physical environment of individual workstations. This finding gave rise to the human relations movement. The same discipline applies to Google's interviews. The ratings turned out to measure whether the interviewer liked the candidate, so the data described the rater, not the candidate. Reading a rating as 'this candidate is strong' would be a misattribution.

• If the outcome changes when your named cause does not, then look for a hidden variable, such as participation or attention (Hawthorne), before crediting the cause.
• If a score comes from a person judging another person, then ask whether it measures the judge's impression rather than the subject's trait, as Google found with interviewers who liked the candidate.
• If your model was tested only on the data that built it, then treat it as untested and hold out fresh data before relying on it.
• If you can forecast but cannot explain the mechanism, then act on the forecast (extra cover on hot days) and label the explanation as an open question needing further research.
• If your data covers one site, one small sample or a subset of the options, then limit the conclusion to that scope and widen the data before recommending broadly. The China site recommendation improved once the other cities' demand data was included.

• Reading a population-level pattern as an individual explanation. The tell: you say 'this employee called in sick because of the heat' when your data only shows daily workforce counts rising with temperature.
• Crediting the intervention when the attention is doing the work. The tell: results improve whatever you change, as at Hawthorne where productivity rose 25% regardless of lighting.
• Treating a rating as a trait of the person rated. The tell: scores track the rater's liking and first impression, and extra interviews add no accuracy (the fourth interview was as good as the tenth).
• Calling a regression 'causal' on the strength of a fit alone. The tell: the write-up asserts a causal relationship while admitting 'we do not know how this relationship precisely works' and tests the line on its own training data.
• Merging systems without checking what each measures. The tell: interviewer ratings, applicant-system records and performance reviews are joined into one table with no note on who generated each field or when.

• ▸Claim: write the relationship in one sentence (X is associated with Y).
• ▸Level of analysis: state whether the claim is about individuals, teams, sites or the whole company, and confirm the data matches that level.
• ▸Data sources and generators: for each variable, name the system and who or what produced the number (subject, rater, process).
• ▸Rival explanations: list at least two, including one hidden variable and one observer or rater effect.
• ▸Variation test: describe what happened to the outcome when the named cause changed, and when it did not.
• ▸Validation: state whether the result was tested on fresh data, and by whom.
• ▸Scope: list which sites, periods or options the data excludes.
• ▸Action versus explanation: split into 'what we can act on now' and 'what needs further research'.
• ▸Confidence and next step: record the claim as predictive or causal and name the next data or test needed.