The basic principle of people analytics learn how to use hr data to drive better outcomes for your business and employees

1. 1Filter every finding through actionability: separate variables you can change (engagement, policies, interventions) from control variables you cannot change (sex, age). Only build recommendations on the changeable ones.
2. 2Never present a finding without a concrete follow-up plan of interventions attached. When turnover is high, don't stop at the number — identify the factors driving it and specify interventions that solve the root cause.
3. 3Choose the delivery channel by audience behavior, not sophistication. Managers forget self-service dashboards for non-daily data; default to a monthly or quarterly PDF report emailed to them, or a short well-timed message in an existing app like Slack.
4. 4Simplify the message before you visualize: keep it stupidly simple. Strip out the statistical models — managers and executives are short on time and won't act on complicated models.
5. 5Design the visual layout deliberately. Decide where graphs go and question default visuals (e.g., don't use a speedometer for an engagement score if a clearer format exists), because presentation affects action more than the data itself.
6. 6Guard against harm when deciding who receives the information. In attrition analysis do not push individual flight-risk probabilities to a manager who might confront the employee directly (e.g., 'there's an 80% chance you'll leave').
7. 7Anchor the recommendation to one of the organization's top three strategic issues so it reaches the CEO/CFO; a finding disconnected from a key challenge becomes an irrelevant HR showpiece.

Credit Suisse (around 48,000 employees) faced high turnover with costs estimated in the tens of millions. Rather than reporting the raw turnover rate, the analytics work isolated the factors that predicted employee turnover — the changeable drivers, not fixed demographics. Those predictors were translated into interventions that reduced the percentage of people leaving. The result was communicated in business terms leaders could act on: a one-point reduction in turnover saved the firm 75 to 100 million dollars. Framing the insight as a dollar-impact recommendation tied to a top business problem is what made managers act, versus a dashboard metric they would have ignored.

• If a variable cannot be manipulated (sex, age), then use it only as a control and never as the basis of a recommendation — pivot to WHY it matters and what you can influence.
• If the data does not play a role in day-to-day business, then send a periodic emailed PDF report instead of relying on a self-service dashboard managers will forget to open.
• If a finding has no concrete follow-up intervention, then do not present it — build the intervention plan first.
• If the analysis does not connect to one of the company's top three strategic challenges, then it is not worth communicating (e.g., don't sell flight-risk analysis when average tenure is fifteen years).
• If a finding is sensitive at the individual level (attrition probability), then restrict who receives it to prevent misuse against specific employees.

• Skipping translation entirely — the fifth pitfall: producing interesting findings on a real business problem that hold no value because no one can act on them. Tell: results describe unchangeable attributes with no intervention attached.
• Handing managers a self-service dashboard and assuming use. Tell: login rates near zero and no decisions reference the data.
• Showing complicated statistical models to time-pressed executives. Tell: meetings stall on methodology instead of moving to a decision.
• Drowning the business in irrelevant numbers by analyzing data without a business-driven plan. Tell: 'intriguing' insights that have no connection to core business — a fancy HR showpiece.
• Presenting numbers that contradict a manager's gut feeling without translating them into a clear visual and narrative. Tell: the manager dismisses the analysis and reverts to intuition.

• ▸Business issue addressed: name which of the top three strategic challenges this serves.
• ▸Core finding in one sentence: state it stupidly simple, no model jargon.
• ▸Changeable driver identified: the manipulable variable behind the finding (not a control variable).
• ▸Recommended intervention(s): the concrete follow-up plan that solves the root cause.
• ▸Estimated business impact: quantify in the audience's terms (e.g., cost/dollars saved).
• ▸Delivery channel: emailed PDF report / dashboard / app message — chosen by audience behavior.
• ▸Cadence: how often it is sent (e.g., monthly, quarterly).
• ▸Visual format decision: which chart and layout, and why the default was rejected.
• ▸Audience and access limits: who receives it and any sensitive data withheld to prevent misuse.

1. 1Start with a decision that currently runs on gut feeling — hiring, promotion, absence planning — and write down the untested assumption behind it (e.g. 'our managers hire world-class performers' or 'more interviews mean better hires').
2. 2Pull the HR data you already hold that connects to the decision and combine it across systems so it can be analyzed; the reproducibility of results improves as you aggregate more data rather than mark your own paper on the same small set.
3. 3Replace subjective judgment with structure: standardize the inputs so different people cannot introduce their own bias. Google did this by having qDroid formulate the interview questions, forcing every interviewer to ask tested, predictive questions instead of their own.
4. 4Convert each subjective assessment into ratings on specific scales, then aggregate to a single number proven to correlate with the outcome you care about (job performance).
5. 5Run the numbers to test the assumption — count time spent, measure whether extra effort actually improves the outcome, and act top-down on the finding even when people resist (Google capped interviews at four).
6. 6Pick outcome measures that reduce the role of luck: favor frequent, upstream signals over rare final results, because more data points squeeze out the noise.
7. 7Validate the model on fresh data, not the data used to build it, so the accuracy estimate is objective — and keep gathering more data to sharpen the algorithm.

Google assumed its managers were hiring world-class performers and that many interviews produced better hires, but neither had been tested. Managers spent five to ten hours interviewing every new hire — over a hundred thousand hours across ten years — some almost full-time. When analysts ran the numbers, the process was time consuming without leading to better hires, and after the fourth interview managers estimated future performance just as well as after the tenth. Google removed human bias by automating interviews through qDroid, which supplies extensively-tested, performance-predictive questions so interviewers don't formulate their own, then stores answers and converts specific-scale ratings into a single predictive number. They forced a top-down decision to cap candidates at four interviews. The bias here was structural: unstructured, self-authored questions and an unquestioned belief that more effort meant better outcomes.

• If a decision rests on an assumption you have never tested against data, then run the numbers before acting — even 'obvious' beliefs like Google's hiring quality can be wrong.
• If interviewers or raters formulate their own questions, then replace them with tested, structured prompts and specific rating scales to strip out individual bias.
• If a successful outcome is rare and complex (long-term absence under 1% of 5,000 employees, or a 1.5-year B2B sales cycle), then measure a frequent upstream proxy (short/mid-term absence, number of contacts or calls) where luck acts as less noise.
• If additional effort (a tenth interview) does not improve prediction beyond an earlier point (the fourth), then cut it — more activity is not more accuracy.
• If you test a model on the same data used to build it, then treat the accuracy as inflated and re-test on fresh data before trusting it.

• Trusting a rare final outcome as your success metric — judging a hockey team by wins (2.3 goals/game) instead of shots-at-goal (10.6/game). Tell: your sample is tiny and results swing wildly game to game.
• Letting interviewers write their own questions. Tell: two interviewers rate the same candidate on incomparable dimensions and no answer maps to a proven predictor.
• Adding more human judgment to feel confident — extra interviews, more reviewers. Tell: estimation accuracy plateaus but time spent keeps climbing.
• Marking your own paper: validating a regression on the same dataset that produced it. Tell: the model looks accurate in-sample but has never faced fresh data.
• Stopping at descriptive reporting (sick-day counts, scorecards) and calling it evidence-based. Tell: you hit the wall of Boudreau — you can describe the past but cannot predict or explain cause, so decisions still default to gut feeling.

• ▸Decision under review: name the people decision currently made on gut feeling
• ▸Untested assumption: state the belief driving it in one sentence
• ▸Outcome metric: the rare final result you currently judge by
• ▸Frequent proxy: the upstream, higher-volume signal that reduces luck/noise
• ▸Data sources to combine: list the systems holding relevant data
• ▸Structure applied: how you standardize inputs (tested questions, fixed scales)
• ▸Single aggregate score: how ratings convert to one predictive number
• ▸Validation set: the fresh data you will test the model on, separate from build data
• ▸Action + owner: what changes based on evidence and who forces the decision

1. 1Frame the people decision as a testable question tied to a business outcome (e.g. 'which candidate should I hire?' or 'which site has the talent supply we need?') rather than a gut assumption.
2. 2Pull the HR data you already hold — interview scores, tenure, travel bookings, check-in times — and connect it to the outcome you care about (future performance, turnover, satisfaction).
3. 3Test your current predictors against actual results the way Google's task force compared interviewer predictions to real new-hire performance; find where added effort stops adding accuracy.
4. 4Remove human bias from the decision by structuring it: use pre-tested questions (like gDroid) that have been proven to predict the outcome, and record answers on specific scales that convert to a single predictive number.
5. 5Enrich the model with driver factors from occupational-psychology literature (age, tenure, seniority, travel distance, marital status, hotel bookings) rather than relying on the obvious demographic fields alone.
6. 6Validate the model on fresh data, not the same set used to build it — testing on the training set is like a student marking their own paper.
7. 7Interpret findings against known research and firm context so an odd result (high turnover in one division) becomes explainable and actionable.
8. 8Prioritise the drivers you can actually influence, and act even before you fully understand the mechanism.

A firm found turnover was unusually high in its international division but could not explain why. Standard fields — age, sex, education, seniority — did not account for it. Analysts added the number of hotel bookings per employee as a proxy for frequent international travel, treating travel as a stress factor grounded in turnover literature. That single factor greatly influenced actual turnover, especially for recently married women in their thirties. Crucially, hotel bookings and travel scheduling were something the firm could adjust relatively easily. The decision improved not because they found a new demographic to screen, but because they connected an operational variable to an outcome and then acted on a lever within their control.

• If added interview rounds or effort stop improving prediction accuracy (Google's fourth interview was as good as the tenth), then cap the process and reallocate the time.
• If a decision currently rests on an untested assumption ('our managers hire world-class performers'), then run the numbers before acting on it.
• If a driver is both statistically significant and something the firm can influence (hotel bookings, scheduling autonomy), then prioritise it over drivers you cannot change.
• If the training or intervention targets the wrong people (customer-service staff instead of the cleaners who actually meet customers), then re-identify who drives the outcome before spending more.
• If a model was tested only on the data used to build it, then withhold judgement until it is validated on fresh data.

• Investing heavily in a process without measuring its effectiveness — Google spent 100,000+ hours interviewing before checking whether it produced better hires; the tell is high spend with no impact figure.
• Training or intervening on the visible group rather than the group that actually drives the outcome — the cleaning firm trained customer-service staff; satisfaction didn't move until they trained the cleaners.
• Relying on gut feeling instead of validating assumptions — the tell is confident claims about people ('we hire the best') that have never been tested against outcomes.
• Overfitting: judging a model on the same data used to create it, giving falsely high accuracy — the tell is a model that looks great in-sample but no fresh-data test exists.
• Using only obvious demographic fields and missing context drivers like travel distance or hotel bookings — the tell is a result you cannot explain ('turnover is high but we don't know why').

• ▸Decision under review: name the hiring/retention/comp/planning choice
• ▸Current basis: is it gut feeling or data, and what assumption underlies it?
• ▸Business outcome: the measurable result this decision should drive
• ▸Predictor(s) in use: what you currently rely on to make the call
• ▸Effort vs. accuracy check: at what point does more effort stop improving accuracy?
• ▸Added driver factors: literature-based variables to test (tenure, travel, marital status, etc.)
• ▸Validation method: fresh dataset or holdout used to test the model
• ▸Actionable lever: which significant driver can the firm actually influence?

1. 1Name a top business challenge outside HR first — location strategy, downtime, attrition — then map which people question feeds it; analytics that answers HR-only questions adds no business impact.
2. 2Attach a dollar or performance figure to the outcome before you analyze, so impact is measurable: identify the cost per unit (per point of turnover, per minute of downtime, per dollar of personnel spend).
3. 3Isolate the driver you can move. Analyze the factors that predict the outcome (as Credit Suisse did with turnover predictors) rather than reporting the outcome itself.
4. 4Convert the finding into a two-part 'so that' statement linking the HR action to a business result to a financial result, per Ulrich.
5. 5Quantify the delta: multiply the change you can produce (one-point turnover reduction, avoided relocation cost) by the per-unit dollar figure to size the impact.
6. 6Translate the number into an actionable recommendation for a decision-maker — a board choice, a policy change — not a statistical model; executives are short on time.
7. 7Choose the delivery format for action: use a scheduled PDF report for data outside daily operations, since managers forget to log into dashboards.
8. 8Track whether the recommendation was acted on and whether the projected saving materialized; impact only counts once the business decides and moves.

Credit Suisse faced attrition costs estimated in the tens of millions but hard to calculate. Rather than reporting turnover rates, its analysts analyzed the factors that predicted employee turnover, then intervened on those drivers to reduce the percentage of people leaving. The measurable result: a one-point reduction in employee turnover saved the firm 75 to 100 million dollars. The impact came from three moves — sizing the per-unit value (each turnover point in dollars), isolating a movable driver, and multiplying the achievable reduction by that value. Compare Alexander in the foreword: he compared workforce plans against external talent supply-and-demand data for a provisional China site, saw the location could cost millions and choke projected growth, and swapped to better cities before the board committed. Both cases show impact = a business decision changed, priced in real money.

• If the question doesn't tie to a challenge outside HR, then reframe or drop it — HR-only analyses add no value to the business.
• If you can't state a per-unit dollar or performance figure (per turnover point, per downtime minute), then find that number before analyzing or the impact stays unmeasurable.
• If the audience is executives short on time, then present the actionable recommendation and its projected impact, not the statistical model.
• If the data sits outside day-to-day operations, then send a scheduled PDF report rather than relying on a self-service dashboard managers won't open.
• If your finding contradicts a manager's gut feeling, then invest extra effort in visualization and delivery, because presentation affects action more than the data itself.

• Reporting the outcome instead of its drivers — you show turnover is high but can't move it; the tell is recommendations that describe the problem rather than a lever.
• Committing to analyses that add no value (the 'fad' trap) — the tell is a project no manager asked about and no one acts on.
• Leaving impact unpriced — you cite a percentage improvement with no dollar figure attached; the tell is a board that can't compare it to other investments.
• Stopping at one 'so that' — 'I ensure uninterrupted staffing' without linking to productivity and downtime cost; the tell is HR describing its work, not its results.
• Dumping results into a dashboard nobody checks — the tell is data that lives online, is never opened, and produces no decision or action.

• ▸Business challenge (outside HR): the strategic problem this touches — location, downtime, attrition
• ▸People question: the specific analytics question feeding that challenge
• ▸Outcome metric + per-unit value: e.g., $75-100M per turnover point, ~$22K-$50K per downtime minute
• ▸Predictive driver: the movable factor your analysis isolates
• ▸Achievable change: the reduction/improvement you can realistically produce
• ▸Projected impact: change multiplied by per-unit value, in dollars or performance
• ▸'So that' chain: HR action so that business result so that financial result
• ▸Delivery format + audience: PDF report or dashboard, and the decision-maker who receives it
• ▸Action taken / result confirmed: what the business decided and whether the saving materialized

1. 1Start with a business question that matters to decision-makers, not an HR curiosity: 'Will leadership training make our managers better?' or 'Does sales training raise sales performance?' If the question would not change a board decision, discard it.
2. 2Convert the people problem into two currencies the rest of the business speaks: a numeric rating and a dollar amount. Frame every finding as ROI on a people policy so you can state, in money, how the policy helps the business earn more.
3. 3Assemble the three skillsets analytics requires — business acumen, data/statistics analysis, and IT/infrastructure knowledge — because these rarely coexist in a traditional HR team. Hire or partner for the missing skillset rather than forcing existing staff to fake it.
4. 4Pull the relevant data (internal workforce plans plus external supply/demand data where the decision needs it), then build the analysis: find cause and effect or a predictive model, not just descriptive summaries.
5. 5Benchmark your model against human judgment, not against perfection. If an algorithm predicts even 30% of future performance, it already outperforms a manager's hunch — use that gap as your argument for evidence-based decisions.
6. 6Translate 'analytical numbers' into actionable data visualizations and choose the delivery channel deliberately: a live meeting, a standing dashboard, or an email each demands different visualization. The way you communicate the result determines its impact.
7. 7Anticipate that the numbers may contradict the manager's or HR practitioner's gut feeling, and prepare to defend the finding with the ROI and the bias evidence rather than retreating.
8. 8Position the output to earn a seat at the table: measure HR's own contribution to an effective organization, show time and money saved, and repeat until HR is quantifying its impact and actively influencing business decisions — the threshold at which it becomes a true business partner rather than a business assistant.

Alexander, the Head of People Analytics at one of the world's largest technology companies, faced a board that had provisionally settled on a multi-million dollar site in China. He took the internal workforce plans — the number and skills needed at opening plus three-year growth projections — and compared them against external talent supply and demand data his team had collected for that city. He saw immediately there were not enough skilled potential employees, and the few available were in high demand from companies already established there: the company was about to make a very expensive mistake that could cost millions and block the projected growth. Rather than only flagging the risk, he pulled the same supply/demand data for other Chinese cities under consideration, found more favourable locations, and worked with analytics counterparts across the organisation. Together they produced a new, better set of recommendations for the board to act upon. This is HR influence in its mature form: data reframing a business decision before it is made, not HR reacting after.

• If a proposed analysis would not change a business decision or add value, kill it — a team that commits to analyses adding no value turns analytics into a fad.
• If a result can be expressed as a dollar amount or ROI, lead with that; if it can only be expressed as an HR metric, translate it before presenting.
• If a finding contradicts a manager's gut, trust the validated model over the hunch — humans are bad decision-makers, and even a 30%-accurate algorithm beats them.
• If you lack one of the three skillsets (business, data, IT/communication), hire for it rather than proceed — a missing skillset causes analytics to fail adoption, get stuck on simple analyses, or add no value.
• If the research concerns engagement, motivation, or attrition, do not import generalized findings from other firms — context differs by organization, so analyze for your firm specifically.

• Treating HR analytics as an efficiency exercise (cost reduction) instead of an effectiveness question — the tell is metrics reported as cost saved with no link to business impact or ROI.
• Running only simple descriptive analyses and never finding cause and effect or building predictive models — the tell is a team perpetually 'stuck' and never influencing a decision.
• Producing valid analyses that managers ignore — the tell is analytics not adopted by the business because results were never translated into actionable visualizations or the right communication channel.
• Generalizing people research across firms despite the reproducibility crisis and importance of context — the tell is applying an external engagement or attrition driver without validating it in your own organization.
• Assuming a CHRO title equals strategic influence — the tell is HR holding a board seat but still doing only operational/tactical work, leaving the position redundant to the CEO's people responsibility.

• ▸Business question: the decision this must change, stated as a question
• ▸Decision owner: which manager/board actually acts on the result
• ▸Data needed (internal): workforce/policy data you must aggregate
• ▸Data needed (external): supply/demand or benchmark data required
• ▸Skillset check: which of business / data / IT-communication you have vs. lack
• ▸Model & finding: cause-effect or prediction, with accuracy vs. human baseline
• ▸Dollar translation: the ROI or dollar impact of acting on the finding
• ▸Bias/context note: unconscious bias reduced or firm-specific context flagged
• ▸Communication plan: visualization + channel (meeting / dashboard / email)

1. 1List the CEO's top 3 business priorities in the CEO's own language (revenue growth, cost, right people with right skills to execute strategy) — not HR's language (support the line manager, be taken seriously, ensure flow of personnel).
2. 2Discard any candidate analytics topic that cannot be tied to one of those three, even if it is popular. Attrition and engagement scores are the usual traps; only keep them if attrition or engagement is an actual core business problem.
3. 3Check the priority against your industry and country context. Ask: what would materially help THIS company? Safety at Shell, talent supply for a new site — not a generic template lifted from another firm.
4. 4Convert the chosen priority into a specific question or hypothesis, because that hypothesis determines what data you select, how you analyze it, and what action follows.
5. 5Pressure-test relevance with the managers who own the problem: do they perceive it as a problem? If turnover is 6% and managers welcome vacancies to promote high potentials, turnover is not your project.
6. 6Bring in compliance at this stage, not later — confirm you can actually access the data (email, survey, individual records) the priority requires before committing resources.
7. 7State the added value explicitly: how solving this priority increases revenue, reduces cost, or secures the right skills. If you cannot state the value, the project is not aligned yet.

Alexander, Head of People Analytics at one of the world's largest technology companies, tied his work to a live top-priority decision: a multi-million-dollar choice of location for a key new operation in China. He took the workforce plan (headcount and skills needed at opening, plus growth over three years) and compared it against external talent supply-and-demand data his team collected for the provisional city. He saw the company was about to make an expensive mistake — not enough skilled candidates existed there, and the few who did were already in high demand from established firms. He then ran the same talent data against the other candidate cities, found more favorable locations, and worked with analytics counterparts across the organization to hand the board a better set of recommendations. The alignment is the point: the analysis attached to a strategic siting decision worth millions, so the insight changed a board-level action rather than producing an interesting but ignorable report.

• If the topic maps to one of the CEO's top 3 priorities, then proceed; if it does not, then park it regardless of how much trade-press attention it gets.
• If managers who own the outcome do not perceive it as a problem (e.g. low, welcomed turnover), then it adds no value — stop and pick a different question.
• If the priority is industry-specific (safety at a chemical company, talent supply for expansion), then weight it above generic HR metrics that any firm could run.
• If you cannot state how solving the question raises revenue, cuts cost, or secures the right skills to execute strategy, then the question is not yet business-aligned — reframe it.
• If compliance or data access can't be confirmed for the priority's required data, then rescope or reprioritize before investing time.

• Starting from the data instead of the business challenge (the Rasmussen and Ulrich warning) — tell: the project began with an available dataset and a search for something 'interesting' rather than a named priority.
• Defaulting to attrition analytics because it's the popular starting point — tell: nobody in the business calls attrition a problem, yet it's the headline of your project.
• Producing technically sound but irrelevant results — tell: the 6% turnover analysis was accurate but managers shrugged because vacancies let them promote high potentials.
• Framing goals in HR-internal terms ('support the line manager,' 'be taken more seriously') — tell: you can't answer 'so what is the business impact?' in one sentence.
• Lifting a priority from another industry or country — tell: you're running safety analytics at a legal firm, or applying a foreign turnover benchmark to a loyal Dutch workforce.

• ▸CEO top-3 priorities: write each in the CEO's language (revenue, cost, right skills for strategy).
• ▸Chosen priority: which one this project attaches to.
• ▸Industry/country context: the specific factor that makes this priority matter here (e.g. safety, expansion).
• ▸Problem owner + their view: name the manager and whether they call it a problem.
• ▸Question/hypothesis: the single testable statement derived from the priority.
• ▸Value statement: one line on how solving it raises revenue, cuts cost, or secures skills.
• ▸Required data + access check: which data the question needs and whether compliance permits it.
• ▸Decision it will change: the specific business action the result should inform.
• ▸Go / No-go: proceed only if value statement, owner view, and data access all clear.

1. 1Map your current team against the five capability contexts—business, marketing, HR, data analytics, IT—and mark each as present, thin, or absent. Business anchors 'ask an analytics question important to the business'; HR supplies process knowledge (recruitment, hiring, firing, compensation); data analytics runs correlation, regression, and structural equation modeling; IT aggregates data from multiple systems; communication capacity translates numbers into visualizations.
2. 2For each context, name the specific failure it prevents so gaps become visible: no business context means analyses add no value (fad); no HR context means struggling to select the right data and interpret outcomes; no data analytics means being stuck at simple analyses with no cause-and-effect or predictive models; no IT means being unable to obtain and aggregate relevant data; no communication means managers don't act on findings.
3. 3Test whether your team can pass the wall of Boudreau: to move from correlation to prediction and causation you need two distinct capabilities—data processing skills to aggregate and structure data across systems, and an advanced statistical skillset to run regression, root cause, and forecasting analyses. If either is missing, you stall at descriptive analytics.
4. 4Close absent contexts by choosing between two routes stated in the source: hire people with different skillsets, or partner with departments that already hold these skills in abundance (Finance and IT are named). Pick partnering when the skill lives elsewhere in the org and the volume of work is low; pick hiring when analytics is becoming a standing center of excellence.
5. 5Assign explicit ownership of communication as its own step—not a byproduct of analysis. Decide the channel (meeting, dashboard, or email) before building visuals, because each channel requires a different visualization, and matching them determines whether findings drive action.
6. 6Where possible, consolidate existing data sources into cloud/business intelligence storage so extraction is easier for data analytics professionals; this reduces the IT-aggregation gap that keeps teams below the wall of Boudreau.
7. 7Structure the team as a multidisciplinary center of excellence rather than an HR-only function, so it can pull in disciplines beyond HR and share knowledge, leadership, and best practices on the subject.

Alexander, Head of People Analytics at one of the world's largest technology businesses, faced a multi-million dollar decision: the company had provisionally settled on a China location for a key new operation. He combined internal capability with cross-functional reach—he held the workforce plans (numbers and skills needed at opening, plus three-year growth projections) and had earlier asked his team to collect external talent supply and demand data for the provisional city. Comparing the two, he saw the location lacked enough skilled potential employees and that those present were in high demand from established competitors, an expensive mistake in the making. Note the skillsets at play: HR/business context to know what workforce the strategy required, data analytics to compare supply against demand, and IT/data-gathering to source external data. He then evaluated alternative cities and worked with analytics counterparts across the rest of the organisation—partnering across departments rather than doing it alone—to produce a new, better set of recommendations for the board to act upon.

• If analyses keep producing insights the business ignores, then the missing context is business (or communication)—re-anchor every project to a stated business priority before selecting data.
• If the team can describe and correlate but cannot predict or find cause and effect, then add advanced statistical and data-processing skills—this is the wall of Boudreau and requires new hires or partnering.
• If you cannot obtain or aggregate the relevant data, then the gap is IT—consolidate sources in a cloud/BI system or partner with IT before commissioning any analysis.
• If managers dispute findings because numbers contradict their gut feeling, then invest in communication capacity—choose the channel first, then build the matching visualization.
• If a needed skill exists abundantly elsewhere in the organisation, partner with that department; if the analytics workload is continuous and strategic, hire the skill in and build a multidisciplinary center of excellence.

• Treating HR knowledge as sufficient—the tell is a solid understanding of HR practices but no one who can choose between correlation, regression, or structural equation modeling, so analysis defaults to the simplest (and often wrong) method.
• Committing to fashionable analyses with no business anchor—the tell is projects that are technically impressive but add no value to the business and quietly get dropped.
• Skipping communication as a step—the tell is findings that live in spreadsheets or a single email, with managers never acting on the data.
• Underestimating the IT/aggregation gap—the tell is the team stuck at simple analyses because data sits in different systems and can't be combined for proper analysis.
• Keeping analytics exclusively inside HR—the tell is a single-discipline team that never involves Finance or IT and stalls at descriptive maturity while believing it lacks 'tools' rather than skillsets.

• ▸Business context: name who translates strategy into the analytics question—present / thin / absent
• ▸HR context: name who supplies recruitment/hiring/firing/compensation process knowledge to select and interpret data—present / thin / absent
• ▸Data analytics context: list the techniques your team can actually run (correlation, regression, SEM, root cause, forecasting)—present / thin / absent
• ▸IT context: state how you obtain and aggregate data across systems (name the storage/BI solution)—present / thin / absent
• ▸Communication context: name who converts results to visualizations and the primary channel (meeting/dashboard/email)—present / thin / absent
• ▸Wall-of-Boudreau position: mark current level—tracking / reporting / correlation / prediction / causation
• ▸Closure route per gap: for each thin/absent context, choose HIRE or PARTNER and name the department or role
• ▸Business priority linked: state the top business priority this team's next project must serve
• ▸Owner and review date: who owns the gap-closing plan and when you re-audit

1. 1Adopt an HRIS as the precondition: level-1 organizations that keep records in an HRIS maintain reliable, consistent employee data, which stops teams from arguing about validity instead of interpreting what the data shows.
2. 2Run the validity check first: ask whether the data actually measures what you intend (e.g., does the appraisal measure performance or manager likability), and whether collection is skewed across parts of the organization.
3. 3Screen for bias in who is represented: confirm data is collected evenly, not concentrated in one group or channel, because uneven collection produces conclusions that only hold for the represented population.
4. 4Run the reliability check: verify the same thing measured repeatedly by different raters and at different times yields the same result; where scales are vague, tighten them and document collection procedures so every rater applies them identically.
5. 5When merging data from two systems, fix mislabeled function or section names and reconcile inconsistent labels before combining datasets, then store the cleaned data separately or feed it back into the source system so quality compounds over future analyses.
6. 6Work the six-step cleaning checklist: (1) check the data is up-to-date, (2) check for recurring unique identifiers, and continue through duplicates, missing records, and outliers.
7. 7Handle missing records deliberately: identify whether the people with missing data share characteristics (e.g., a whole department on an outdated performance system that omits questions), because that shared pattern skews results toward other departments and threatens generalizability.

The city of Boston built an app that let drivers report road bumps via GPS so the road service could fix them; a spokesperson said the data gave the city real-time information to fix problems and plan long-term investments. The flaw was in validity: the app was used mainly by younger people in more affluent communities, while poorer communities lacked equal access to smartphones and mobile data. That uneven collection is a significant bias — the data represented affluent roads far better than poor ones, so investment planning built on it would systematically underserve poorer areas. The tell was not in the numbers themselves but in asking who generated the data. Applying the validity questions — does the data represent what we want to measure, and are there biases in how it was collected — would have surfaced the skew before it drove funding decisions. The lesson: valid-looking real-time data can still measure the wrong population.

• If level-1 teams keep debating the data's validity instead of what the data shows, then stop analyzing and fix the source system (adopt or clean the HRIS) before going further.
• If two raters give the same subject different ratings, then treat the scale as vague or the procedure as undocumented — tighten the scale and document collection methods rather than averaging the disagreement away.
• If the data comes from only one channel or group (like an app used by affluent smartphone owners), then flag it as biased and narrow your claims to the represented population.
• If records are missing and the missing set shares a characteristic (same department, same outdated system), then do not treat the remaining data as representative — the gap will skew results.
• If data is outdated, then discard or refresh it; stale data produces irrelevant results that can corrupt the entire analysis.

• Assuming a metric measures performance when it measures something else — Google's interviews reliably measured only whether the interviewer liked the candidate, not future performance; the tell is a measure that correlates with rater preference rather than the target outcome.
• Trusting real-time, high-volume data without checking who produced it — the Boston bump app looked comprehensive but was skewed toward affluent smartphone users; the tell is data concentrated in one demographic or channel.
• Rater bias from inconsistent procedures — one manager reviews six months, another two weeks, producing divergent, unreliable ratings; the tell is the same subject scored differently by different people or at different times.
• Ignoring missing records that cluster — a department on an outdated system omits certain questions, so its people vanish from the dataset; the tell is a shrinking population and results that lean toward the departments still reporting.
• Merging datasets without reconciling labels — combining two systems with mislabeled functions or section names before cleaning; the tell is duplicate or inconsistent categories that make aggregation meaningless.

• ▸Up-to-date: Confirm every record's date and flag anything outdated for refresh or removal.
• ▸Unique identifiers: Check for recurring unique IDs and remove duplicates.
• ▸Missing records: List missing data and note whether the missing set shares a trait (department, system, group).
• ▸Validity — represents the target: State what you intend to measure and confirm the data reflects it, not a proxy (e.g., likability instead of performance).
• ▸Validity — collection bias: Identify who or what was excluded from collection and whether coverage is even across the organization.
• ▸Reliability — repeatability: Confirm the same thing measured multiple times produced the same result.
• ▸Reliability — documented procedures: Record whether collection methods were clearly documented and instructions followed each time.
• ▸Outliers: Flag values that deviate sharply and decide whether to investigate, keep, or exclude them.
• ▸Merge integrity: Reconcile mislabeled function/section names across systems and store cleaned data separately or feed it back into the source.

1. 1Score your organization first using the four self-assessment statements (HR provides actionable solutions based on data analysis; dedicated CHRO on the board; HR continually develops predictive models; HR engages in scenario-based strategic workforce planning), scoring 0 for strongly disagree up to 3 for strongly agree, then total to place yourself: 0-5 = Level 1, 6-11 = Level 2, 12-18 = Level 3, 19+ = Level 4.
2. 2If at Level 1, get a dedicated HRIS operating so you hold accurate, consistent employee records and can reliably produce operational metrics: headcount, attrition, labor cost, training cost.
3. 3To move from Level 1 to Level 2, stop generating static traditional reports and start producing proactive reports that support strategic decisions — apply benchmarks to make reports meaningful, report metrics multidimensionally, and deliver them to management and executives in clear dashboards.
4. 4To reach Level 3, build a centralized HR analytics department and integrate your various systems, typically via a business intelligence (BI) system that compounds data from multiple sources into a database usable for analysis, then begin answering business questions through statistical analysis.
5. 5When presenting Level 3 statistical work, do not show executives complicated models; focus on key business issues and translate every result into an actionable solution, since managers are short on time.
6. 6To reach Level 4, add dedicated analytical capability to your team (a more advanced statistical skillset) so you can build predictive models that show potential impact and forecast future scenarios.
7. 7Apply Level 4 predictive capability to strategic workforce planning: forecast how the workforce will develop, where leadership succession is jeopardized, and which positions you will need to fill in the near future so you can mitigate risk earlier.

Alexander, the Head of People Analytics at one of the world's largest technology businesses, works late reviewing a provisional multi-million-dollar site location in China. He holds internal workforce plans stating the number and skills of employees needed at opening plus projected growth over the first three years — that is his internal data. He then compares those plans against external talent supply-and-demand data he had his team collect for the provisional city, and quickly sees the problem: there are not enough skilled potential employees there, and the few who exist are in high demand from firms already established in that city. This is Level 3/4 work — combining integrated internal and external data to forecast a future constraint before it materializes. He pulls the supply-and-demand data for the other candidate cities, finds more favorable locations from a talent perspective, and works with analytics counterparts across the organization to bring the board a new, better set of recommendations — averting a mistake that could have cost millions and blocked projected growth.

• If your self-assessment total is 0-5, then you are Level 1 and your first move is securing an HRIS and reliable operational reporting — not chasing predictive models.
• If you can produce accurate reports but cannot combine data across systems, then you are stuck below Level 3 — build a centralized analytics function and BI-based data integration before attempting statistical analysis.
• If you are presenting statistical results and the audience is executives, then strip the model and lead with the key business issue and a concrete recommended action.
• If leadership succession risk or future position needs are decisions you must inform, then you require Level 4 predictive/scenario capability, which means adding advanced statistical skills to the team.
• If you believe you are optimally organized for analytics success, then treat that belief with suspicion — Bersin (2016) found 92% of companies believe they are not optimally organized, so honest placement usually sits lower than assumed.

• Skipping to predictive ambitions while stuck in operational reporting — the tell: you talk about predicting performance but still only report headcount, attrition, labor cost and training cost (only 8% of organizations had actually used analytics to predict performance in 2016).
• Trying to run analyses without integrating data — the tell: your data lives in separate systems with no centralized HR analytics department, so you cannot compound sources for analysis.
• Assuming HR already has the skills to advance — the tell: no advanced statistical skillset on the team, which is exactly what stops HR from puncturing the 'wall of Boudreau' into analytical and predictive capability.
• Reporting metrics as static traditional numbers instead of benchmarked, multidimensional dashboards — the tell: reports are generated for hygiene, not to proactively support strategic decisions.
• Overwhelming executives with complex models — the tell: you present statistical detail rather than a translated, actionable solution tied to a key business issue, and decisions stall.

• ▸Statement 1 score (HR provides actionable solutions based on data analysis): enter 0-3
• ▸Statement 2 score (dedicated CHRO on board of directors): enter 0-3
• ▸Statement 3 score (HR continually develops predictive models for strategic decisions): enter 0-3
• ▸Statement 4 score (HR does scenario-based strategic workforce planning using predictive models): enter 0-3
• ▸Total score & level: sum the above; 0-5=L1, 6-11=L2, 12-18=L3, 19+=L4
• ▸Current-state evidence: name the reports/analyses you actually produce today (e.g. headcount, attrition, benchmarks, statistical models)
• ▸Capability gap: state what your level lacks (HRIS, BI/data integration, centralized analytics team, advanced statistical skillset)
• ▸Next-level target action: write the single concrete step to advance one level (per how-to)
• ▸Business question to prove value: name one key business issue your next-level capability will answer with an actionable solution

1. 1Start from a business priority framed as a testable question, not a data pull — e.g. 'Can a hiring manager predict employee performance?' Bad questions produce analyses that add no value to the business (the 'fad' failure).
2. 2Map which systems hold the data the question requires. Google needed interviewer ratings, the Applicant Tracking System, and the Performance Review System — inventory each source before you commit to an analysis.
3. 3Aggregate the sources using programming skills and knowledge of your IT infrastructure. Merge on unique identifiers and confirm data labels match across the merged datasets.
4. 4Clean the data through the six-step checklist: verify it's current, resolve duplicate/reoccurring unique identifiers, match labels across fields, count missing values, flag numerical outliers, and remove invalid outputs.
5. 5Select the technique that fits the question — correlation, regression, or structural equation modeling — recognizing some methods handle missing data and some struggle. Descriptive reporting (sick days, scorecards, benchmarking) is not enough to find cause-and-effect or build predictive models.
6. 6Interpret the output against the business, expecting it to sometimes contradict the manager's gut feeling — that contradiction is the point of a fact-based approach.
7. 7Transform 'analytical numbers' into actionable visualizations and match the channel to the audience: a meeting, a dashboard, or an email each require different visualization.
8. 8When a needed skillset is absent in HR, hire for it or partner with departments that hold it in abundance — Finance and IT.

Google assumed its managers hired world-class performers but had never tested it. Rather than trust the hunch, the head of HR crunched the numbers to see whether the interview process actually predicted future performance. To do this they aggregated data across three systems: interviewer ratings, the Applicant Tracking System, and the Performance Review System. They then applied an analytical technique — correlating interview scores with later performance, the kind of question answerable by regression or structural equation modeling — because literature suggested managers could not in fact predict performance. The numbers confirmed the hunch was wrong, which let HR justify new hiring practices. Because interviewing consumes so much employee time, even small improvements produced large gains, and selecting better candidates built a stronger workforce that added to long-term profitability.

• If the analysis doesn't map to a top business priority, then stop — you are building a fad that adds no value.
• If missing values are over-represented in one department or one part of the organization, then treat your results as non-generalizable until you understand what those people share.
• If your chosen algorithm struggles with missing data, then either switch to a method that tolerates it or narrow the population deliberately and disclose the limit.
• If the numbers contradict the manager's gut feeling, then present the numbers and the validation, not a softened version — that contradiction is the evidence.
• If the team can run only descriptive reporting (scorecards, benchmarks, sick-day counts), then you lack the analytical capability to find cause-and-effect or build predictive models; acquire the skill before promising more.

• Getting stuck on simple descriptive analyses — reporting sick days and scorecards while never finding cause-and-effect. Tell: outputs describe what happened but never explain why or predict what's next.
• Committing to analyses no one asked for. Tell: the work is technically impressive but no manager acts on it — a 'fad' with no business owner.
• Selecting the wrong technique or building an invalid model. Tell: results can't survive scrutiny, and you can't defend why you chose correlation over regression or SEM.
• Failing to aggregate all relevant data. Tell: an outdated performance system in one department omits questions, so that department is silently absent from the dataset, skewing results toward others.
• Neglecting communication and visualization. Tell: correct 'analytical numbers' sit unread because they weren't turned into an actionable visual matched to the channel — HR analytics fails to be adopted by the business.

• ▸Business question — state the priority as one testable question a manager cares about.
• ▸Source systems — list every system holding required data (e.g. ATS, Performance Review System).
• ▸Aggregation owner — name who has the programming/IT-infrastructure skill to merge the sources.
• ▸Data-cleaning status — mark each of the six steps done: currency, duplicate IDs, label match, missing counts, outliers, invalid removal.
• ▸Missing-data risk — note any department/group over-represented in missing values and why.
• ▸Technique chosen — name the method (correlation / regression / SEM) and why it fits the question and the missing-data profile.
• ▸Gut-feeling conflict — record which manager assumption the result confirms or overturns.
• ▸Communication plan — specify channel (meeting / dashboard / email) and the matching visualization.
• ▸Skill gap & source — list any missing skillset and whether you hire or partner (Finance/IT) to close it.