The AI Marketing Canvas: A Five-Step AI Plan for Marketers

1. 1Reframe your marketing questions from deterministic to probabilistic: stop asking 'what is the best group to target' and start asking 'what is most likely the best group to target,' then let computational methods analyze the data to make the prediction.
2. 2Fund continuous learning as a line item, not a nicety: pay for AI conference attendance, buy and expense AI marketing books, and run lunch-and-learns with external speakers so the team keeps pace as the technology accelerates.
3. 3Install fast-failure tolerance by running Agile experiments in step 2 — pick a few marketing activities, use third-party AI tools to get quick outcomes, and treat failed experiments as learning rather than punishable mistakes.
4. 4Force collaboration and shared learning so best practices circulate; require teams to document what works and share learnings so everyone gets smarter together rather than hoarding insight.
5. 5Quantify the ROI of the new AI way versus the old way at every experiment, building a repeatable body of proof — because you are asking management to make a big bet on an approach they won't automatically understand.
6. 6Use the 'risk of ignoring' argument internally: position inaction as the larger cost, since brands clinging to outdated mindsets risk falling irrevocably behind as the landscape accelerates.
7. 7Name an AI marketing champion in step 3 and begin insourcing capability — identify internal data science resources to design models and run experiments in-house, keeping only best-of-breed vendors for what they do best.

Estée Lauder Companies faced a wake-up call after reporting a 6 percent drop in organic net sales for Q2 fiscal 2025. CEO de La Faverie admitted on the earnings call, 'We lost our agility. We did not capitalize on the higher growth opportunities quickly enough' — a direct diagnosis of a culture that had lost speed. His 'Beauty Reimagined' plan put AI at its heart: 'We are ready to hardwire AI through the organization,' committing to use it to target consumers more effectively and unleash creativity in brand teams. ELC then partnered with Adobe's Firefly suite so creative teams spend less time on repetitive work like versioning digital ads for different channels and formats, freeing them for ideation and new campaign development. This shows the culture shift in practice: agility and speed restored, repetitive tasks automated, and human talent redirected to strategic, high-value work.

• If proving ROI on a GenAI investment is hard, then still act — weigh it against the 'risk of ignoring,' where the cost of inaction is far greater.
• If a marketing task is repetitive (like versioning ads across channels and formats), then hand it to AI and redirect the team to strategic work like ideation.
• If it makes strategic sense to reduce vendor dependence, then engage internal data science resources to build models in-house; otherwise keep the best-of-the-best partners doing what they do best.
• If your organization has lost agility and isn't capitalizing on opportunities quickly enough, then treat culture — speed, experimentation, fast failure — as the fix, not just new tools.
• If you want teams to adopt continuous learning, then you (the leader) pay for the conferences, books, and lunch-and-learns rather than expecting them to self-fund.

• Treating AI as optional; the tell is a leadership team still debating whether to start while high-performing peers already use it (84 percent for creative, 52 percent for strategy).
• Clinging to deterministic thinking — demanding the single 'right' answer instead of the most probable one; the tell is teams rejecting model outputs because they aren't certainties.
• Investing in GenAI but expecting instant payoff; the tell is CMOs declaring investments 'haven't paid off yet' because no repeatable ROI proof was built through experimentation.
• Ignoring the cultural component and buying tools only; the tell is 27 percent-style stalled adoption where people lack the willingness and ability to fail fast and collaborate.
• Losing agility through slow decision-making; the tell is ELC's own confession — not capitalizing on higher-growth opportunities quickly enough despite seeing them.

• ▸Deterministic-to-probabilistic shift: name one recurring question your team asks and rewrite it in 'most likely' terms.
• ▸Fast-failure norm: describe how failed experiments are currently handled and one change to make failure safe.
• ▸Continuous-learning budget: list the conferences, books, and lunch-and-learns you will fund this quarter.
• ▸Collaboration mechanism: state where and how teams will document and share learnings.
• ▸ROI proof plan: name the experiment and the old-way vs. new-way metric you will compare.
• ▸Risk-of-ignoring statement: write one sentence on what inaction costs your brand.
• ▸AI marketing champion: name the person accountable for driving the shift.
• ▸Insource vs. vendor call: list one capability to build in-house and one best-of-breed partner to keep.
• ▸Speed target: define one repetitive task to automate and where freed time gets redirected.

1. 1Consolidate first-party data as the foundation: capture every click, search, and interaction across touchpoints and route it into a system (an advanced customer data platform) that can act on terabytes-to-petabytes of your own internal data — this is what distinguishes Wave 3 from Wave 2's rules-based logic.
2. 2Move off if-then rules. Wherever you have a static rule like 'if no purchase in 3 months, then send we-missed-you email plus 10% off,' replace it with a model that predicts next-best message, offer, channel, and timing per individual based on recent behavior.
3. 3Instrument the four decision variables the AI must resolve for each customer in real time: the right next message, the right offer, the right channel, and the right timing — then let the system create, send, and track results in a closed loop.
4. 4Add dynamic asset generation so content adapts per person. Follow CarMax's model: dynamically adjust images and text based on individual search behavior so each customer sees the most relevant inventory and faces less cognitive overload.
5. 5Deploy generative AI to remove versioning bottlenecks. As ELC does with Adobe Firefly, let AI handle repetitive tasks like versioning digital ads across channels and formats so human teams shift to ideation and campaign development.
6. 6Build internal 'generative AI brains' (Mondelēz's approach) that surface customer insights, test ideas, generate content at scale, recommend performance changes, and optimize channel allocation — treat personalization as an internal operations capability, not just an external output.
7. 7Benchmark against the best digital experience anywhere, not just competitors. CarMax competes against every digital experience a consumer has, so measure personalization relevance against that bar.
8. 8Run adoption incrementally so failure teaches instead of destroys — avoid the million-dollar, year-long initiative that delivers no consumer value; test in small increments to learn safely.

Two customers receive the identical offer; Customer A opens but doesn't click, and Customer B also opens but doesn't click. Under Wave 2 rules-based logic they would both trigger the same next email. Under Wave 3, an AI-powered system (an advanced customer data platform) reads each customer's recent behavior separately and determines the right next best message, offer, channel, and timing for each — then generates and sends the differentiated assets, tracks the results, and feeds outcomes back into the model. CarMax runs a comparable pattern on CarMax.com, dynamically adjusting images and text on the basis of individual search behavior so each shopper sees the most relevant inventory. Starbucks does it inside its app, proactively recommending new drinks and food based on each customer's past preferences and purchase history. The common thread: identical starting inputs, individualized outputs, decided by prediction and generation rather than a shared rule.

• If you can write your personalization as an if-then rule, then it belongs to Wave 2 — upgrade it to a predictive model before calling it personalization at scale.
• If demand for personalized experiences outstrips human capacity to create and deliver them, then AI is the required ally, not an optional enhancement.
• If a task is repetitive versioning of assets across channels and formats, then hand it to generative AI and redeploy human time to ideation and strategy (the ELC/Firefly split).
• If your data foundation is thin, then prioritize first-party direct-to-consumer data — Airbnb outperformed state-of-the-art models by using its own data and custom-built algorithms.
• If an initiative would cost a million dollars and a year before delivering any consumer value, then break it into incremental tests so failure stays cheap and instructive.

• Staying frozen: acknowledging AI's potential but lacking a structured framework, so awareness never becomes action. Tell — endless discussion, no live personalization in production.
• Aiming too high with tactics only — launching an expensive initiative that burns a million dollars and a year without delivering consumer value. Tell — spend precedes any measured customer benefit.
• Treating AI as the tech team's problem. The CNBC warning applies: a CEO can't just say 'my tech guys will educate me.' Tell — leaders can't explain how AI changes a customer decision.
• Mistaking rules for personalization — keeping if-then logic (three-month lapse email + 10% coupon) and believing it's one-to-one. Tell — two identically-behaving-then-diverging customers still get identical follow-ups.
• Benchmarking only against direct competitors, not the best digital experience in any industry. Tell — relevance feels adequate internally while consumers defect to smoother experiences elsewhere.
• Ignoring quantification — no clear metrics on AI's impact leads to hesitation and incomplete adoption. Tell — pilots run indefinitely with no profit-growth or overachievement measure.

• ▸First-party data source: name the internal click/search/interaction data feeding your system.
• ▸Rule being retired: the current if-then rule you will replace with prediction.
• ▸Next-best-message model: how the system decides the individual message.
• ▸Next-best-offer model: how it selects the offer per customer.
• ▸Next-best-channel + timing: how it chooses where and when to deliver.
• ▸Dynamic asset generation: which images/text/versions generative AI will produce automatically.
• ▸Human refocus: the strategic tasks freed up once repetitive versioning is automated.
• ▸Benchmark experience: the best digital experience (any industry) you measure relevance against.
• ▸Success metric + increment: the profit/overachievement metric and the small, safe test that proves it.

1. 1Split trust into two audiences from the start: human customers who need emotional confidence, and their AI agents who evaluate you on price, performance, ethics, reputation, and value alignment. Build separate criteria for each because 'trust, in this future, becomes programmable.'
2. 2Publish machine-readable brand and product signals so an agent given a prompt like 'stylish guest bathroom makeover for under $2,500, finished by June 1' can find, evaluate, and act on your offer. If your systems can't 'communicate fluently and persuasively with those agents,' you won't be considered.
3. 3Secure consumer data at every step of the AI interaction. Inventory where sensitive data flows through agentic handoffs and add safeguards specifically at points where an agent handles research, evaluation, purchasing, and logistics on the customer's behalf.
4. 4Stand up AI governance: write clear guidelines on how your systems operate and schedule regular audits so you can trace where things went wrong when an agent's actions cause a bad experience or ethical dilemma.
5. 5Make decisions explainable. Because it's difficult to understand the reasoning behind an autonomous agent's decisions, build logging and reasoning-trace capability so you can pinpoint failures rather than guess.
6. 6Optimize your brand signals for machine reasoning as deliberately as you optimize campaigns for human emotion — design 'systems that are as transparent and trustworthy to AI agents as they are compelling to people.'
7. 7Set goals and guardrails for your own outbound agents (campaigns, spend, messaging), then let them execute — but keep humans owning the guardrails so trust breaches don't propagate autonomously.

A customer tells their personal AI agent: 'Find me a stylish guest bathroom makeover for under $2,500, finished by June 1.' From that moment the agent, not the customer, handles research, evaluation, purchasing, and logistics — the customer may never see your home page, category pages, or Instagram ads. A home-goods brand that has structured its product data, pricing, availability, delivery timelines, and sustainability/ethics claims into machine-readable form gets surfaced, compared, and selected; a competitor whose signals live only in human-facing creative is filtered out and never considered. The agent weighs price, performance, ethics, reputation, and alignment with the customer's values before it negotiates and buys. If the winning brand's systems mishandle the data passed during that transaction, the breach undermines trust between the consumer and the brand for future interactions. Winning here is about being trustworthy to the algorithm, not just persuasive to the person.

• If an interaction will be handled by a customer's AI agent, then optimize for machine decision-making frameworks (price, performance, ethics, reputation, value alignment), not just human emotion.
• If your brand's systems can't communicate fluently and persuasively with agents, then treat that as being invisible — no consideration, no sale — and fix data legibility before spending on creative.
• If you deploy autonomous agents on your own side, then define goals and guardrails first and keep humans controlling the guardrails, because you can't easily trace an agent's reasoning after the fact.
• If sensitive customer data passes through an agentic handoff, then require security safeguards at that step, since mishandling or misuse undermines trust between consumers and brands.
• If an agent's action causes an unintended bad experience or ethical dilemma, then run an audit against documented operating guidelines rather than assuming a one-off error.

• Optimizing only for human emotion and experience while ignoring the agent's decision-making models — tell: strong brand affinity surveys but your products never surface in agent-mediated shortlists.
• Treating agentic AI as an efficiency gain rather than a new operating model — tell: you bolt agents onto existing workflows without rethinking roles, governance, or data legibility.
• Deploying autonomous agents with no reasoning trace or audit trail — tell: when a bad customer experience happens you can't pinpoint where it went wrong.
• Letting sensitive data flow through agent interactions without step-by-step safeguards — tell: security is treated as a one-time perimeter, not applied at every handoff, leaving breach exposure.
• Assuming trust is only a human-facing feeling — tell: no transparency or value-alignment signals are exposed to the agents evaluating whether your brand is trustworthy.

• ▸Human trust signals: emotional and experiential claims your brand makes to people.
• ▸Agent trust signals: machine-readable price, performance, ethics, reputation, and value-alignment data.
• ▸Agent legibility test: can an agent find and act on your offer from a natural-language goal? (yes/no + gap).
• ▸Data flow map: every point where sensitive customer data moves through an agentic handoff.
• ▸Safeguards per step: the security control applied at each handoff point above.
• ▸Governance guidelines: written rules for how each AI system operates.
• ▸Audit cadence: schedule and owner for regular checks that systems function properly.
• ▸Reasoning traceability: mechanism to explain/log why an agent made a given decision.
• ▸Guardrail owner: the human accountable for goals and guardrails on your outbound agents.

1. 1Map each of your existing customer touchpoints against the four pillars—acquisition, retention, growth, advocacy—and mark which are still linear (attract, convert, hope for repeat) versus which respond to real-time behavior.
2. 2Build the data foundation first: collect enough clean first-party data across the business to train machine-learning models. Use RFM value tables and a data lake to consolidate; treat clean data as the precondition, not an afterthought.
3. 3For acquisition moments, stop relying on broad demographic assumptions—configure AI to dynamically adjust messaging, visuals, and offers on the basis of real-time consumer behavior and intent (programmatic, PPC, CTR optimization).
4. 4For retention moments, deploy predictive models to anticipate needs and proactively address concerns before they cause churn—automated chatbots for support, predictive retention scoring, personalized 24-hour support (the Disney animatronic name-pronunciation model).
5. 5For growth and advocacy, use behavioral data to build relationships and incentivize the next best action—Disney's MagicBand tracked movements and purchasing habits so cast members could nudge guests toward another attraction.
6. 6Sequence the effort through the canvas steps: experiment with third-party tools for quick wins, quantify impact to justify budget, then expand across more activities and deepen into one or two moments in the transformation step.
7. 7Anchor every AI moment to a behavioral objective—before personalizing, state what change in consumer behavior you are trying to effect, so tech serves the relationship rather than existing for its own sake.

Disney Research and Imagineering built the MagicBand, a theme-park-wide pass that tracked guests' movements, analyzed purchasing habits, and reported real-time data back to Disney. This converted a static park visit into a fluid, data-driven moment: cast members could anticipate customer behavior and make quick decisions—adding staff at pressure points or incentivizing guests to head to a less crowded ride (retention and growth). The result was improved park efficiency and capacity, higher operating margins, and continued strong traffic trends. More recently, Disney is training animatronic cast members to correctly pronounce guests' names, addressing the disconnection guests feel when AI voices fail at this, and creating immersive, personalized relationships (growth and advocacy). As chairman Josh D'Amaro framed it, the purpose of the AI tech 'at the end of the day is to create a connection with our guests. To make them feel something.' The tech is the mechanism; the relationship is the goal.

• If a touchpoint still follows attract-convert-hope, then it is a static journey and is your first candidate for AI-moment conversion.
• If you cannot state the behavior change you want to trigger, then do not build the personalization—AI is not a stand-alone tool, so define the end before the mechanism.
• If your first-party data is not clean (RFM tables, consolidated data lake), then stay in the foundation step; you lack the quality data to train models that supercharge moments.
• If an AI capability spans two pillars—like MagicBand's retention-and-growth or animatronics' growth-and-advocacy—then prioritize it, because cross-pillar moments compound relationship value.
• If you have quick experimental wins but no quantified impact, then you cannot make the budget case for expansion—measure before you scale.

• Treating AI as an efficiency tool only—the tell is a roadmap full of time-and-cost savings with no line item for personalization or loyalty outcomes.
• Personalizing without a behavioral objective—the tell is 'we're using AI to personalize messaging' with no answer to 'to what end, what change in consumer behavior?'
• Building AI moments on dirty data—the tell is models producing accuracy risk and off-target outputs because clean, first-party data was never established in the foundation step.
• Keeping journeys linear while calling them AI—the tell is broad demographic segmentation and batch campaigns rather than messaging that evolves in real time on behavior and intent.
• Optimizing one pillar in isolation—the tell is strong acquisition metrics (CTR, PPC) with churn quietly rising because retention and advocacy moments were never instrumented.

• ▸Touchpoint: name the specific customer interaction
• ▸Pillar(s): mark acquisition / retention / growth / advocacy (allow multiple)
• ▸Current state: is it linear (attract-convert-hope) or already data-driven?
• ▸Behavior change goal: the exact consumer behavior you want to trigger
• ▸Data required: which first-party signals feed this moment (RFM, real-time behavior, intent)?
• ▸Data readiness: is the data clean and model-trainable now? yes/no
• ▸Canvas step: foundation / experimentation / expansion / transformation / monetization
• ▸Impact metric: how you will quantify the change to justify budget
• ▸Real-time trigger: what live signal fires the personalized action

1. 1Follow the five-step Canvas in order (Foundation, Experimentation, Expansion, Transformation, Monetization) so AI enhances rather than disrupts existing strategy — the final step, Monetization, is explicitly where you measure and maximize AI's impact on marketing ROI and secure long-term competitive advantage.
2. 2Tie every AI initiative to a specific change in consumer behavior across one of the four customer moments; before deploying, answer 'to what end?' — personalization is a means to durable, profitable growth, not an end.
3. 3Baseline the financial state you are trying to move: name the metric under pressure (e.g., organic net sales, profit-growth margin) so you can prove incremental lift later, since the passages warn that proving ROI is a genuine challenge.
4. 4Benchmark your AI adoption pace against high performers: Gartner found high-performing orgs adopt AI faster, using it for creative development (84%) and strategy development (52%) versus 77% and 48% for peers — set adoption targets against those figures.
5. 5Split use cases into external (customer experience/personalization) and internal (marketing operations) so you capture both revenue-side growth and efficiency gains — the passages show both the Starbucks-style external personalization and Mondelēz-style internal 'generative AI brains.'
6. 6Build a road map that demonstrates early success to unlock further investment: take stock by asking the right questions, lay a foundation, then execute — Brett Groom's sequence is 'demonstrate success, and get the support and investment you need.'
7. 7Position for winner-take-all dynamics: the authors' evidence points to winner-take-all outcomes, so treat AI investment as a defensive necessity (business survival) as much as an offensive growth play.

On February 4, 2025, thirty days into his tenure, Estée Lauder CEO Stéphane de La Faverie faced a 6 percent drop in organic net sales for Q2 FY2025 in an increasingly competitive beauty market. He diagnosed the cause plainly — 'We lost our agility. We did not capitalize on the higher growth opportunities quickly enough' — and launched a 'Beauty Reimagined' plan to restore sustainable sales growth. The plan combined increased advertising and marketing spend, rapid innovation, and a leaner org, but its core commitment was to 'hardwire AI through the organization,' especially in marketing. ELC targeted AI at better consumer targeting and globally personalized messaging — right ad, right consumer, right time — and greater creative output. It then signed a strategic partnership with Adobe's Firefly suite so brand teams could stop spending time on repetitive tasks like versioning digital ads across channels and shift toward more engaging, faster campaigns. The move ties a concrete financial problem (declining net sales) to specific AI use cases (targeting, personalization, creative versioning) aimed at returning the company to profitable growth.

• If you cannot name the consumer behavior change and the customer moment an AI use case targets, then do not fund it — AI is a mechanism for the relationship, not a stand-alone tool.
• If a declining financial metric is the trigger (like ELC's organic net sales drop), then anchor the AI plan to restoring that specific number and increase marketing spend behind proven use cases.
• If competitors are adopting AI faster and outcomes trend winner-take-all, then treat AI investment as a survival requirement, not an optional experiment.
• If an AI initiative shows early promise in Experimentation, then scale it across functions and channels (Expansion) before attempting to reshape workflows and talent (Transformation).
• If you need more investment, then first demonstrate documented success on an early use case so you can 'get the support and investment you need.'

• Adopting AI because 'every marketer thinks they need to use it' without knowing why — the tell is initiatives with no defined behavior change or business outcome attached (Brimmer's warning).
• Skipping steps and jumping to Transformation before laying the Foundation — the tell is disrupted existing strategies and stalled teams instead of enhanced ones.
• Investing in GenAI but never proving payoff — the tell matches Gartner's finding that 27% have limited/no adoption and many CMOs feel investments 'have yet to pay off' because ROI proof was never built in.
• Losing agility and moving too slowly on growth opportunities — the exact self-diagnosis de La Faverie made at ELC before the 6% sales drop.
• Treating AI as only external (customer-facing) or only internal (operations), missing half the value — the tell is efficiency gains with no growth lift, or personalization with no workflow productivity.

• ▸Financial metric under pressure: state the baseline number you must move (e.g., organic net sales -6%).
• ▸Target customer moment: name which of the four moments this initiative addresses.
• ▸Consumer behavior change: describe the specific behavior shift you intend to cause.
• ▸Canvas step: mark current step (Foundation / Experimentation / Expansion / Transformation / Monetization).
• ▸Use-case type: classify as external (personalization/CX) or internal (marketing operations).
• ▸Adoption benchmark: record your AI use % vs. high-performer marks (84% creative, 52% strategy).
• ▸Early success proof: name the demonstrated result you will show to secure more investment.
• ▸ROI measure: define how you will measure incremental profitable growth and competitive advantage.
• ▸Investment ask: state the funding/support requested once success is demonstrated.

1. 1Confirm you have completed steps 2-4: most AI/ML capabilities are in-house (or the model is under marketing's control), you own the data feedback loops 100 percent, and the AI is powering most of the customer relationship. Do not attempt monetization from off-the-shelf, third-party dependence.
2. 2Audit the proprietary AI capabilities you built for internal use and ask which ones, if exposed externally, would create value for other firms like yours. Separate the algorithm from the system — the sellable asset is the accessed, experienced, valued system, not the model alone.
3. 3Apply Lafley and Martin's where-to-play/how-to-win test: name the specific external market segment (adjacent customers, suppliers, or peers), and define the defensible advantage that makes your solution profitable, scalable, and hard to copy.
4. 4Convert the internal capability into a market-facing product form: a platform, a service, or licensed software. Redesign how customers access and experience the capability, not just what the model outputs.
5. 5Choose a licensing/monetization architecture from the three core types: permanent (one-time activation), time-based (renewable subscription), and use-based (consumption pricing, e.g., per unit consumed). Match the license type to how the customer derives and measures value.
6. 6Price around demonstrated customer ROI so the payment scales with realized value — bill on actual usage where possible so no use means no fee and high precision means higher customer return.
7. 7Pilot with a defined external customer set, measure adoption and recurring revenue, then scale. Recognize this step is rare and optional — many firms win by mastering earlier steps, so proceed only with strategic clarity and in-house innovation capacity.

John Deere illustrates the move from AI user to AI seller. For decades its model was: build equipment, sell it, occasionally service it. Having built AI on top of its machines for internal use, it commercialized those capabilities as solutions-as-a-service: precision agriculture kits, autonomous navigation, real-time data analytics, and a licensing model. See & Spray uses computer vision to detect and treat only the weeds that need attention; rather than charging up front, Deere bills farmers monthly based on how many acres are actually sprayed — no use, no fee; high precision, higher ROI. Its licensing architecture spans permanent, time-based, and use-based licenses (e.g., StarFire guidance, ExactShot fertilizer micro-dosing), turning AI into renewable, recurring revenue. The passages stress that Deere didn't just build better models — it redesigned how those models are accessed, experienced, and valued, which is why its step 5 maturity is potent and rare. Contrast this with GE's Predix: an ambitious AI-powered industrial IoT platform aiming to be a top-ten AI platform, which stumbled despite deep industrial expertise and rich data.

• If your AI still depends on third-party off-the-shelf products or the model isn't under your control, then do not attempt step 5 — return to transformation and bring capabilities and feedback loops fully in-house first.
• If a proprietary capability creates strategic advantage AND would deliver value to external customers like you, then it is a monetization candidate; if it only drives internal efficiency, keep it internal.
• If customers derive value in proportion to consumption, then use a use-based license (e.g., per acre sprayed); if value is continuous, use a time-based subscription; if value is fixed at activation, use a permanent license.
• If your team believes the algorithm is the product, then stop — the sellable asset is the system: how the capability is accessed, experienced, and valued.
• If you lack strategic clarity on where to play and how to win, then don't take step 5 — winning at earlier steps is a legitimate endpoint.

• Chasing platform ambition without the system redesign — the tell is a bold platform vision (like GE's top-ten-AI-platform goal for Predix) that leans on technical prowess and rich data but never re-architects how external customers access and value the capability.
• Skipping to step 5 on off-the-shelf tools you don't own — the tell is that you can't control the model or the data feedback loops, so you have no defensible, proprietary asset to sell.
• Selling the algorithm instead of the system — the tell is a product pitch centered on model accuracy rather than on customer experience, access, and realized ROI.
• Charging up front regardless of usage — the tell is customer resistance because pricing doesn't track the value they actually receive; compare to Deere billing only for acres sprayed.
• Treating step 5 as mandatory maturity — the tell is forcing external commercialization when the firm has no strategic clarity or in-house innovation capacity, ignoring that few firms venture this far and many succeed at earlier steps.

• ▸Proprietary capability: name the AI capability built for internal use you want to commercialize
• ▸In-house control check: is the model and its data feedback loop owned 100 percent by you? (yes/no)
• ▸External market: which customers like you would pay for this, and why
• ▸Where to play / how to win: the segment plus the defensible advantage
• ▸System redesign: how customers will access, experience, and value it (not just model output)
• ▸License type: permanent, time-based, or use-based — matched to how value is realized
• ▸Pricing basis: the unit of value you bill against (e.g., per acre used)
• ▸ROI proof: the wins/business case from steps 2-4 supporting external sale
• ▸Go/no-go: given difficulty and strategic clarity, do you proceed or master earlier steps?

1. 1Automate collection of high-volume, time-sensitive data where analytics or speed wins — e.g., pull benchmark data on a fixed cadence rather than manually (the insurance example updates quotes every fifteen minutes).
2. 2Centralize analytics and reporting so a single owner sees the cross-channel big picture: which products sold in which channel, to which customers, and where customers switched between products.
3. 3Connect your databases and collect first-party customer data at every relationship moment you can, following the Starbucks Rewards model of capturing behavior directly from loyal users.
4. 4Impose data-quality processes that make input (collection) and output (delivery) uniform and complete: every record 'clean,' attributed to the right customer and the right product.
5. 5Organize the data around the individual customer rather than by function or division, so machine-learning models train on a complete customer view instead of siloed fragments.
6. 6Build the transfer path to the machine: for a large firm, wire the database directly to the ML system; for a small direct-to-consumer firm, export to a .csv and upload.
7. 7Prioritize standardization early if you've grown through acquisition or run many channels/divisions — connecting and standardizing is where the effort concentrates, but benefits accrue faster once the data is in order.

Starbucks launched Deep Brew in 2019 to personalize offers and recommendations on its mobile app, but the data foundation behind it took fifteen years of continuous refinement in and around the Rewards app. Rewards is the mechanism that captures first-party data on what customers buy, how often, and where and when — the raw material Deep Brew feeds to AI and machine-learning systems. That data is useless without an efficient, accurate way to process it, so Starbucks built the digital infrastructure to amass millions of loyal users as the main data source. The clean, connected data lets the machine learn and predict what customers actually want, which is then expressed as personalized app recommendations. The lesson the authors draw: if you didn't start fifteen years ago like Starbucks, the next best time is today.

• If data is unstructured (like written property reviews), then apply an algorithm to convert it to structured numeric values before it can enter a model — as Airbnb did by scoring review text into a polarity metric with qdap.
• If you're a large firm with many divisions or growth-by-acquisition, then front-load connecting and standardizing data, because that's where the difficulty lives; if you're a small direct-to-consumer firm, then a .csv export may suffice.
• If a data field can't be reliably attributed to the correct customer and product, then treat it as unclean and fix the input process before feeding the machine.
• If analytics intelligence or speed would give an advantage in a data-heavy process, then automate that process rather than run it manually.
• If reporting can't show a big-picture view across channels, then centralize before you invest further in AI — a category manager needs the complete historic customer picture to see cross-channel and switching behavior.

• Feeding unstructured raw data straight into a model — the tell is a model that won't run or produces nonsense because text/words were never converted to numeric, structured values.
• Data organized by function or channel instead of by customer — the tell is you can't assemble a single complete view of one customer's history across channels.
• Misattribution in collection — the tell is reviews or purchases mapped to the wrong customer or product, so 'clean' fails at the input stage.
• Assuming AI is the starting point rather than the foundation — the tell is investing in models before you've amassed enough quality first-party data to train them.
• Underestimating standardization cost after acquisitions or channel sprawl — the tell is stalled projects because databases won't connect or fields don't match across divisions.

• ▸Automated data processes: list high-volume/time-sensitive collections you've automated and their cadence.
• ▸Centralized reporting: name the owner who can see the cross-channel, historic customer picture.
• ▸Connected databases: list which systems are wired together and which remain siloed.
• ▸First-party capture points: identify each relationship moment where you collect customer data directly.
• ▸Data organization: confirm records are keyed to the individual customer, not the function/channel.
• ▸Structuring step: note any unstructured data (text, reviews) and the algorithm used to make it numeric.
• ▸Data-quality rule: state how you verify input/output is uniform, complete, and correctly attributed.
• ▸Transfer method: specify how data reaches the ML system (direct database link or .csv export).
• ▸Standardization gap: list division/channel/acquisition mismatches still to be resolved.

1. 1Confirm you have completed Step 1 Foundation first: enough quality first-party data collected across the business to begin training models. Experimentation applies AI tools on top of that data.
2. 2Map your four customer relationship moments — acquisition, retention, growth, advocacy — and identify specific marketing activities within them where a value pocket exists (a task that is repetitive, costly, or slow enough that a quick win is plausible).
3. 3Divert a small, contained budget to skunkworks initiatives rather than seeking large upfront investment. Keep scope narrow: a few marketing activities, not a full-function rollout.
4. 4Select third-party/vendor AI tools rather than building in-house at this stage. Use a vendor checklist (Figure 14) to evaluate options and rely on the technical vocabulary from chapters 4 and 5 to be a smart consumer and communicate with data science contacts.
5. 5Adopt an Agile approach to running the initiative — short cycles aimed at quick learning rather than periodic, campaign-based planning.
6. 6Define what 'quick learning' and 'quick outcome' mean for each experiment before you start, so you can tell whether the experiment produced a usable result.
7. 7Document what works — the outcomes and learnings — because Step 3 Expansion requires you to quantify impact to build the business case for the additional budget you will later request.

Estée Lauder Companies illustrates where experimentation with vendor tools leads. Under its Beauty Reimagined plan, ELC committed to scale AI across the organization, particularly in marketing, to target consumers more effectively and personalize messaging globally. Rather than building tools from scratch, ELC announced a strategic partnership with Adobe to adopt its Firefly suite of AI tools. The value pocket was concrete: creative teams spent excessive time on repetitive work like versioning digital ads for different channels and formats. Firefly handles that versioning, freeing teams for strategic tasks such as ideation and new campaign development. This is the experimentation pattern — apply a third-party tool to an identified value pocket in a customer relationship moment to get a quick, measurable outcome — though ELC's ambition to 'hardwire AI through the organization' points toward the later expansion and transformation steps.

• If your first-party data is not yet collected at quality, then stay in Foundation — experimenting without data yields no durable learning.
• If a marketing activity is repetitive, high-volume, and time-consuming (e.g., ad versioning), then it is a strong value-pocket candidate for a vendor tool experiment.
• If a vendor cannot be evaluated against your checklist and you cannot converse with them using AI/ML vocabulary, then delay adoption until your team can be a smart consumer.
• If an experiment produces a quantifiable win, then document it for the Step 3 business case; if it does not, then kill it fast rather than expanding it.
• If you are being asked to build or deeply insource the model now, then you have jumped ahead of Experimentation into Expansion or Transformation.

• Skipping Foundation and experimenting on thin data — the tell: experiments generate activity but no reliable learning because there is no quality first-party data underneath.
• Buying into vendor-driven hype and 'AI expert' advice without a structured framework — the tell: a scattered, ad hoc collection of tools rather than experiments tied to specific value pockets and moments.
• Over-scoping the skunkworks into a full rollout — the tell: budget and timeline balloon and Agile speed disappears, reverting to periodic campaign-based planning.
• Failing to define or measure outcomes — the tell: you cannot quantify impact when Step 3 arrives, so you cannot justify additional budget and adoption stalls.
• Treating efficiency as the only use case — the tell: every experiment is about saving time, ignoring the other marketing-use cases for AI.

• ▸Customer relationship moment: name one — acquisition, retention, growth, or advocacy.
• ▸Marketing activity: the specific task being targeted (e.g., ad versioning).
• ▸Value pocket: why this task is costly/slow/repetitive enough to warrant an experiment.
• ▸Vendor tool: the third-party tool chosen and how it scored on the vendor checklist.
• ▸Foundation check: confirm the first-party data needed exists at quality.
• ▸Quick-learning goal: the one question this experiment must answer.
• ▸Success metric: the quantifiable outcome that defines a win.
• ▸Agile cycle length: the short timebox for the experiment.
• ▸Documentation owner: who records results for the Step 3 business case.

1. 1Quantify the impact of your step 2 experiments first: document lift, cost savings, and revenue from each AI-powered activity so you have hard numbers to justify additional budget for expansion.
2. 2Select the winning experiments and scale them across more customer relationship moments (acquisition, retention, growth, advocacy) — do more of what worked, less of what didn't, and keep documenting results for your business case.
3. 3Appoint an AI marketing champion who owns the expansion agenda, coordinates campaign design/run/evaluate cycles, and keeps model development under marketing's control.
4. 4Begin insourcing: hire or develop data science competency inside marketing so you gradually reduce reliance on third-party vendors and own the models rather than renting them.
5. 5Follow the Airbnb path when insourcing — start from an off-the-shelf model, then customize it to your proprietary first-party data instead of building from scratch, and set a measurable accuracy target before deploying.
6. 6Run campaigns as controlled, incremental experiments so failure stays cheap and instructive; avoid the million-dollar, year-long build that delivers no consumer value.
7. 7Preserve a documented record of every model, result, and cost so that at step 4 you can make an evidence-based build-vs-buy decision and bring most AI/ML capabilities in-house.

Airbnb illustrates expansion through insourcing. Because listing photos make or break a rental decision, in 2018 Airbnb's own data scientists took an off-the-shelf neural net model, ResNet50, and customized it to categorize its massive database of listing photos by room type using computer vision and deep learning. They experimented with training options and reliable data-labeling methods, achieving 95 percent certainty that an image classified as outdoors was in fact outdoors. This served two scaled purposes across customer moments: grouping same-room photos into a simple home tour and validating host-provided information. By using its own direct-to-consumer data and custom-built algorithms, Airbnb outperformed industry state-of-the-art models. The lesson for expansion is that owning proprietary data plus a customized model beats depending on a generic vendor solution.

• If an experiment produced measurable lift, then scale it into additional customer moments; if it can't be quantified, then don't expand it — keep it in experimentation until it can.
• If you still rely on a vendor at expansion, then require that the model's development stays under marketing's control so you retain the option to insource later.
• If an off-the-shelf model plus your first-party data can hit your accuracy target, then customize it rather than build from scratch (the Airbnb ResNet50 pattern).
• If no single person owns the AI agenda, then appoint an AI marketing champion before scaling further.
• If an initiative would cost heavily and take a year with uncertain consumer value, then break it into incremental, cheap-to-fail experiments instead.

• Scaling before quantifying: expanding experiments you can't measure — the tell is a budget request with no lift, cost, or revenue numbers behind it.
• Aiming too high: launching an expensive initiative that takes a million dollars and a year but delivers no consumer value — the tell is scope inflation with no incremental milestones.
• Staying vendor-dependent: outsourcing model development so the vendor, not marketing, controls the model — the tell is you can't modify or explain the model yourself.
• No champion: treating AI as a fragmented series of disconnected experiments with no owner — the tell is duplicated pilots and no cohesive expansion roadmap.
• Building from scratch unnecessarily: ignoring proven off-the-shelf models you could customize with your own data — the tell is long development cycles reinventing capabilities that already exist.

• ▸Proven experiment: which step-2 activity you're scaling
• ▸Quantified impact: measured lift / cost saving / revenue to date
• ▸Customer moments targeted: acquisition, retention, growth, and/or advocacy
• ▸AI marketing champion: named owner and their mandate
• ▸Data science competency: in-house skills present vs. gaps to hire/develop
• ▸Model control: vendor-led, marketing-controlled, or fully in-house
• ▸Build approach: off-the-shelf-plus-customization or build-from-scratch, with accuracy target
• ▸Vendor dependence: current reliance and plan to reduce it
• ▸Budget case: documented results supporting the next funding request

1. 1Appoint the champion as you transition from step 2 (Experimentation) to step 3 (Expansion) — not earlier. The trigger is that you have 'a bunch of AI and machine-learning plates spinning' and quantified results from your step 2 experiments worth scaling.
2. 2Staff the role with a 'marketing technologist': someone who understands data and marketing, is excited about AI/ML possibilities, AND has technology in their background including Agile-based software project-management experience. Do not fill it with a pure marketer or a pure data scientist — the point is translation between the two.
3. 3Assign the translator mandate explicitly: the champion functions as the interface between marketing and data science, converting marketing value pockets into problems a data science team can predict or generate against, and converting model outputs back into marketing decisions.
4. 4Task the champion with quantifying the impact of your step 2 experiments so you can document what works, do more of it, and assemble evidence for a budget request (both CapEx and OpEx).
5. 5Have the champion identify resources, cultivate vendor relationships, and hunt for new value pockets across marketing activities as you expand AI beyond the first channel.
6. 6Direct the champion to build the business case that promotes AI activities across silos and, when ready, goes to management to secure additional resources.
7. 7On the technology side, have the champion manage the Agile process (regular test sprints) and the implementation of algorithms, and begin insourcing some development and capabilities.
8. 8Prepare the champion to run the eventual build-vs-buy analysis: weighing internal resources to create proprietary in-house models against engaging an outside strategic partner.

Compare.com shows the kind of translation work a champion coordinates, even though it began at the experimentation stage. Compare's completion rate on its insurance questionnaire fell from 18 percent in January 2016 to 12 percent in March — a 33 percent decline — a firm-level strategic problem because Compare is only paid when a customer accepts a quote. The designated value pocket was clear: increase the number of visitors who complete the questionnaire. The team converted this into two predictive questions — does showing predicted rates earlier increase completion, and which customer groups respond best and where — then engaged an intern to build a basic algorithm predicting potential auto-insurance savings earlier in the flow. Notice the translator's operating questions: where to present the estimate, to whom (test vs control groups), how many questions are sufficient for a reliable estimate, and what copy draws customers to the predicted quote. This is exactly the champion's job at scale: figure out what you are trying to predict, then work with a data science team to build and test it inside an Agile test-and-learn culture. The Compare case demonstrates why the champion needs both marketing framing (the value pocket, the copy) and technical fluency (the algorithm, the test design).

• If you are still in step 2 with only ad hoc third-party AI tools, then hold off appointing a champion; if plates are spinning across multiple activities and you are entering expansion, then appoint one now.
• If a candidate understands marketing but has no Agile software project-management experience, then they are not qualified — the role explicitly requires technology in their background.
• If your step 2 experiments are not yet quantified, then the champion's first job is measurement, because no business case survives management without documented impact.
• If model development can stay under marketing's control and internal resources can build proprietary models, then favor building in-house; if not, then engage an outside strategic partner — the champion runs this build-vs-buy analysis at step 4.
• If an AI initiative stalls at a silo boundary, then route it through the champion, whose mandate is to advocate across silos and reinforce importance to internal staff.

• Hiring a pure marketer or pure data scientist. The tell: meetings where marketing and data science talk past each other with no one translating value pockets into predictions and back.
• Appointing the champion too late (after transformation) or too early (during initial experimentation). The tell: either scaling chaos with no coordinator, or a champion with nothing yet to expand.
• Skipping impact quantification of step 2 experiments. The tell: a budget request to management with enthusiasm but no documented results, so CapEx/OpEx is denied.
• Treating the role as purely marketing and ignoring the Agile/algorithm-implementation responsibilities. The tell: sprints drift, algorithms don't ship, and vendor relationships go unmanaged.
• Letting the champion lead in isolation. The tell: the passage warns it's everyone's responsibility to get on board and yours to lead the charge — if only the champion cares, adoption dies at the silo walls.

• ▸Readiness trigger: name the step 2 experiments currently 'spinning' and their status.
• ▸Candidate profile: evidence of marketing fluency, data fluency, and Agile software project-management experience.
• ▸Translation scope: the specific marketing-to-data-science handoffs this person owns.
• ▸Value pockets to pursue: list the marketing activities targeted for AI expansion.
• ▸Quantified step 2 results: documented impact numbers to carry into the business case.
• ▸Vendor relationships to cultivate: current third-party tools and prospective partners.
• ▸Business case owner tasks: what must be proven to management to secure CapEx and OpEx.
• ▸Agile/algorithm responsibilities: sprint cadence and algorithms slated for implementation.
• ▸Build-vs-buy checkpoint: criteria for insourcing models vs engaging an outside strategic partner.

1. 1Confirm you have completed Expansion before starting Transformation: AI must already run across a broad set of marketing activities with insourcing underway, and you must have quantified experiment impact. Transformation is step 4 in the sequential canvas — do not attempt it out of order.
2. 2Map every marketing activity against the four customer relationship moments — acquisition, retention, growth, advocacy — and mark which are AI-powered versus manual. The target state for Transformation is a complete set of activities automated across all four moments, not a scattered subset.
3. 3Reshape the workflow, not just the tools: replace legacy periodic, campaign-based planning (built on intuition, experience, scheduled cycles) with continuous, real-time optimization that AI's speed and automation require. Redesign roles and decision-making frameworks around this cadence.
4. 4Make the build-and/or-buy decision for each strategic capability (see Figure 19). Move from third-party vendor tools used in Experimentation toward bringing strategic AI capabilities in-house where they are core to your competitive position; keep buying where the capability is commodity.
5. 5Install and empower the AI marketing champion and team introduced during Expansion to own the transformed workflows, so automation across moments has clear organizational accountability.
6. 6Rebuild talent strategy to match: staff for continuous AI operation and in-house development, not one-off campaign execution, and integrate AI into decision-making frameworks rather than bolting it on ad hoc.
7. 7Verify the transformation is cohesive, not fragmented: each automated moment must trace back to an intended change in customer behavior across the journey. Prepare metrics to carry into step 5, Monetization, where you measure and maximize AI's ROI.

Consider the AI-transformed gym-joining journey described in the passages. In the traditional model, a prospect saw an ad, visited, spoke with a salesperson, and signed up — a manual, sequential, human-driven path. In the transformed model, AI-powered systems detect a consumer's interest in fitness from online behavior (acquisition), dynamically personalize landing pages to match their preferences, recommend membership tiers based on data insights (growth), and follow up with automated yet personalized email sequences. This is Full Transformation because a complete set of activities is automated across customer relationship moments rather than one isolated experiment. The workflow itself changed: personalization now evolves in real time instead of being planned in campaign cycles. Coca-Cola and Ancestry are cited in the passages as brands documented at these later canvas steps, showing the pattern is observed, not theoretical.

• If a capability is core to your competitive advantage, then build it in-house; if it is commodity or you need speed to learn, then buy from a vendor (build and/or buy, Figure 19).
• If your marketing still runs on periodic, campaign-based planning, then you are not transformed yet — real-time continuous optimization is the marker of Transformation.
• If AI covers only some marketing activities or only one or two customer moments, then you are still in Expansion, not Transformation; Transformation requires a complete set across acquisition, retention, growth, and advocacy.
• If you cannot name the change in customer behavior a given automated moment is meant to produce, then remove or redesign it — the canvas requires customer-centric, not tech-driven, automation.
• If you have not quantified experiment impact and installed an AI marketing champion, then hold: Transformation must build on Expansion's groundwork.

• Fragmented, ad hoc adoption: automating disconnected experiments instead of a cohesive set across all four moments. Tell — no single map showing every activity and its moment, and initiatives that don't share metrics.
• Skipping the sequence: jumping to Transformation without Foundation data, Experimentation learning, and Expansion insourcing. Tell — models lack quality first-party data and there is no in-house champion to own workflows.
• Automating tools while keeping legacy processes: bolting AI onto intuition-led, campaign-cycle planning. Tell — decisions still happen on a periodic calendar rather than continuously in real time.
• Buying everything, building nothing: outsourcing even strategic capabilities to vendors, leaving no proprietary in-house advantage. Tell — competitors using the same vendor stack achieve identical results.
• Tech-first, customer-blind transformation: personalizing 'to what end' is never answered. Tell — activities exist because AI enables them, not because they change a customer behavior in the journey.

• ▸Activity: name one marketing activity
• ▸Customer moment: acquisition, retention, growth, or advocacy
• ▸Current state: manual, vendor-tool, or AI-automated
• ▸Target behavior change: what customer action this should shift
• ▸Build or buy: mark strategic (build) or commodity (buy)
• ▸Workflow cadence: periodic/campaign vs. real-time continuous
• ▸Owner: the AI marketing champion or team accountable
• ▸Data readiness: quality first-party data available? yes/no
• ▸Impact metric: how ROI will be measured in Monetization