The Retail Executive's AI Stack: What to Buy, What to Build, and What to Kill Before Your Next Board Meeting.

Quick answer

Buy commodity AI for personalization engines, generative content, and demand forecasting, these are solved problems where off-the-shelf vendors deliver 3.2x higher early ROI than custom builds (Forrester, 2026). Build only where your proprietary data creates a defensible edge that a vendor cannot replicate. Kill anything sitting in the 51% of your stack you are already not using. The average enterprise actively uses fewer than half its contracted martech tools (Gartner, 2025), and that dead weight is the first thing your board will find.

TL;DR

Six numbers that define the retail AI stack decision in 2026:

• 49% martech utilization, five-year low.

  The 2025 Gartner Marketing Technology Survey found martech utilization at 49%, with only 15% of organizations qualifying as high performers who meet strategic goals and demonstrate positive ROI (Gartner, 2025).

• 80% of AI projects produce no measurable business value.

  RAND Corporation's analysis of 2,400+ enterprise AI initiatives found an 80.3% overall failure rate. In 2025 alone, more than $547 billion of $684 billion invested in AI failed to deliver intended business value (Pertama Partners, 2026).

• AI personalization drives 10-15% revenue uplift, consistently.

  McKinsey benchmarks personalization revenue lift at 10-15% for retailers that implement it, with fast-growing companies generating 40% more revenue from personalization than slower-growing peers (McKinsey, 2025).

• 70% of enterprise AI use cases are solved by off-the-shelf tools.

  McKinsey's 2025 analysis found 70% of enterprise AI use cases are adequately served by existing vendor solutions. Companies that piloted a buy-first approach before committing to custom builds reported 3.2x higher ROI (Forrester, 2026).

• CMOs allocate 15.3% of marketing budget to AI, with only 30% organizational readiness.

  Gartner's 2026 CMO Spend Survey found CMOs now allocate an average of 15.3% of marketing budgets to AI initiatives. At the same time, only 30% say their organizations have mature or fully developed AI readiness capabilities (Gartner, 2026).

• Retail AI personalization failures trace to one root cause: bad data integration.

  60% of retail SME AI personalization initiatives fail due to poor data integration, not tool quality (Workinsiders, 2025). Data fragmentation is the top-cited blocker to AI marketing ROI across Gartner and eMarketer research (Gartner, 2025).

Why the Stack Problem Is Getting Worse, Not Better

The martech ecosystem crossed 15,384 distinct solutions in 2025, a 9% year-over-year increase across 49 categories (Chiefmartec, 2025). The average enterprise now runs 91 martech tools. Actively used: fewer than 40% of them. That utilization number has dropped in five consecutive Gartner annual surveys. Meanwhile, the 2026 Gartner CMO Spend Survey found that 62% of CMOs plan to invest more in marketing technology this year, even as martech as a share of the total marketing budget hit a five-year low of 19.4%, down from 26.6% in 2021 (Gartner, 2026). The contradiction is arithmetic: more tools being bought at lower relative spend while utilization falls. Something has to give, and it will be the tools that cannot connect to a unified data layer and produce a documented revenue outcome. The forcing function is AI itself. Disconnected technology environments block AI agents from delivering consistent customer experiences (ADAPT CIO Edge Survey, 2025). CMOs who do not rationalize first cannot build effectively on top.

The Three-Tier Retail AI Stack: A Working Map

Think in three architectural tiers, not in vendor categories. Tier 1 is the data foundation: your customer data platform or cloud data warehouse, first-party identity graph, and purchase and behavioral signal capture. Tier 2 is the intelligence layer: the models and decisioning engines that read from Tier 1 and produce outputs, personalization, pricing, demand forecasting, content generation. Tier 3 is the activation layer: email, SMS, paid media, site experience, and in-store systems that execute the decisions Tier 2 produces. Most retail stacks have too many competing tools at Tier 3 and a broken Tier 1. AI cannot fix a broken Tier 1. Generative AI traffic to U.S. retail sites grew 4,700% year-over-year as of mid-2025 (Adobe Digital Insights, 2025), that volume of inbound intent signal is wasted if your data layer cannot capture and act on it. Build your audit starting at Tier 1, not at the shiny Tier 2 purchases.

What to Buy: Solved Problems Where Vendors Win

A use case belongs in the buy column when it is commercially solved, does not encode your proprietary advantage, and where vendor training data dwarfs anything you can produce internally. For most mid-market retailers, that covers five categories.

• Personalization and recommendation engines.

  AI-driven personalization delivers 10-15% revenue uplift on average (McKinsey, 2025). Personalized product recommendations drive 26-31% of e-commerce revenue on average (Salesforce/Barilliance, per Vovv.ai, 2026). Vendor platforms trained on cross-retailer behavioral data outperform anything a single $200M retailer can build in under 24 months. Buy.

• Generative content production.

  46% of DTC brands used AI-generated product visuals in 2026. Personalized creative lifts click-through rates 22-40% on paid social (Vovv.ai, 2026). The underlying models are commodity; your brand voice and creative direction are the differentiator. Buy the platform, own the creative standards.

• AI-native customer service deflection.

  AI chatbots now resolve up to 86% of customer questions without human intervention (Tidio, 2025). Retailers see $3.50 back for every dollar invested in AI customer service (Ringly.io synthesis, 2026). The economics of buying a proven platform are unambiguous.

• Demand forecasting and inventory optimization.

  AI reduces forecasting errors by 20-50% in supply chain management per McKinsey's State of AI 2025. A $500M retailer carrying $100M in inventory saves $15-30M in carrying costs through AI-driven replenishment (The Thinking Company, 2026). Vendor solutions with cross-industry training data get there faster than internal builds.

• Paid media and audience optimization.

  Platform-native AI (Meta Advantage+, Google Performance Max) operates on signal volumes no retail brand can match independently. 85% of enterprise AI budgets in 2025-26 went to platform selection and integration rather than ground-up model training (McKinsey, 2025). Allocate accordingly.

What to Build: The 20% Where Your Data Is the Moat

The build case rests on one question: do you own proprietary data that a vendor cannot replicate, and does that data encode your competitive advantage? For most retailers, this is a short list, shorter than most AI roadmaps acknowledge. Custom AI model development costs range from $150,000 to $5 million or more; off-the-shelf solutions typically run $20,000-$200,000 per year for comparable scope (Technobrave, 2026). The math only works when the proprietary output is worth the delta. Retailers who built custom AI without first validating with a vendor proof-of-concept wasted an average 14 months (Gartner, 2025). Three valid build cases exist for most retail CMOs.

• Proprietary loyalty and lifetime value modeling.

  If your loyalty program contains 5+ years of transaction-level data with unique behavioral signals, category switching, seasonal elasticity, reactivation patterns, that dataset is yours. A vendor model trained on generalized retail data cannot replicate it. Build the LTV and churn prediction layer on top of your own warehouse.

• Category-specific assortment intelligence.

  Retailers with deep private-label or specialty category expertise hold assortment signal that no horizontal AI vendor has indexed. A specialty outdoor retailer's gear-compatibility graph or a home goods chain's regional style-preference model are build candidates. The proprietary taxonomy is the moat.

• Cross-channel attribution with owned-media signal.

  If you operate a retail media network or have a high-volume loyalty email program, your attribution signal is a strategic asset. Building a closed-loop measurement model on your own data avoids the structural bias baked into vendor attribution tools that serve competing advertisers.

What to Kill: The Audit Criteria Your Board Will Use Anyway

S&P Global's 2025 Voice of the Enterprise survey found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the prior year, a 147% increase in abandonment rate (S&P Global, 2025). Boards are now doing what CMOs should have done 18 months ago. Get ahead of it. Kill a tool when it meets any of these three criteria.

• Active utilization below 40%.

  Gartner's 2025 survey found that only 15% of organizations qualify as martech high performers. The separating factor is not the size of the stack, it is data integration depth and documented revenue outcomes. A tool with sub-40% active use has already been voted out by your team.

• Duplicate capability without a unified data connection.

  68% of enterprise CMOs plan to double AI spending in 12 months, but only 22% have defined clear integration roadmaps for AI tools inside existing stacks (Gartner, 2025). If two tools produce the same output and neither feeds a central data layer, one of them is overhead.

• No documented revenue moment it owns.

  Score every tool against one question: what specific revenue decision breaks if this disappears tomorrow? If the answer requires more than one sentence and references indirect influence, the tool is a cost center. Kill it before the next budget cycle.

• Vendor AI bundling that forces capability you cannot use.

  Major vendors introduced AI features bundled into existing plans at 10-20% price premiums in 2025-26, regardless of whether customers use them (SaaS Management Index, 2026). If you are paying the AI uplift on a platform where your team has not activated the AI features, that is a clean kill candidate for vendor renegotiation.

The Sequencing Problem Most CMOs Get Wrong

The right sequence is: data layer first, kill redundancies second, buy proven use cases third, build proprietary intelligence last. Nearly every failed AI initiative inverts this. Generative AI pilots fail to scale at a 95% rate (MIT Project NANDA, 2025). The models are rarely the weak point. The data layer beneath them is broken, or the organizational process connecting model output to execution decisions does not exist. Gartner's 2026 CMO survey found 70% of CMOs say becoming an AI leader is a critical goal, and the same 70% admits internal processes are not mature enough to implement and scale AI effectively (Gartner, 2026). That is not an ambition problem. That is a sequencing problem. The CMOs posting 10-15% revenue lift from personalization did not start with an AI strategy. They started with a data audit and a utilization cull, then bought into proven platforms, then built on top of the signal their clean data layer generated.

What Could Go Wrong

Honest risks, stated plainly:

• You rationalize the stack without fixing the data layer beneath it.

  Cutting tools does not clean data. If your customer identity resolution is broken, every AI platform you buy will underperform against its benchmark. Data integration failure is the proximate cause of 60% of retail AI personalization failures (Workinsiders, 2025).

• The 14-month custom build trap.

  Enterprises that built custom AI without first validating with a vendor proof-of-concept wasted an average 14 months before realizing a commodity solution was adequate (Gartner, 2025). Sunk cost pressure then makes it harder to kill the build and buy the platform.

• Vendor lock-in on the buy side.

  Forrester's 2025 analysis warned that consolidating to a primary AI platform 'concentrates risk and neuters your negotiation leverage.' Buying 80% of your intelligence layer from a single vendor creates contractual exposure that will surface at renewal.

• AI spend outpaces organizational readiness.

  CMOs now allocate 15.3% of marketing budgets to AI, but only 30% report mature AI readiness in their organizations (Gartner, 2026). Tools do not fix process gaps. Buying ahead of internal capability produces the 80% failure statistic.

• Board pressure accelerates the kill list arbitrarily.

  42% of companies abandoned most AI initiatives in 2025, a 147% increase from 2024 (S&P Global, 2025). Some of those kills were correct. Many were reactive budget cuts that eliminated initiatives weeks before they would have reached their ROI threshold. Have a documented performance timeline for every AI investment before the next board meeting, not after.

The J.Caresse Point of View

The buy-build-kill framework sounds like a procurement exercise. It is actually a test of organizational honesty. Most mid-market retail CMOs we work with arrive with a stack that grew through opportunistic buying, a pilot here, a vendor demo there, an AI feature bundled into a renewal. The result is a portfolio where Tier 1 (data) is fragile, Tier 2 (intelligence) is duplicative, and Tier 3 (activation) is where everyone is focused. That inversion is why 80% of AI investments produce no measurable EBIT impact (McKinsey, 2025). The fix is not another tool purchase. It is a structured audit that assigns every dollar in the AI stack to a documented revenue outcome, and terminates anything that cannot make that case in one sentence. The contrarian point worth stating directly: for most retailers in the $50M-$500M range, the build column should be nearly empty in 2026. Proprietary data at that revenue scale is often thinner than it appears, loyalty programs with 18 months of data and 30% active member rates do not constitute a moat. Buying a proven personalization platform with cross-retailer training data will outperform a custom build for at least the next three years. The build case becomes real at scale, with 5+ years of clean transaction data and a team with the engineering capacity to maintain a production model. Commit to that sequence honestly, and your AI stack becomes an asset on the board slide rather than a liability.

Key Takeaways

Six actions for the CMO who has a board meeting in the next 90 days:

• Run a utilization audit before any new AI purchase.

  The average enterprise uses fewer than 40% of its contracted martech tools actively (Gartner, 2025). Identify every tool below 40% active utilization and assign a 60-day remediation or kill decision.

• Buy personalization, content generation, and demand forecasting.

  These are commercially solved with benchmarked ROI. AI personalization delivers 10-15% revenue uplift (McKinsey, 2025). Vendor platforms with cross-retailer data reach that benchmark faster than any internal build at the $50M-$500M revenue scale.

• Build only on top of 5+ years of clean, proprietary transaction data.

  70% of enterprise AI use cases are adequately served by off-the-shelf solutions (McKinsey, 2025). Restrict the build column to LTV modeling, category-specific intelligence, or closed-loop attribution where your data is demonstrably unique.

• Kill anything that cannot answer the revenue moment question.

  Score every tool: what specific revenue decision breaks if this disappears tomorrow? Tools that require indirect-influence arguments to justify their existence are overhead. Vendor AI bundles with 10-20% AI price premiums on unused features are the first kill candidates (SaaS Management Index, 2026).

• Fix the data layer before the next tool purchase.

  Data fragmentation is the top blocker to AI marketing ROI across Gartner and eMarketer research (Gartner, 2025). 60% of retail AI personalization failures trace to poor data integration (Workinsiders, 2025). No AI platform will hit its benchmark on a broken identity graph.

• Document a performance timeline for every AI investment before the board meeting.

  42% of companies abandoned most AI initiatives in 2025, up 147% from the prior year (S&P Global, 2025). Many of those kills were reactive, not strategic. A documented ROI timeline with a defined decision gate is the difference between a strategic kill and a board-mandated one.