Walk into almost any retail AI conversation and you will hear the same opening move. The CFO asks for the software cost. The CDO asks for the roadmap. Everyone circles the investment as if it were the thing that carries risk. That reflex is the mistake, and it is remarkably consistent across merchandising, retail media, and the boardroom itself.
The pattern underneath is simple and expensive: retailers price what a decision costs to make while pretending the current state costs nothing to keep. The status quo never shows up on a slide, so it never gets defended, and it never gets killed. It just runs, quietly, quarter after quarter, in markdowns, stockouts, decayed data, and talent stuck doing work an agent should be doing at 3am. This essay is about that hidden bill, why it stays hidden, what it actually costs, and the operating model that finally puts it on the P&L where it belongs.
The pattern: everyone prices the move, nobody prices standing still
Three very different situations tell the same story.
In merchandising, a McKinsey survey this year found that 71 percent of merchants say AI merchandising tools have had limited to no effect on their business so far. Read in a boardroom, that number pushes leaders toward wait-and-see, as if the category itself were unproven. It measures something narrower than that. It measures deployment discipline, because most of those merchants bolted a tool onto dirty product data and called it a strategy. The instinct to wait treats the investment as the risk and the delay as safe. The delay is where the loss lives.
In retail media, Walmart generated $6.4 billion in advertising revenue, up 46 percent year over year, and the C-suite reflex is to stand up a media network and start selling sponsored search. That copies the output and skips the input. What actually differentiates Walmart sits one layer below the revenue line, in the decision to treat first-party data as a governed product rather than as exhaust from moving inventory. Retailers who benchmark their media network against Walmart are pricing the wrong asset while their own shopper signal decays in a dozen unjoined systems.
In the boardroom, Deloitte described the dynamic plainly this spring: capital markets are rewarding retail and consumer products companies for AI announcements even though company financials do not yet reflect the impact, which means you are being paid for the press release ahead of the proof.. That is the most dangerous incentive structure in the sector right now, because it lets a CIO spend real budget, publish a real headline, and never once quantify what the delay behind the headline is costing.
Same disease in three organs. The investment gets scrutinized to the decimal. The status quo gets a free pass. I made the merchandising version of this concrete in a full worked example for a Tier 2 specialty retailer, and the math is worth sitting with because it generalizes.
What it actually costs: three leaks and a decaying asset
Take that Tier 2 specialty retailer: roughly $800M in annual revenue, around 40,000 active SKUs, eight buyers, and assortment, pricing, and promotion still running on spreadsheets, vendor emails, and merchant intuition. You can model the cost of leaving that untouched with three leaks that any skeptical CFO will accept.
The first leak is excess markdown. This retailer takes markdowns on roughly 30 percent of units. Assume a conservative 1.5 point margin recovery from better-timed assortment and price actions, and on $800M you are looking at $8M to $10M annually left on the table. The second leak is stockout-driven lost sales, where manual reorder cycles let best sellers go dark for days; even a modest 0.5 percent revenue recovery from tighter inventory sensing is $4M of demand that walked to a competitor. The third leak is buyer hours, eight people spending half their week building reports instead of negotiating vendors and shaping assortment. Add only the defensible portion of those three and you clear $14M a year. The agentic program that closes them, done with discipline, costs a fraction of that. The investment decision was never close. The spreadsheet simply hid the losing side.
That $14M is the point. It is a bill you are already paying, and the Cost of Doing Nothing model just makes it visible.. The reason it stays invisible is structural: nobody in the organization is forced to book it, so nobody does.
The data version of the same cost is quieter and compounds faster. When shopper data lives as a byproduct of transactions rather than as a joined, governed, production-ready product, the signal decays every quarter while a competitor who built the plumbing compounds theirs. You can watch that compounding in the market share of retail media itself. Amazon Ads commanded 79.7 percent of the U.S. retail media market in 2025, Walmart Connect ranked second at 8.0 percent, and Target Roundel sat at 1.5 percent. In 2026, Amazon and Walmart are forecast to capture 89 percent of incremental retail media spending for the year. Two companies taking eighty-nine cents of every new dollar is a story about better data infrastructure, built earlier, while everyone else treated the asset as exhaust.. I unpacked why that gap is really a data-layer gap in the Walmart analysis, and the lesson for a boardroom is that the leak and the moat are the same asset viewed from two directions.
Why the bill stays hidden
Three forces keep the status quo off the ledger, and each one is fixable.
The first is accounting convention. A P&L books what you spend, not what you forfeit. Markdowns show up as margin pressure with a hundred plausible explanations, and none of them get labeled “the assortment decision we did not automate.” Lost sales never appear at all, because you cannot invoice demand that walked out the door. The cost is real and the ledger has no line for it, so the organization behaves as though it does not exist.
The second is the reward structure. When capital markets pay for announcements, the rational move for a CIO is to write the optimism memo: open with agentic AI, cite a cost-reduction range, list pilots in merchandising and contact centers, ask for more budget. The board nods, the release publishes, and the stock might tick. Then the variance report lands, and the conversion problem surfaces. The average enterprise now runs 14 AI projects simultaneously, up from 8 in 2023, and most organizations report that fewer than half are delivering measurable business value. Global enterprise AI spending is projected to reach $665 billion in 2026, while three out of four AI deployments fail to achieve their projected return on investment. The spending is happening. The conversion is not, and the reward structure rewards you for the spending anyway.
The third force is that the technology genuinely works now, which paradoxically makes bad deployment cheaper to hide. In 2026, always-on agentic systems can autonomously detect trends, competitive moves, pricing anomalies, inventory issues, and customer sentiment, and can reprice SKUs, rebalance inventory, or fine-tune promotions without waiting for human batch cycles. A merchandising agent is only as good as the product information underneath it, though, so if silhouettes, materials, and size ranges are inconsistent, the agent inherits your mess and amplifies it at machine speed. The 71 percent who saw limited effect are living proof. The capability arrived, the foundation did not, and the failure reads as a technology verdict when it is a data verdict.
The operating model that reverses it
Fixing this is a discipline, and it has four moving parts that work across merchandising, media, and the board memo alike..
Start by pricing inaction directly. Build the Cost of Doing Nothing as a line item, not a rhetorical device. Quantify the margin you forfeit by waiting and put a dollar figure on it. IMRG data shows a 2.3 percentage point market share loss over 24 months for retailers without AI personalization; translate that to your revenue base and the conversation changes, because wait-and-see becomes the visibly expensive option. Most retail organizations sit at Stage 2 of AI maturity, with proven pilot ROI but no business case for enterprise investment. CODN is what closes that gap, since it reframes the enterprise investment against a status quo that finally has a number attached.
Second, treat first-party data as a product with the same seriousness. When data is a product, you can sell off it more than once, and ads are only the first SKU. The same joined asset powers measurement, personalization, supplier collaboration, and agents. Walmart’s own trajectory makes the sequence explicit: advertising and membership income together accounted for roughly one-third of operating profit in Q4 FY2026, and the company is now testing ads inside its AI chatbot, Sparky, which executives say is already influencing higher basket sizes. An agent that recommends and converts inside a conversation only works if the data underneath is clean and real-time. So the question for a 2027 plan is sharp and testable: is your first-party data governed, joined, and production-ready, or is it still a quarterly export someone cleans by hand? The retailers auditing product attributes before they deploy are answering that question with action. On April 14, 2026, David’s Bridal joined Shopify’s Agentic Storefronts for ChatGPT and Microsoft Copilot while auditing attributes such as silhouette, neckline, fabric, and size range, so its assortment shows up when shoppers research through AI. Fix the foundation, then deploy the agents.
Third, self-fund the staircase instead of asking for a moonshot. Retail P&Ls cannot absorb a single large capital commitment, and they do not have to. Start with quick-win use cases that break even fast, then reinvest the returns into larger initiatives. The sequencing is knowable: recommendations and chatbots break even in 8 to 12 weeks, demand forecasting and inventory optimization reach payback in 4 to 6 months, and dynamic pricing requires 6 to 9 months because of governance setup and model tuning. Each stage funds the next, which turns AI from a bet the CFO fears into a ladder the CFO can climb.
Fourth, put a name on every dollar and a ceiling on every workflow. This is where memos usually go quiet and where the money leaks. Uncontrolled usage, nicknamed tokenmaxxing, has produced real overruns, including one company that reportedly spent $500 million in a single month after failing to set usage limits. Model your two or three production use cases with full cost accounting, including data infrastructure as a shared cost, implementation, ongoing MLOps, change management, and governance. A memo that hides the MLOps and governance lines is fiction. Cross-functional steering committees that identify, prioritize, and align use cases to enterprise goals are emerging as the mechanism that makes this stick. The failure rate is a governance story, and governance is what closes the gap between spending and value.
I laid out this exact four-part structure as the board memo every retail CIO should write before the budget locks, and the framing matters: the memo is a positioning statement that says we measure inaction, we ship to production, we self-fund, and we own the spend. Read it alongside the CODN worked example and the shape of the operating model is clear.
The window is closing on a knowable clock
The deadline is not abstract. By 2030, analysts project that 25 percent of global e-commerce sales will be enabled by AI agents, and 55 percent of digital consumers will begin product research on large language model platforms. Your assortment either shows up in that flow or it does not exist for those shoppers, and the data foundation that gets you into the flow takes quarters to build, not weeks.
The organizations pulling ahead already priced their status quo, productized their data, laddered their funding, and named their owners. Everyone else is still treating inaction as free while paying for it in full. The next four quarters will separate the two groups cleanly, because ambition is free in 2026 and announcements are cheap, while governed value is the only durable moat.
So build the CODN model this week, before the 2026 budget locks and before the analysts ask why the P&L still has not moved. Once the status quo has a number, the decision tends to make itself.