AI for Retail: A Founder's P&L Playbook

2026-05-29 · Tommaso Maria Ricci

Why AI for Retail Stopped Being a Science Project and Became a P&L Decision

I have spent more than twenty years building companies, and I have watched plenty of technology waves crest and crash from the operator chair, not the analyst desk. AI for retail is the first wave I have seen that pays for itself inside a single quarter when it is deployed with discipline, and quietly burns cash for eighteen months when it is treated as an innovation showcase. That gap, between the disciplined and the decorative, is the whole story.

I run businesses out of Miami. I have deployed machine learning against real inventory, real shrink, real customers who abandon carts and walk out of stores. So this is not a survey of vendor decks. It is what actually moves margin, what wastes money, and how a retail executive should sequence the work so the board sees results before the patience runs out.

Let me be blunt about the stakes. McKinsey has estimated that generative AI alone could add somewhere in the range of 240 to 390 billion dollars in value annually to the retail and consumer goods sector, on top of the value classic analytics already delivers. Most retailers will capture almost none of it, not because the technology fails, but because the organization never gets past the pilot.

What AI for Retail Actually Means When You Strip Out the Hype

When people say AI for retail, they usually picture a chatbot. That is the least interesting part. The money sits in the unglamorous machinery behind the storefront: how much you order, how you price it, how you place it, how you prevent it from walking out the door, and how you keep the customer coming back.

Here is the mental model I use. Retail is a sequence of forecasts. You forecast demand, you forecast price elasticity, you forecast which customer wants which product, you forecast when a shelf will go empty, you forecast which employee should work which shift. Every one of those forecasts is currently being made, today, by someone using a spreadsheet and a gut feeling. AI does not introduce forecasting into retail. It just makes the forecasts an order of magnitude better and removes the human bottleneck.

The retailers winning with AI share one trait: they picked forecasts that were measurably bad and replaced them with forecasts that were measurably better. They did not chase novelty. They chased error reduction in places where error costs real money.

A practical way to frame it for your team:

Where are we guessing? Demand, price, placement, staffing, fraud.
What does each wrong guess cost? Stockouts, markdowns, lost baskets, shrink, overtime.
Which guess, if improved by twenty percent, moves the P&L most?

Answer those three questions honestly and you have your AI roadmap. Everything else in this article is detail on how to execute against that roadmap without lighting money on fire. If you want the broader operating logic before you go vertical, my practical guide to AI for small business lays out the same discipline at a smaller scale.

Demand Forecasting and Inventory Optimization: The Quiet Engine

If you do only one AI project in retail, do this one. Demand forecasting is where AI for retail proves its worth fastest, because the baseline is almost always terrible and the cost of being wrong is brutal on both sides.

Think about what bad forecasting costs you. Overstock ties up cash, fills warehouse space, and eventually gets marked down to clear, destroying margin. Understock empties shelves, sends customers to a competitor, and the lost sale never shows up in any report because it never happened. Retailers routinely run inventory accuracy below seventy percent in stores, and industry studies have long pegged the combined cost of overstock and out-of-stock conditions at well over a trillion dollars globally per year.

Classic forecasting uses moving averages and seasonality curves. It treats every SKU as an island and assumes next week looks like last year. Machine learning forecasting does something fundamentally different:

It learns cross-product relationships. When umbrellas sell, so do certain jackets. The model finds these without anyone programming them.
It absorbs external signals. Weather, local events, paydays, school calendars, competitor promotions, even pollen counts for pharmacies.
It forecasts at the store-SKU-day grain, not the regional-category-month grain, so replenishment becomes surgical.

What does this look like in numbers? In my experience and consistent with what McKinsey and Gartner have published, well-executed AI demand forecasting cuts forecast error by twenty to fifty percent, reduces inventory holding by ten to twenty percent, and lifts product availability by a few points. A few points of availability on a high-velocity category is not a rounding error. It is the difference between hitting the quarter and missing it.

The trap to avoid: do not forecast everything at once. Start with your fastest-moving, highest-margin categories where data is dense. Slow-moving long-tail SKUs have sparse data and the models struggle. Win on the head of the curve first.

Dynamic Pricing and Markdown Optimization: Where Margin Hides

Pricing is the single highest-leverage decision in retail and the one most often made by feel. A one percent improvement in price realization typically drops almost entirely to the bottom line, because there is no incremental cost. There is no other lever in the business with that property.

AI changes pricing in two distinct ways, and you should treat them as separate projects.

1. Base price and elasticity. The model estimates how sensitive demand is to price for each product, segment, and channel. It finds the products where you are leaving money on the table because customers would happily pay more, and the products where a small cut drives outsized volume. Most retailers discover they have been underpricing their loyal-customer staples and overpricing their price-sensitive traffic drivers.

2. Markdown optimization. This is the one I push hardest, because the upside is immediate and visible. Every retailer with seasonal or perishable goods runs markdowns, and almost all of them mark down too late, too uniformly, and too steeply. AI markdown engines answer a precise question: given current sell-through, remaining inventory, days until end-of-season, and demand decay, what is the exact discount on this exact SKU in this exact store today that maximizes total recovered margin?

The results are consistent. Retailers deploying AI markdown optimization typically recover several points of gross margin on clearance inventory and sell through faster, freeing floor space and cash. Gartner and HBR have documented gross margin improvements in the low-to-mid single digits from disciplined price and promotion optimization, and on a thin-margin business that is enormous.

A word of caution that most vendors will not give you. Dynamic pricing without guardrails destroys trust. If a customer sees a price swing wildly hour to hour, or feels punished for loyalty, you lose more than you gain. Constrain the model. Set floors, ceilings, and fairness rules. Never let the algorithm optimize a single transaction at the expense of the relationship. The retailers that got burned on dynamic pricing got burned because they removed the human judgment instead of augmenting it.

If pricing intelligence sits inside a broader revenue motion for you, the logic connects directly to my AI for sales guide, because price and conversion are two sides of the same equation.

A focused preliminary assessment, two weeks of looking at your actual data, will usually surface exactly which categories are bleeding margin through lazy pricing, and that alone tends to justify the entire engagement.

Personalization and Recommendation Engines Online and In-Store

Personalization is the topic everyone thinks they understand and almost nobody executes well. The famous statistic, that recommendation engines drive a large share of revenue at the platform giants, is true and also misleading, because it makes mid-size retailers think they need the same infrastructure. They do not.

Let me separate the channels, because they behave very differently.

Online personalization is mature and accessible. A modern recommendation engine does three jobs: it ranks the products a given visitor is most likely to buy, it personalizes search results so the same query returns different products for different people, and it sequences merchandising so the homepage and category pages reorder themselves per visitor. The lift is real. Well-tuned personalization commonly adds ten to thirty percent to conversion and meaningfully raises average order value. The sports retailer case I keep in mind grew sales roughly thirty percent largely on the back of AI-driven marketing and personalization working together, not personalization in isolation.

In-store personalization is the frontier and the harder problem. You do not have a login, a cart, or a click trail. What you have is the loyalty card, the receipt history, the app, and increasingly the store associate with a tablet. The winning pattern is not creepy facial tracking. It is arming the associate and the loyalty app with the same recommendation intelligence the website uses, so a known customer gets relevant offers and the associate knows what this person buys before the conversation starts.

The connective tissue across both channels is a clean customer data foundation. If your online behavior, loyalty history, and in-store purchases live in three disconnected systems, no recommendation engine can help you. The model is only as good as the identity resolution underneath it. I go deeper on the marketing side of this in my AI marketing strategy frameworks and tools breakdown.

Computer Vision in Stores: Shrink, Planograms, and Frictionless Checkout

This is where AI for retail gets physical, and where I see the most overspending on the wrong things. Computer vision in stores is powerful, but it is also where vendors sell hardware-heavy fantasies. Let me separate what works today from what is still a demo.

Loss prevention and shrink. Shrink is a brutal, growing problem. The NRF has reported that retail shrink represents over one hundred billion dollars in annual losses for the industry, and theft, both external and internal, is the largest component. Computer vision attacks this at the point of sale and on the floor. Camera systems detect scan-avoidance at self-checkout, flag suspicious ticket-switching, and identify known shoplifting patterns. The honest truth: these systems generate real ROI when focused narrowly on self-checkout loss, where the math is clean and the events are frequent. Broad store-wide surveillance AI is far harder to justify on ROI alone.

Planogram compliance. This one is underrated and quietly excellent. A planogram is your shelf plan, and the dirty secret of retail is that shelves drift out of compliance constantly. Products go to the wrong slot, facings collapse, out-of-stocks go unnoticed. Computer vision, whether from fixed cameras, associate phone photos, or shelf-scanning robots, checks actual shelves against the plan and flags gaps. Brands pay for premium placement, and if you cannot prove compliance you cannot defend the trade dollars. The ROI here is both lost-sales recovery from caught out-of-stocks and protected trade revenue.

Frictionless and computer-vision checkout. The grab-and-go, walk-out-without-scanning concept is real but capital-intensive and operationally fragile at scale. My counsel to most retailers: do not lead here. Let the giants subsidize the maturation of this technology. Where vision-assisted checkout pays off sooner is in accelerating, not eliminating, the checkout, catching scan errors and produce misidentification at staffed and self-checkout lanes.

The pattern across all three: start where the loss event is frequent and the camera angle is controlled. Self-checkout and shelf gaps qualify. Open-store theft detection rarely does on day one.

Supply Chain and Replenishment: From Forecast to Shelf

A demand forecast is worthless if it does not flow into action. This is the gap that kills most retail AI value. The model predicts beautifully, and then a human manually places the same orders they always did. Supply chain AI closes the loop between prediction and execution.

The components that matter:

Automated replenishment. The forecast drives purchase orders and store transfers automatically, within guardrails, so inventory positions itself ahead of demand instead of reacting to stockouts.
Lead-time prediction. Suppliers are unreliable in different, predictable ways. AI learns each supplier's true lead-time distribution and buffers accordingly, instead of using one static safety-stock number for everyone.
Allocation and rebalancing. When a SKU sells out in one store and sits dead in another, the model orchestrates transfers before you have to mark anything down.
Disruption response. When something breaks, a port delay, a weather event, the system re-plans across the network in hours, not in a week of frantic meetings.

The numbers that justify this are consistent with what Gartner and Deloitte have published on AI-enabled supply chains: meaningful reductions in logistics and inventory carrying costs, double-digit improvements in forecast-to-fulfillment accuracy, and a sharp drop in the manual planning hours your team spends firefighting.

The mistake I see constantly: retailers buy a forecasting tool and a separate replenishment tool from different vendors and they never integrate. The handoff between forecast and order is where the value lives. If those two systems do not talk, you have bought two expensive dashboards. For the deeper architecture of getting these pieces to work as one system, see my AI supply chain optimization guide.

Customer Service and Conversational Commerce

Now the chatbot, the part everyone starts with and the part I tell people to start last. Not because it lacks value, but because it is the easiest place to embarrass your brand and the hardest to get a clean ROI on if you lead with it.

Here is the honest assessment. Generative AI has genuinely transformed customer service in the last two years. The earlier generation of rigid, menu-driven bots was rightly hated. Large language models can now actually understand a messy customer question, pull the relevant order and policy data, and resolve a real issue. Forrester and McKinsey have both documented that AI can deflect a large share of routine contacts and cut average handle time substantially for the contacts that still reach a human.

Where it genuinely works in retail:

Order status, returns, and tracking, which are the bulk of contact volume and almost fully automatable.
Product discovery and conversational commerce, where a shopper describes what they want in plain language and the assistant narrows from ten thousand SKUs to three relevant ones.
Agent assist, which I rate higher than the customer-facing bot. The model drafts the response, surfaces the policy, and suggests the resolution while a human agent stays in control. Faster, safer, and customers cannot tell the difference except that the answer arrives quickly and correctly.

The non-negotiable rule: never deploy a customer-facing AI agent that can take an action, issue a refund, change an order, without tight constraints and an audit trail. The failure mode is not the bot saying something silly. It is the bot doing something expensive. My AI customer service business guide goes through the guardrails in detail.

If you are evaluating whether to let agents take autonomous actions at all, the broader question of agentic systems is worth understanding before you commit, which is why I wrote about what agentic AI is and how it works.

Marketing, Customer Lifetime Value, and Churn

This is where I have personally seen the most dramatic top-line results, and it deserves its own section because retailers systematically underinvest here relative to the supply-chain glamour projects.

The core shift is from campaign thinking to customer thinking. Traditional retail marketing blasts the same promotion to the whole list. AI marketing predicts, for each individual customer:

Their likelihood to purchase in the next window, so you spend acquisition and retention budget where it converts.
Their lifetime value, so you stop treating a one-time bargain hunter and a high-value loyalist the same way.
Their churn risk, so you intervene before they quietly disappear rather than after.
The next best offer, the specific product and incentive most likely to move this person now.

The sports retailer I mentioned grew sales roughly thirty percent, and the engine behind it was exactly this: predictive segmentation feeding personalized, well-timed marketing instead of generic blasts. On the hospitality side, I have seen a hotel move annual revenue from around nine million to ten million euro by applying the same logic, predicting which guests to target with which offer at which moment, rather than discounting indiscriminately.

The economics are compelling because retention is cheap relative to acquisition. HBR has long cited that increasing retention by a few points can lift profits substantially, and AI churn prediction is the most direct lever on retention I know. The model tells you the fifty customers about to leave this month who are worth saving, and your team acts on a list of fifty instead of a vague worry about the whole base.

This connects tightly to the personalization work above, because the same customer data foundation powers both. Build it once, monetize it across marketing and merchandising.

Store Operations and Workforce Scheduling

The least glamorous AI application in retail and one of the most reliably profitable. Labor is typically the largest controllable cost in a store, and it is almost universally scheduled badly.

The problem is structural. Managers build schedules from habit and last week's pattern, not from predicted traffic. The result is predictable: overstaffed dead hours that burn payroll, and understaffed rushes that lose sales and exhaust the team. AI workforce scheduling fixes this by forecasting foot traffic and transaction volume at the fifteen or thirty minute grain, then building schedules that match labor to demand while respecting availability, skills, labor law, and fairness.

The wins compound:

Payroll savings of several percent from cutting overstaffed hours.
Higher conversion during peaks because there are enough associates to actually sell.
Lower turnover, because predictable, fair schedules are one of the top drivers of frontline retention, and turnover is a massive hidden cost.

Beyond scheduling, store operations AI also handles task management, telling associates what to do and when based on real conditions, and increasingly uses the computer vision shelf data discussed earlier to trigger restocking tasks the moment a gap appears. The compounding theme across this whole article: the data you collect for one use case feeds the next. That is why sequencing matters, and why a coherent rollout beats a scatter of disconnected pilots. The discipline of running these projects to value, rather than to demo, is the subject of my AI implementation business practical framework.

The AI for Retail Readiness Scorecard: A Yes or No Self-Assessment

Before you spend a dollar, run your organization through this. Answer each question honestly with yes or no. This is the assessment I run mentally before I greenlight any retail AI project, and it predicts success better than any vendor demo.

Data foundation

Can you pull two or more years of clean transaction-level sales history per SKU per store?
Is your inventory data accurate enough that you trust the on-hand numbers?
Do online behavior, loyalty, and in-store purchases connect to a single customer identity?
Can a new data feed reach your analytics environment in days, not months?

Organization

Is there a single executive who owns the outcome, not just the technology?
Do the store and supply-chain teams who will use the output actually want it?
Are you prepared to change a process based on what the model says, even when it contradicts a senior person's instinct?
Can you free up at least one analyst or data person to own this internally?

Economics

Have you identified a specific, measurable cost the model is meant to reduce?
Have you agreed in advance how you will measure success, and against what baseline?

How to read your score:

Eight or more yes: you are ready. Pick the highest-value use case and move now.
Five to seven yes: you can start, but fix the no answers in parallel, especially anything in the data section.
Four or fewer yes: stop. Spend the next sixty days on data and ownership before any model. Buying AI now would waste money. Most failed retail AI projects fail right here, and no algorithm fixes a broken data foundation.

The most common pattern I see: companies score low on data, high on enthusiasm, and buy software anyway. That sequence has a near-perfect record of disappointment.

A Realistic 30-60-90 Day Implementation Roadmap with Budgets

Here is how I would actually sequence a first retail AI deployment, with budget ranges grounded in what mid-market retailers genuinely spend. These are real ranges, not vendor wish lists. Adjust up for enterprise scale and down for a single-region operator.

Days 1 to 30: Prove the data and pick the target.

Audit data availability against the scorecard above.
Choose one use case with a clear, measured cost to attack. For most retailers this is demand forecasting on top categories or markdown optimization.
Establish the baseline. Document current forecast error, current markdown recovery, current shrink rate, whatever you intend to improve.
Stand up a clean data pipeline for that one use case only.
Budget: 15,000 to 50,000 dollars, mostly internal time plus light data engineering. Resist buying platform software this month.

Days 31 to 60: Build and backtest.

Deploy the model, whether a vendor solution or a focused internal build, against historical data first.
Backtest relentlessly. Would the model have beaten your actual decisions over the last year? If it cannot win on history, it will not win live.
Run it in shadow mode alongside your current process. Compare its recommendations to what your team would have done.
Budget: 30,000 to 120,000 dollars, depending on build versus buy and category breadth. Vendor subscriptions in this range typically run 2,000 to 15,000 dollars per month for a focused module.

Days 61 to 90: Go live, narrow, and measure.

Turn the model loose on a controlled slice: a set of stores, a category, a customer segment.
Keep a control group running the old way. Without a control group you will never prove the value, and you will lose the next budget fight.
Measure against the baseline you set in week one. Report the delta in dollars, not in model accuracy metrics the board does not care about.
Budget: 20,000 to 80,000 dollars for rollout, integration, and change management.

All in, a serious first deployment lands somewhere between 75,000 and 250,000 dollars over ninety days for a mid-market retailer, and should return a multiple of that within the following two to three quarters if you picked the use case correctly. If your projected return does not clearly exceed the spend, you picked the wrong use case. Pick again. I treat ROI projection as a gate, not a formality, which is why I wrote a full AI ROI for business guide.

The Real Cost Breakdown Nobody Puts in the Proposal

Vendors quote you the license. The license is rarely the biggest number. Here is the honest, fully-loaded cost structure, because the hidden costs are what blow up budgets and credibility.

1. Software licenses. The visible cost. For a focused module, expect 2,000 to 15,000 dollars per month for mid-market, scaling to six figures monthly for enterprise platforms. Generative AI usage costs are usage-based and can surprise you on customer-service volume, so model them carefully.

2. Infrastructure. Cloud compute and storage for training and serving models. Often underestimated. For most retail use cases this is modest, a few thousand dollars a month, but computer vision with live video is the exception and can be the largest line item because of bandwidth and edge hardware.

3. Data. The silent budget killer. Data cleaning, integration, identity resolution, and ongoing pipeline maintenance routinely consume more than the software license, especially in year one. Budget for it explicitly or it will ambush you. This is the line where I have seen the most projects quietly double their cost.

4. People. You need at least one internal owner who understands both the business and the data. Whether that is a hire, a reallocation, or fractional support, it is not optional. A model with no internal owner becomes shelfware within two quarters. Plan for ongoing analyst time, not just a one-time setup.

5. Change management. The cost almost no proposal includes and the one that most often determines success. Training store managers, rebuilding processes, and overcoming the very human resistance to a machine second-guessing decisions. Underfund this and the technology sits unused while you keep paying for it.

A useful rule from the field: for every dollar of software license, plan one to two dollars across data, people, and change management. Proposals that show only the license are not telling you the real number. For a structured way to think about the whole adoption budget at the organizational level, see my enterprise AI adoption framework.

Common Mistakes That Kill Retail AI Pilots

I have watched more retail AI pilots die than succeed, and the autopsies all read the same. These are the killers, in rough order of how often they strike.

1. Starting with the shiny use case instead of the valuable one. Teams launch a flashy chatbot or a futuristic vision project because it demos well, while ignoring the boring forecasting work that would have paid for everything. Lead with value, not with novelty.

2. No baseline. If you did not measure the problem before deploying, you cannot prove improvement after. The project dies in the budget review because nobody can say what it earned. Measure first, always.

3. Pilot purgatory. The pilot works, everyone is pleased, and then it never scales because no one planned for integration, governance, or rollout. A pilot that cannot scale is a science project, not an investment. Plan the path to production before you start the pilot.

4. Broken data, ignored. Teams know the data is messy and proceed anyway, hoping the model will be smart enough to compensate. It will not. Garbage in, confident garbage out. Fix the data or do not start.

5. Removing the human entirely. The pricing disaster, the rogue refund bot, the customer-trust collapse. These come from replacing judgment instead of augmenting it. Keep a human in the loop wherever an action has real cost or real brand risk.

6. No internal owner. A vendor cannot own your outcome. If there is no one inside who lives with the results and tunes the system, it decays. Models are not appliances. They need an owner.

7. Optimizing a metric that hurts the business. The classic: a markdown model that maximizes sell-through by torching margin, or a personalization engine that boosts clicks while training customers to wait for discounts. Optimize for profit and lifetime value, not for vanity metrics.

Avoid these seven and you are already ahead of the large majority of retailers attempting this work. None of them are technology problems. All of them are discipline problems, which is good news, because discipline is free.

How to Choose a Vendor Without Getting Burned

The market is loud, and most of the noise comes from companies that rebranded as AI vendors in the last twenty-four months. Here is the filter I apply, and the questions that separate the real from the repackaged.

Demand proof on data like yours. Ask for results from a retailer of similar size, category, and data maturity. Not the enterprise flagship logo, a peer. If they can only show giants, their product may not fit your reality.

Insist on a backtest against your history before you sign. A serious vendor will run their model on your historical data and show whether it would have beaten your actual decisions. One that refuses is selling hope.

Interrogate integration honestly. How does this connect to your point-of-sale, your inventory system, your customer data? The demo is always smooth. The integration is where projects die. Get specifics, get timelines, get them in writing.

Understand the pricing model fully. Flat versus usage-based matters enormously, especially for anything generative. Model your real volume. Ask what happens to cost when you scale, and what the renewal looks like.

Check who owns the data and the model outputs. You should own your data and the insights derived from it. Read the contract. Some vendors quietly claim rights you would never agree to if you noticed.

Favor focus over breadth. A vendor that does demand forecasting brilliantly beats a suite that does ten things adequately. You can integrate best-of-breed. You cannot un-mediocre a do-everything platform.

Test their support before you need it. Run a real problem past their team during evaluation. The responsiveness you get while they are selling is the best it will ever be. Plan accordingly.

One more, from hard experience: be wary of any vendor that promises to remove humans from the loop. The good ones augment your team. The dangerous ones promise to replace your judgment, and that promise is exactly where the expensive failures originate.

The 24-Month Outlook for AI in Retail

I will be specific rather than vague, because vague predictions are useless for planning. Here is where I believe AI for retail is genuinely heading over the next two years, and where I think the hype is running ahead of reality.

Agentic systems move from demo to cautious production. The shift from AI that recommends to AI that acts, placing orders, adjusting prices, resolving service tickets end to end, is real and accelerating. Within twenty-four months, leading retailers will run autonomous replenishment and pricing agents within tight guardrails. The retailers who built clean data foundations now will be the only ones able to trust an agent to act on their behalf. This is the single biggest reason to fix your data this year.

Personalization collapses the online and in-store divide. The artificial wall between channels disappears as unified customer data and on-device intelligence let the store recognize and serve the known customer the way the website already does. The technology is arriving faster than most retailers' data foundations can absorb it, which will widen the gap between the prepared and the unprepared.

Computer vision gets cheaper and more practical. As models get more efficient and edge hardware drops in price, shelf monitoring and loss prevention become accessible to mid-market retailers, not just the giants. Frictionless checkout, by contrast, will remain a niche bet for most, subsidized by the few willing to absorb the operational complexity.

The generative service layer matures. Conversational commerce and agent-assist become standard, not differentiating. The advantage shifts from having a bot to having a bot grounded in clean, real-time product and customer data, which again points back to the data foundation.

The winners and losers separate decisively. This is my strongest conviction. The gap between retailers who built disciplined AI capability and those who ran disconnected pilots will become a competitive chasm over these two years, because the advantages compound. Better forecasts produce better data produce better models produce better forecasts. Start the flywheel now or watch competitors who started earlier pull permanently ahead.

If you take one thing from this outlook: the binding constraint over the next two years is not the technology. The models are good and getting better and cheaper without you. The constraint is your data foundation and your organizational discipline. That is the work, and it is entirely within your control.

For the executive framing of why this can no longer be delegated downward, I made the full argument in why every CEO needs an AI strategy, and the broader business case sits in my generative AI for business guide.

Where to Start, Practically

If you read all of this and feel the pull to do everything at once, resist it. The retailers who win do not deploy ten use cases. They deploy one, prove it in dollars, fund the next from those dollars, and compound. Sequencing beats ambition every time.

So here is the concrete first move. Run the readiness scorecard this week. If you score eight or higher, pick demand forecasting or markdown optimization, set a baseline, and start the ninety-day clock. If you score lower, spend sixty days on data and ownership before touching a single model, because that spend will return more than any algorithm you could buy today.

The technology is no longer the hard part. The hard part is choosing the right first fight, measuring it honestly, and having the discipline to scale what works and kill what does not. That is operator work, not vendor work, and it is exactly where the durable advantage in retail will be won.

Where the fastest returns hide is rarely where the noise is loudest, and a short, focused look at your own numbers will usually point straight at them. That diagnostic is the cheapest, highest-leverage step you can take before committing real budget, and it is the one I would take first.

For the wider industry context behind these shifts, three sources worth your time are the National Retail Federation research hub, the Deloitte retail and distribution insights center, and the analysis published by Harvard Business Review.