Every retail business has heard the pitch. AI will personalise the experience, optimise the inventory, automate the customer service. Most have a pilot or two running. Few have something in production that the team actually uses every day.
The gap is not technology. The gap is pattern selection. Some AI use cases in retail are mature, well understood, and ship in weeks. Others depend on data you do not have, integrations you would have to build first, or a behaviour change in your team that is harder than the AI itself.
Here is what we have seen actually ships in Indian retail businesses, across eight specific patterns. The order matters. The first three ship in weeks. The next three ship in months. The last two often never land at all, for reasons worth understanding before you start.
1. Demand forecasting at SKU and store level
The single highest-leverage AI pattern in retail, and almost always the one to start with. The data already lives in your POS and ERP. The model is well understood. The decision it informs (replenishment) is one your team makes every day.
Typical impact across the businesses we have built this for: 15 to 25 percent reduction in out-of-stock incidents, 8 to 12 percent reduction in working capital tied up in slow-moving inventory. The math compounds because better forecast leads to better promotion planning leads to better assortment decisions.
What goes wrong: teams try to build a perfect model before they have any model. Ship a 70 percent accurate forecast in week two and improve it. Do not wait six months for 90 percent.
2. Customer service AI on WhatsApp, email, and web chat
The other ships-in-weeks pattern. A single AI agent handles the routine customer questions, the order status checks, the return requests, the basic complaints. It escalates the rest to a human.
The leverage is enormous because customer service in retail is mostly the same five questions in different words. An agent that answers four of those five questions reliably reduces your support team's load by 60 to 70 percent. The team then handles the genuinely complex cases properly instead of drowning in volume.
What goes wrong: brands try to make the agent handle everything. Better to handle 80 percent of the queries 95 percent correctly than 100 percent of the queries 70 percent correctly. The 5 percent error rate at scale becomes the brand reputation issue, not the cost saving.
3. Order intake automation
For brands that sell through distributors, retailers, or modern trade, a lot of orders still arrive as email, WhatsApp messages, phone calls, and PDFs. An AI agent reads the order in any format, validates it against price lists and stock, writes it into the ERP, and confirms back to the customer. The same workflow that used to take a sales rep two hours of admin a day takes the agent two minutes.
This is a low-glamour pattern, but the team time it gives back lets the sales reps actually sell. ROI shows up as faster order cycles, fewer fulfilment errors, and a noticeable drop in the time between "customer wants to order" and "order is in the dispatch queue".
4. Promotion ROI measurement
Covered in detail in our piece on trade promotion ROI. Ships in months because the data joining is harder than people expect, and because the team has to change how it briefs schemes for the analysis to work.
This pattern is worth doing because the size of the prize is large. Most brands recover 20 to 40 percent of their trade spend by stopping the schemes that do not pay back.
5. Assortment intelligence
What to stock where, at what depth, with what range. The right model uses sales data, demographics, store footfall, and the data of comparable stores to recommend assortment changes. Done well, it reduces the long tail of SKUs that never sell at a given store while protecting the head of the tail that drives the visit.
Months to ship because the data shape is harder, the team change is bigger, and the success metric (margin per square foot, not just sales) is something most retailers do not measure consistently yet.
6. Personalisation for repeat customers
Cohort the customers by behaviour, predict their next purchase, time the offer right, measure the lift. This works well in DTC and online retail, less well in offline retail where you do not always know who the customer is.
Months to ship because the customer ID has to be solid before anything else works, and most brands do not have that yet.
7. Planogram compliance via computer vision
Now we are in the often-a-year category. A camera in the store takes pictures of the shelf, the model identifies which SKUs are where, compares to the planogram, and flags compliance issues. The promise is large. The delivery is hard because: the camera install is operational change, the model needs significant fine-tuning per category, and the closed-loop action (a rep being notified and fixing the shelf) requires field force discipline that does not always exist.
Worth doing if your category is high-value and the planogram non-compliance is costing measurable share. Not worth doing if you are looking for quick wins.
8. Voice-to-data capture for field teams
Field reps speak voice notes in any local language, the system transcribes, structures the data, and writes it into your operational database. Beautiful idea. Slow to land because language coverage is uneven across Indian languages, accents vary widely, and the value of the data depends on the rep's discipline to capture consistently.
We have shipped this. It works when the rep population is small enough to train, when the language coverage is matched to the workforce, and when there is a manager actually reading the resulting data weekly. Not a starter project.
The pattern that lands first, almost every time
Demand forecasting. The data is already in your systems, the team already makes the decision it informs, the impact is measurable, the model is well understood, and the risk of looking dumb is low. Start there. Ship the forecast in six weeks. Use it to inform replenishment for one product line, one region. Measure the impact against the prior method. Once that is working, move to the next pattern.
Build or buy
For each pattern, the build-versus-buy decision matters. The shorthand is:
If the pattern is industry generic, buy a platform and add the integration layer. Customer service AI is a good example. Many vendors offer the model, the issue is wiring it into your stack.
If the pattern is specific to your data shape, build. Demand forecasting at the right level of granularity, for example, depends on your SKU hierarchy and your store taxonomy, which are unique to your business. A platform vendor will sell you a generic model that is 70 percent as good. A bespoke build with a senior team is 90 percent as good, and the integration is yours forever.
The integration layer is always custom. No SaaS reads your ERP, your retailer feed, and your finance system out of the box. The integration is where most projects fail, regardless of the underlying AI.
What to do this quarter
Pick demand forecasting or customer service AI. Run it as a six-week proof on one product line or one customer touchpoint. Measure against a baseline you can defend in a finance review. If it works, expand to the next product line, the next touchpoint, the next region. If it does not work, the cost was six weeks and you learned what is missing in your data or your team setup.
If you want help picking the pattern and shipping the six-week proof, talk to us. We have done it many times.