AI is genuinely useful in exactly three places in a distribution business, and oversold in about ten others. Here is how to tell them apart.
Strip out the marketing and AI earns its place in a distribution business in three specific spots.
Understanding what a retailer wrote. A message like "5 peti parle g bhejo kal tak" mixes languages, scripts, trade units and brand shorthand. This is the one place where rules genuinely run out — the variation is unbounded, and every attempt to enumerate it fails. It is the strongest case for a model in the entire stack, and it is covered in depth in why your commerce chatbot needs to understand Hinglish.
Ranking who to chase, and when. Given payment history, order frequency and seasonality, a model can order your receivables list by likelihood of going late. On 40 accounts a good collections person does this in their head. On 400 they cannot, and the ranking is worth having.
Suggesting the next SKU. Which line is this retailer most likely to add, given what comparable shops nearby stock? This is a recommendation problem with a clean signal, and it maps directly onto the range-selling gap described in fixing the four leaks in a beat.
Notice what these have in common: high variation, high volume, and a human who cannot hold all the data in their head. That is the test.
Anything with a stable rule. Applying a scheme, checking a credit limit, calculating GST, deciding whether stock exists. These are deterministic. A rule is faster, cheaper, auditable and, when it breaks, fixable in an afternoon. Putting a model here buys you unpredictability.
Anything you cannot measure. "AI-powered insights" that nobody acts on is a subscription, not a system.
Anything where you lack the data. This is the big one. A model that predicts lost orders needs a record of lost orders. Most distributors have never recorded one — the salesman simply did not bill that shop, and no row exists anywhere saying why.
Sequence matters more than model choice.
Most distributors want to start at step 4 because it sounds most like AI. Step 4 built on step 1's absence produces confident nonsense.
Numbers to re-run with your own, not a benchmark.
Take an office where two people spend a combined three hours a day converting WhatsApp messages into orders in Tally. At ₹25,000 a month each across 26 days and 8-hour days, three hours is roughly ₹1,800 a day, or about ₹47,000 a month of salaried time on retyping.
Automated parsing will not remove all of it. Assume it handles the two-thirds of messages that are straightforward and leaves the ambiguous third for a human. That is roughly ₹31,000 a month of time returned — against whatever the system costs you.
The part that does not show up in that calculation, and usually matters more: orders that arrive after office hours get processed immediately instead of the next morning, so the delivery makes the same day's route. Whether that is worth more than the salary saving depends on your route economics — the Delivery Cost Calculator is a starting point for putting a number on it.
Three questions that separate real systems from demos.
"What happens when it is wrong?" Every system misreads orders. The good ones confirm before acting, show the resolved quantity in pieces, and make correction a one-word reply. The bad ones fail silently and ship the wrong carton.
"What data does it need from me, and do I have it?" If the answer requires twelve months of clean sell-out data you have never collected, the honest timeline is a year longer than the pitch.
"What does it do on day one, with no history?" Cold-start behaviour tells you whether the product was built for a real distributor or for a demo with seeded data.
For most distribution businesses, the highest-return technology project in year one is not AI. It is capturing orders and payments cleanly and automating the deterministic messaging around them — the work described in three WhatsApp workflows that move money.
That work is unglamorous, it pays immediately, and it produces exactly the dataset that makes the AI step worth doing in year two. Skipping it does not accelerate the AI; it just means the model has nothing good to learn from.
Every AI application in distribution depends on data most distributors have never captured. Worth being concrete about what each needs.
| Application | Data required | Do you have it? |
|---|---|---|
| Order parsing | Real retailer messages | Yes — in WhatsApp |
| Collections ranking | Invoice dates, payment dates per retailer | Usually, in the ledger |
| Reorder prediction | Order history by SKU by outlet | Usually, in billing |
| Lost-order analysis | Why an outlet was not billed | Almost never |
| Demand forecasting | Sell-out, not just sell-in | Almost never |
The pattern is clear. The applications that work are the ones running on data your billing system already produces as a by-product. The ones that disappoint need data somebody would have had to deliberately capture, and nobody did.
That is why "start capturing clean data" is not a stalling tactic — it is the actual first step. A year of order and payment history with timestamps is worth more than any model you could buy today.
Months 1–3: capture. Orders with timestamps, SKUs and quantities. Payments dated against invoices. Outlets visited, not just outlets billed — that last one is the field almost nobody records and it unlocks the coverage analysis in fixing the four leaks in a beat.
Months 4–6: automate the deterministic. Invoice delivery, payment reminders, balance lookups. No AI. Immediate return, and it generates more clean data as a side effect.
Months 7–9: add parsing. Free-form order messages into structured orders, with confirmation before anything is acted on. This is the highest-value AI step for most distributors.
Months 10–12: add ranking. Collections priority and reorder prediction, now trained on nine months of your own clean data rather than on assumptions.
Most distributors want to start at month ten. Started there, with months one to three skipped, the model has nothing good to learn from and produces confident nonsense that is worse than no model at all.
Three questions, and if a proposal fails any of them the answer is not yet.
Is the variation genuinely unbounded? If you could write the rules down given a week, write the rules down. They are faster, cheaper, auditable, and fixable in an afternoon when they break.
Does the data already exist as a by-product? If it needs somebody to start capturing something new, that capture is the project — and it has to run for months before the model has anything to learn from.
Can you tell when it is wrong? A system whose errors are invisible will accumulate them silently. If you cannot describe how a failure would surface, you are not ready to deploy it.
Order parsing passes all three for most distributors. Demand forecasting usually fails the second. Anything described as "insights" usually fails the third.
Worth noting that some of this is not a future decision. Two places it is already routine in Indian distribution:
Speech recognition on voice notes. A meaningful share of retailer messages arrive as voice, and transcription is now good enough to be useful — with the confirmation step described in why your commerce chatbot needs to understand Hinglish.
Fuzzy matching on product names. Turning "surf ex" into a specific SKU is a matching problem that classical techniques handle well, and it does not need a large model at all.
Both are narrow, both have obvious success criteria, and neither requires you to restructure anything. That is the profile of AI worth adopting: it disappears into a workflow rather than becoming one.
If you take one action from this: start recording why an outlet was not billed.
It is one field, filled by the salesman, from a short list — shop shut, no requirement, credit block, price objection, competitor scheme. Nobody captures it, which is why nobody can analyse lost sales.
In six months it is the most valuable dataset you own, and it costs nothing to start today.
Beyond the three earlier, one more separates real products from demos:
"Show me what it does with my last week of WhatsApp messages."
Not a curated demo. Your actual messages, with your retailers' spellings, your SKU names, your units. A system built for real Indian distribution will handle most of it and fail gracefully on the rest. A system built for a pitch will fail in ways the vendor has never seen, and their reaction to that failure tells you everything.
The three questions worth asking before any AI project.
You need it for the parsing step, if your retailers send free-form messages in mixed language. You do not need it for the rest — matching a parsed SKU to your catalogue, checking stock, applying a scheme and creating the voucher are all deterministic. Use AI for the messy input and rules for everything downstream, because rules are debuggable and AI is not.
It can rank accounts by risk using payment history, order patterns and seasonality, and that ranking is usually better than intuition on a large base. What it cannot do is know that a shop is closing because of a family dispute. Treat the output as a prioritised list for human follow-up, not as a credit decision.
Being applied to a problem that was really a data problem. If nobody records why an order was lost, no model can tell you. Most distributors get more value in year one from capturing clean order and payment data than from any model trained on what they currently have.
The FlowKartAI team builds WhatsApp-native ordering for Indian B2B distributors and the kirana stores they serve. We write about distribution economics, GST compliance, and the practical side of putting AI in front of retailers who have never opened an app.
FlowKartAI parses natural language WhatsApp messages into ERP-ready orders in seconds.
मुफ़्त डेमो आज़माएं