Nobody on your team can read a model architecture diagram. You do not need to. Five questions do the job instead.
Every AI vendor pitch looks similar: a dashboard, a chart trending upward, language like "insights" and "intelligent automation." None of that tells you whether the system will correctly read "2 dz surf excel bhej do jaldi" from a retailer who has never typed a full sentence in his life. That is the actual job, and it is the one thing the deck cannot show you.
You do not need a technical team to judge this. You need five questions, asked in order, and the willingness to walk away when a vendor cannot answer the first one plainly.
"Run this against my last week of actual WhatsApp order messages."
Not a curated example. Your retailers, your spellings, your SKU shorthand, your trade units. A system built for real Indian distribution will handle most of it and fail visibly on the rest — asking for clarification rather than guessing. A system built for a pitch will fail in ways the vendor has never anticipated, because their demo data was clean and yours is not.
The vendor's reaction to a failure matters as much as the failure itself. "Let us look at that and get back to you" is a normal, honest response to a genuine edge case. Silence, or blaming the retailer's message for being "unclear," is not.
Every order-parsing system misreads something eventually — a quantity, a SKU, a unit. The question is what happens next.
The systems worth using confirm the resolved order back to the operator before it becomes a voucher, in plain language, with the quantity in pieces rather than trade units that hide an error (a misread "2 case" versus "2 piece" is a 24x difference, not a rounding error). Correction should be a one-word reply, not a support ticket.
The systems to avoid ship the misread order straight into your ERP and let the delivery van sort out the discrepancy. If a vendor cannot describe their confirmation step clearly and immediately, assume there isn't one.
Some AI applications need data you already generate as a by-product of billing — order history, payment dates, SKU catalogues. Others need data almost nobody has ever captured deliberately — why an order was lost, what a retailer's true sell-out was, which competitor scheme pulled a sale away.
| Vendor claims to need | Do you likely have it? |
|---|---|
| Your WhatsApp order messages | Yes |
| Invoice and payment dates | Usually, in the ledger |
| Order history by SKU by outlet | Usually, in billing |
| Why specific orders were lost | Almost never |
| True sell-out per outlet | Almost never |
If the pitch depends on the bottom two rows and you do not have that history, the honest timeline is measured in the months it takes to start capturing it, not the weeks the sales deck implies. This is the same data gap covered in more depth in AI in B2B distribution and in reorder prediction versus a reorder point rule — most AI pitches quietly assume a dataset that does not exist yet.
Cold-start behaviour is the fastest way to tell a real product from a demo built on seeded data. Ask exactly what the system does on the first day, with no order history and no trained model behind it.
A system built for real distributors degrades gracefully — it might fall back to simple rules until enough history accumulates, and say so plainly. A system built for a pitch either has no answer, or quietly assumes history that will not exist for months, which means the actual results will lag the sales conversation by however long that history takes to build.
Before signing, ask what happens to your data — your catalogue, your order history, your retailer messages — if you cancel in six months. A vendor who cannot answer this cleanly, or whose contract makes export difficult, has built a system designed to make you stay rather than one designed to earn it.
This matters more for AI systems than ordinary software, because switching cost compounds: the longer a model has learned from your specific catalogue and retailers, the more painful an exit becomes if the relationship goes wrong later.
Numbers to re-run with your own, not a benchmark.
If two people spend a combined three hours a day retyping WhatsApp orders into your billing system, at a combined ₹25,000 a month each across 26 working days, that is roughly ₹47,000 a month of salaried time. A vendor's pitch usually claims all of it back. A more honest starting assumption — automated parsing handling the straightforward two-thirds and leaving the ambiguous third for a human, per the worked example in AI in B2B distribution — is closer to ₹31,000 a month of time returned. Compare that figure, not the vendor's larger one, against what the system costs, and use the ROI Calculator to see the payback period at your own numbers.
If a vendor cannot show their system handling your actual messages, cannot describe what happens when it is wrong, and cannot name the data it needs from you honestly — the pitch deck's chart trending upward is not evidence of anything except that someone was paid to design a chart.
Questions distributors ask before signing an AI vendor contract.
Ask the vendor to run it live against your last week of actual WhatsApp messages, not a curated demo. Your retailers’ spellings, your SKU names, your trade units. A system built for real Indian distribution handles most of it and fails visibly on the rest. A system built for a pitch fails in ways the vendor has never seen, and how they react to that failure tells you more than the demo did.
Nothing beyond time, if the vendor is confident in their product. A live test against a week of your own messages costs the vendor an afternoon and costs you nothing. Any vendor unwilling to run that test before a contract is the clearest possible signal to walk away, whatever the pitch deck claims.
Long enough to see the system handle a real edge case, not just the happy path — typically 2–4 weeks of live use on a subset of your retailers. A trial confined to demo data or a single perfect scenario tells you nothing about how the system behaves when a message arrives in three languages with a typo in the quantity, which is the normal case, not the exception.
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