Not in the product today. Nothing in Stemzo is predicted or decided by a model — route ordering is manual and payroll deductions are explicit rule-based arithmetic. Stemzo is building natural-language querying in Hindi, Bengali and English, drafted collection follow-ups, and plain-language anomaly explanation, and will describe a feature as AI only once it can be demonstrated running on a customer's own data.

Is AI in distribution management software real or marketing?

Mostly marketing, in this category. Demand forecasting is usually a moving average, route optimisation is usually a fixed beat plan entered once by a supervisor, anomaly detection is usually a percentage threshold, and smart reorder is usually a subtraction. The two features that genuinely do involve trained models — document OCR and natural-language querying over your own data — are the ones vendors tend to describe in the flattest language.

How can I tell whether a vendor's AI claim is real?

Ask to type your own question instead of watching the demo script — real natural-language querying survives an unscripted question. Ask for the model's error rate and what it was measured against. Ask the vendor to split their feature list into features that run a model and features that run a rule. Ask where your data goes when the AI feature runs, and whether it is retained. And ask to see it working on your own data during a pilot rather than on the vendor's demo account.

What is the difference between AI and rule-based automation in distribution software?

A rule is written by a person and always produces the same output for the same input — a reorder level, a percentage threshold, a payroll formula. A model is fitted to data, produces outputs with a measurable error rate, and can change as the data changes. Rule-based automation is often the better choice for operations work because it is inspectable: when a payroll deduction is wrong you can trace the exact rule that produced it, which is not true of a model.

What AI actually does in distribution software

01The short answer

Very little software sold to Indian distributors as “AI” contains a model that predicts or decides anything. Most of it is threshold rules, moving averages and lookup tables — useful, often worth paying for, but not learning from your data.

In Stemzo today, nothing is predicted or decided by a model. Route ordering is manual. Payroll deductions are explicit, rule-based arithmetic. We will call a feature AI on the day we can show it running on your data in a demo.

02What vendors usually mean when they say AI

What it is called	What it usually is	Is it actually AI?
Demand forecasting	A moving average of the last three to six months, sometimes with a seasonality multiplier	Rarely. Genuine forecasting needs a fitted model and a measurable error rate — ask for both
Route optimisation	A fixed beat plan entered once by a supervisor, replayed weekly	No. Real optimisation re-solves against today’s orders and constraints
Anomaly detection	A threshold — flag anything more than X% off last month	No. That is a rule, and a useful one, but it does not learn
Smart reorder / auto-indent	Reorder level minus current stock	No. That is a subtraction
OCR on invoices and PO documents	A genuinely trained model, usually a third-party API	Yes — and the most commonly under-claimed real AI in this category
Chat / ask-your-data	A language model over your database	Yes, when it exists. Ask to type your own question, not to watch a scripted demo
Image-based shelf audit	A trained vision model	Yes, and expensive to build. Verify it on a photo you take in the room

The pattern: the features most often labelled AI are the ones least likely to contain any. The two that genuinely do — document OCR and natural-language querying — are the ones vendors describe in the flattest language.

03What Stemzo does today

Everything below is rule-based, deterministic and inspectable. That is a feature, not an apology — when a payroll deduction is wrong you can trace exactly which rule produced it, which is not true of a model.

Two-way Tally sync. Vouchers move both directions. The hard part is not intelligence, it is correctness: an edited voucher gets renumbered by Tally, and matching on voucher number turns that into a duplicate GRN you find three weeks later. We match on something stable instead — the long version is here.
Receiving and despatch. Recorded against the actual document, by the person who handled the goods.
Proof of delivery. Captured at the door, timestamped, attached to the delivery — not a photo in a WhatsApp thread.
Route and beat planning. Entered by your supervisor. Manual, and we say so.
Approvals and expenses. Explicit rules you configure.
Payroll and attendance. PF, ESI, TDS, gratuity. Arithmetic, applied consistently.

No model is in that list. Nothing above learns.

04What we are building

In order, and only shipping when it survives a live demo on a customer’s own data.

1. Ask your data, in the language you actually use. “Kal Vishal ka kitna gaya?” → the answer. Typed or spoken, in Hindi, Bengali or English. The data is already in the system; this is a language layer over queries that already run. It is first because it is the one that changes a supervisor’s day, and because it cannot be faked in a demo — you type the question.

2. Drafted collection follow-ups. Outstanding ledger plus ageing, turned into a message per party, in the right language and the right register for that relationship. A human reads it and presses send. The value is in the drafting, not the sending.

3. Anomaly explanation in words. Not “anomaly detected” — that is a threshold with a badge. Something closer to: “This party ordered 40% less this month than the previous three. Their last three deliveries were all short-shipped.” The explanation is the product; the flag is the easy half.

Each is genuinely generative. Each sits on data the system already holds. None is on this page as a claim until it is on a screen.

05Nine questions to ask any vendor who says AI

“Can I type my own question, right now, instead of watching the demo script?” The single fastest test. Real natural-language querying survives an unscripted question; a scripted demo does not.
“What is the model’s error rate, and measured against what?” A forecasting feature with no stated accuracy has not been evaluated. If nobody can answer, there is no model.
“Which of these features runs a model, and which runs a rule?” Ask for the list split into two columns. A vendor who cannot split it does not know either.
“Does it get better as we use it, and what specifically changes?” “Self-learning” should have a mechanism behind it. Ask what quantity updates, and how often.
“Whose model is it?” A wrapper around a third-party API is completely legitimate — most good products are. But it changes who your data reaches, and what happens if that vendor’s pricing changes.
“Where does our data go when the AI feature runs?” If a language model is involved, something leaves your server. Ask what, to whom, and whether it is retained or used for training.
“What happens when it is wrong?” Every model is wrong sometimes. Ask whether a human approves before anything is acted on, and whether you can see why it decided what it decided.
“Can I see it working on my data during the pilot, not on your demo account?” Demo data is clean. Yours is not. This is where most AI claims quietly fail.
“Which logos on that integration wall are live today?” Not an AI question, but the same instinct, and the most revealing question on the list. We ask it too. Our own wall carries three integrations — Tally, Google Sheets and Google Drive — because those are the three that exist.

You are welcome to ask us all nine. Question 1 we will fail today, and we would rather you found that out here than in week three.

06Why we are saying all this

We run a distribution business in Siliguri. We have sat on the buyer’s side of that demo — the one where the software claims things. You do not find out in the meeting. You find out in week three, when you need the thing it promised.

This page used to say other things. Self-learning. Anomaly detection. Machine-optimised. It also carried thirteen integration logos, including SAP, Salesforce and AWS, that we do not integrate with, and two products in the list that were never built. All of it came off.

What is left is what exists.