Why we tell most businesses they are not ready for AI yet

A wholesale distributor called us last year wanting an AI agent to take orders on WhatsApp. Reasonable request. They had 400-odd retail customers, three phones behind the counter, and two people who spent their entire day typing orders that had arrived as voice notes.

We asked the question we always ask: where does an order go after someone types it?

The answer was a notebook. Then, at the end of the day, a spreadsheet — sometimes. Prices lived in a second spreadsheet that was six weeks out of date, so in practice prices lived in the head of the owner's brother, who was the only person who knew what each customer had been promised. Stock was counted on Saturdays.

There was nothing an AI agent could usefully connect to. If we had built one, it would have taken an order beautifully and then had nowhere to put it.

The uncomfortable sequence

There is an order to this, and it is not the order people want.

A machine can only automate work that exists somewhere it can read. If your prices are in one person's memory, no software can quote. If your stock count is a Saturday ritual, no system can tell a customer whether something is available on Tuesday. If your orders live in a notebook, nothing downstream can happen automatically, because there is nothing downstream.

This is the single most common reason AI projects in this region fail, and it has nothing to do with AI. The project fails at the point where the clever system has to touch the actual business and finds nothing there to hold on to.

What we did instead, and what it cost

We did not build the agent. We built the boring thing first, and it took about five weeks.

One product list. Every item, one price per customer tier, one place, one person allowed to change it. This took four days and was the single most valuable thing in the project, because it ended the daily argument about what things cost.

Orders captured where they already arrive. Not a new app for the retailers — they were never going to install one. The order still comes in on WhatsApp. But the person behind the counter now enters it into a form on their phone that takes twenty seconds, and that form writes to a real record with a customer, a date, items, quantities and a price that comes from the product list rather than from memory.

Stock that moves when orders move. Not a full warehouse system. Just: an order dispatched reduces the count. Saturday's manual count became a reconciliation against a number that already existed, rather than the only source of truth.

Mobile money reconciled automatically. Payments arrive on MoMo. Previously someone matched screenshots against the notebook. Now the payment attaches itself to the order, and what is left is a short list of genuine exceptions rather than an hour of matching.

One screen showing today. Sales, outstanding payments, what is running out. The owner looks at it on his phone at 7am.

None of that is artificial intelligence. All of it is the plumbing that has to exist before intelligence is any use.

What changed for them

The two people typing orders still type orders — but they finish by four instead of by seven, and they no longer spend the last hour reconstructing what happened. The end-of-month close went from three days of archaeology to an afternoon. And the owner's brother went on holiday for two weeks, which had not happened in four years, because the prices were no longer only in his head.

That last one is worth naming plainly. Most small businesses here have a single point of failure who is a person. Digitalisation, before it is about efficiency, is about that person being allowed to be ill.

And then, the agent

Eight months later, we built the WhatsApp agent they originally asked for. It took eleven days, because by then there was a product list to quote from, a stock number to check, an order record to write to and a payment flow to trigger.

It would have been impossible in week one and it was straightforward in month nine. The agent was not the hard part. The agent was never the hard part.

How to tell which stage you are at

Answer these about your own business, honestly:

• If your best-informed employee were unreachable for two weeks, what would stop?
• Can you find out what you sold last Tuesday without asking anyone?
• Is there one price list, and does everyone quote from it?
• When money arrives, does anything happen automatically, or does someone match it by hand?
• Could a new employee find out the status of a customer's order without interrupting somebody?

If those answers are uncomfortable, you do not have an AI problem yet. You have a record problem, and it is cheaper to fix, faster to fix, and worth more than the AI project you were about to commission.

We would rather tell you that in the first call than discover it in week six of a build you paid for.