Most owners we talk to have seen two versions of "AI": a chatbot on a website, and a demo where someone types a question into a spreadsheet. Neither is what an agent does in a company of 50 people.
Start with the Monday-morning question
Every company of that size has five or six numbers that somebody checks by hand. The owner opens the ERP export, the ops manager scrolls through a sheet, and someone notices that a customer's margin dropped three weeks after it happened. An agent is the thing that does that same check every day, on every row, and tells the right person when something moved.
That is the whole idea. Not a conversation. A routine that runs without anyone asking.
Three examples from the day-to-day
1. The margin leak alert
Take a distributor selling a few thousand SKUs to a few hundred customers. Prices are set once a year; costs move every month. Somewhere in there, a product family is being sold below its landed cost, or a customer with a "temporary" discount has kept it for two years. Nobody is lying. Nobody is looking either.
The agent reads yesterday's invoices and the current cost list, computes margin per customer and per SKU, compares it with the last eight weeks, and when a line crosses the threshold you set, say gross margin down by more than two points on volume that matters, it sends one message: which customer, which SKUs, what it costs per month, and the three most likely causes it can see in the data (a cost increase that was never passed on, a discount code still active, a freight term that changed). The commercial manager reads it in a minute and decides. The agent does not change a price.
2. Stock ageing
Warehouses hide money. A pallet that has not moved in 120 days costs space, capital and eventually a write-off, and it is invisible in a stock report that shows totals.
The agent walks the stock ledger every night, ages each item by its last movement, and keeps a short list: what is about to cross 90, 120 and 180 days, what is already there, and what that is worth at cost. Once a week it sends the list to the owner of the category with a plain suggestion: promote it, return it to the supplier, transfer it to the site that still sells it, or take the write-off now instead of in December. The decision stays with the person. The list stops being a quarter-end surprise.
3. Pipeline health
Sales pipelines rot quietly. A proposal sits in "sent" for seven weeks, the forecast still counts it, and the owner finds out at the month-end review.
The agent reads the CRM daily, looks at how each deal is moving compared with the deals that actually closed in the last year, and flags the ones that have gone stale, the ones where the next step is overdue, and the ones that are bigger than usual with nobody senior on them. It sends the sales lead a short note on Monday, and sends the owner one line: how much of this month's forecast sits in deals that have not moved for more than three weeks.
What an agent is not
It is not a chatbot. A chatbot waits for a question; an agent has a job and does it whether or not anyone asks. You can put a conversational layer on top ("why did margin drop at this customer?") and it is useful, but it is the last 10%, not the product.
It is not autonomous, at least not at first. Every agent we build keeps a human in the loop: the agent proposes, a named person decides, and every message it sends is logged so you can check later whether it was right. Autonomy is something an agent earns, one decision type at a time, after you have watched it be right for a few weeks. Our rule is simple: each agent has an owner, an evaluation and an off switch.
It is not a replacement for a report. Reports answer "what happened"; the agent answers "what needs attention today, and who should look". Most companies need far fewer reports once the agents are running.
What you need in place first
This is where most AI projects stall, so it is worth being honest about it.
One clean source for the numbers the agent will read. If the ERP, the warehouse system and the CRM disagree about what a customer is called, the agent will disagree with everyone. You do not need a full data warehouse on day one. You need the three or four tables that matter, loaded daily, with definitions that finance and operations have agreed on. On Google Cloud that is a small BigQuery dataset, and at this size it runs for a few euros a month.
A threshold and an owner for every signal. "Margin dropped" is not a signal. "Gross margin down more than two points at a customer above a set monthly revenue, sent to the commercial manager" is. Writing those sentences is most of the design work, and it has to be done with the people who will receive the messages, not for them.
A channel people already read. Email, Slack, Teams, a WhatsApp group, whichever your team actually opens. An agent that posts into a portal nobody visits is a dashboard with extra steps.
Someone to say "that one was wrong". In the first month the agent will send things that look important and are not. The person receiving them needs five minutes a week to mark what was useful. That feedback is how the thresholds get tuned, and it is the difference between an agent that gets quieter and sharper and one that gets muted.
Where to start
Pick one number that costs you money when it moves and that somebody checks by hand today. That is your first agent. Build it until the person who receives the message trusts it, then add the second. Three or four agents of that kind, each with an owner, usually cover what a 50-person company needs to stop being surprised by its own data.
If you want to talk through which number that should be in your case, we are happy to do it in a 30-minute call. No deck, no commitment.