AI in a company of 10 to 250 people: where it pays off and where plain automation will do

How AI differs from plain automation

Automation is rules. When an order changes status, a text goes out. When stock drops below ten units, a purchase order is created. The same condition always gives the same result, and the program cannot get it wrong if the rule is written properly.

AI handles what cannot be put into rules: text, images, speech and ambiguity. It reads an order from an email a customer wrote in their own words, recognises that a ticket about “a washing machine that jumps” is a complaint and not a price enquiry, and finds the penalty clause in a contract even though nobody called it “penalties”.

The price of that flexibility is that AI is sometimes unsure and sometimes wrong. So where a mistake costs money, AI proposes and a person approves. A rules-based program needs no such supervision.

If the process can be described with rules, automation is enough. It is cheaper, predictable and needs no oversight. AI makes sense when the input is something a rule cannot handle.

Four uses that pay off in a small or mid-sized company

In manufacturing, distribution, service and trading companies AI pays back fastest in four places.

• Reading documents. Purchase invoices, orders in PDF, contracts, reports, consignment notes. AI pulls the data out and enters it into the system, and a person checks the preview and clicks “approve”. What the screen looks like and what happens with errors is on the page about reading invoices and documents.
• Sorting and first handling of requests. An email from a customer lands with the right person, tagged with its topic, urgency and the data already gathered, instead of sitting in a shared inbox until someone reads it. Requests about the same thing are merged into one.
• Answering repeat customer questions. Order status, delivery date, where the invoice is, how to make a complaint. AI answers from the customer’s data in the system and from company procedures, and hands hard cases to a person. The limits are on the page about answering customer questions.
• Searching company knowledge. Procedures, contracts, quotes, emails. A staff member asks in plain words and gets an answer that points to the document it came from. Details on the page about searching company knowledge.

In all four the input is text or an image from a person, and the output is data in the system that someone is waiting for today.

Where AI makes no sense

I will say so on the first call if that is the case. The most common situations:

• The process has clear rules. Discount calculation, payment reminders, generating documents from a template. A plain program does that, cheaper and without mistakes.
• There is little data. Ten invoices a month the bookkeeper retypes faster than they can check what AI read. The effect appears at hundreds of documents or requests.
• A mistake is unacceptable and nobody will check. If no one has time to review AI’s proposals, and an error means a wrong shipment or a wrong transfer, stay with rules.
• The point is to “have AI”. A function that takes work off nobody costs money and never pays back. I always ask how many hours a month it should save, and for whom.

How to build AI into the system your team works in

A separate chat window where staff are supposed to paste an invoice stops being used after a week. AI works when it sits in the same panel the team already uses: a “read” button next to the document, a suggested reply in the ticket, a search box above the list of procedures.

That is why I build it as part of the system, not as an extra tool. The system is made of modules, meaning self-contained parts: orders, documents, customer service. I add AI to the module it is meant to work in, and where a mistake costs money, the result waits for a person’s approval. If there is no such system yet, I start with its first part and add AI to that. How such a system gets built is on the page about internal systems for companies.

Data and confidentiality

Before we start we settle three things. Which data may go to AI at all and which never may, for example customers’ personal data or contract terms. Where AI runs: with an outside provider whose terms do not allow them to use your data for their purposes, which I check before we start, or on a server we choose together when the data is sensitive. Who in the company sees the results, because searching company knowledge must not show a staff member a contract they could not open themselves.

All of that comes to you in writing with the scope, in language your lawyer and your bookkeeper will understand.

How I start

1. One process. We pick the one with the most documents or requests and the longest wait for a result. We do not write an “AI strategy for the company”.
2. Measure before. How many documents a month, how many minutes each, how many people. Without that nobody can say whether the function paid back.
3. A function in the system. I build it in the module the team works in, with an approval screen. It is usually ready to use within the first month.
4. Measure after. The same month, the same numbers. If there is an effect, we add the next process. If there is none, I say so plainly.

You pay after accepting a working function, not up front and not for presentations.

Risks and running costs

AI has costs that do not go away after launch. Providers charge per use, usually for every document read or reply generated. At hundreds of documents a month that cost is usually small next to the hours of work, but we work it out before we start and you get it in writing.

The second issue is quality. Providers change their tools, and documents from a new supplier can look different from the ones we set the reading up on. So every function has a screen that shows how many results a person corrected. When that number rises, you know something needs tuning before anyone spots an error on an invoice.

If you want to check whether AI makes sense in your process, describe in the form below where you have the most documents or requests and how many there are each month. I will get back to you.