AI for business: from scattered experiments to decisions that hold up

What exactly is AI for business?

AI for business is applying language models and related technology to your own company information, with a bounded goal and with a person who stays accountable. That is the whole definition. Everything on top of that is detail.

In practice that detail falls apart into three kinds, which often get mixed up even though they have completely different consequences.

Type	What it does	Who notices	Typical gain
Assistant	Waits for a question from an employee and helps with writing, summarizing, searching or programming	the individual employee	one to five hours per week per person
Automation	Runs a fixed series of steps without anyone starting it: reading in invoices, sorting mail, moving data across	a department	less manual work, fewer errors
Agent	Has its own assignment and its own access to sources, comes forward with a signal on its own and backs it with source and amount	the management team	faster decisions with better evidence

Most companies start with the assistant, because it costs nothing and works right away. That is fine. The problem is that it also stops there. Everyone gets a little faster individually, and the company does not change. The question that really matters only comes up with the third kind: which decisions do you now take too late or on too little information?

What can AI actually do in a company?

Below is what we see working most often in practice, per part of the business. Not what is theoretically possible, but what runs within weeks and whose return you can measure.

• Margin and post calculation. Putting hours written next to signed orders and reporting the difference per project. Additional work that was delivered but not invoiced is almost always the biggest leak in project businesses, and it almost always only stands out in the post calculation, weeks later.
• Capacity and planning. Putting the schedule next to time tracking and warning as soon as a project starts to run out of line, instead of at handover.
• Cash. Tracking payment behavior per customer, putting outstanding items on one overview every week and naming the customers where it chafes.
• Sales. Analyzing lost quotes on price, lead time and offer, and tracking the win rate per segment.
• Purchasing. Putting price agreements next to actual invoices. Suppliers do not always stick to their own volume discounts.
• Knowledge. Capturing what sits in the heads of two or three key people, and making that searchable for the rest.
• Customer contact. Answering recurring questions with an agent that reads the real documentation, not with a chatbot that makes something up.

What differs per sector is mainly the order. For installation companies we worked that out in AI in the installation sector.

What stands out about this list: it is rarely about technology and almost always about information that is already in the company, but in the wrong place, at the wrong moment or in the wrong system.

What is the difference between an AI agent, a chatbot and ordinary automation?

This distinction determines what you buy and what you may expect from it, so it pays to have it sharp.

	Chatbot	Automation	AI agent
Starts	when you ask a question	at a fixed event	on its own, at its own rhythm
Knows	what you give it in the conversation	what the rule says	what the connected sources say
Can handle	anything, but without context	only what was thought of up front	situations that were not thought of up front
Delivers	an answer	a completed step	a signal with source, amount and proposal
An error means	a strange answer	a stuck process	a badly prepared decision, so review is mandatory

We work out this distinction further in what are AI agents. The practical rule: the more autonomous the system, the heavier the demands on traceability. An agent that cannot show where a figure comes from does not belong in a management meeting.

Where do you start?

Not with the technology and not with a list of tools. Start with a decision that is now taken too late or on too little information. Four questions are enough to decide whether a topic is suitable as a first application.

1. Is it a recurring decision? Something that comes back every week or every month. One off questions do not lend themselves to automation.
2. Does the information sit somewhere? In a system, in an export, in email, in documents. If nobody knows, an agent cannot find it either.
3. Can you measure the difference? In days of lead time, in hours of digging, in euros. Without a baseline measurement you will not know afterward whether it worked.
4. Does it hurt? If the answer to the previous three is yes but nobody loses sleep over it, choose something else. Attention is the scarce resource, not the technology.

In our decision scan we walk through eight places where decisions get stuck in midsize companies: margin and price, capacity and planning, cash and working capital, sales and quotes, knowledge held by key people, information at the moment of decision, decision speed, and coherence between departments. In most companies two or three of these score noticeably worse than the rest, and that is almost always where you should start. Six questions every leadership team should be able to answer is a quick way to test that yourself.

What does an AI implementation look like over time?

A good first implementation takes weeks, not quarters. We work with 45 days, and that term is not arbitrary: it is two full management cycles, long enough to prove something and short enough to hold the attention. Anything that takes longer quietly turns into a program, and programs are rarely finished. The full setup, including what we measure and when we stop, is in the 45-day pilot.

1. Day 1. Establish which challenge you take on and why.
2. Day 2 to 7. Diagnosis and baseline measurement. How many days sit between signal and decision today? How many hours of digging come before it? How complete are the sources? Without this week you cannot claim anything afterward.
3. Day 8 to 30. Building. One application, on your own sources, read only. Halfway through, something runs that the leadership team can look at.
4. Day 30 to 44. Proving it with real users and real questions. Every outcome is tested: is the figure right, is it traceable, would you have known this anyway?
5. Day 45. Deciding. Sharpen, scale up or stop.

Stopping is a normal outcome. An approach in which stopping is not an option is not a pilot but a purchase with a delay.

What does implementing AI cost, and when does it pay back?

On the Dutch web the amounts for a first AI implementation run from ten thousand to a hundred thousand euros. That spread says mostly something about what all gets sold under the same name. It is more useful to look at where the costs come from.

• Opening up the sources. The largest item, and the only one that really varies. A system with a decent connection costs hours, a system without one costs days.
• Building the application itself. Fairly predictable if the scope is one challenge.
• Model usage. For an application that runs a handful of analyses a day, you are talking about tens of euros per month, not thousands.
• Time from your own people. Count on two to three hours per week from someone who knows the challenge, plus two sessions with the leadership team. This is the item that is forgotten most often and most often the reason a project runs aground.
• Maintenance afterward. An agent that nobody keeps up is ignored within six months.

You do not calculate the payback on the invoice but on the baseline measurement. Three figures are enough: how many hours of digging disappear per week, how many days shorter a decision takes, and how many euros of value found have been confirmed by someone accountable. That last condition is strict and that is deliberate: an estimate from a model does not count; only once someone puts their name under an amount is it real.

Why do AI projects run aground in midsize companies?

There is little mysterious about it. Five causes explain nearly everything we come across.

1. It started with the tool. Someone bought licenses, and then people went looking for a problem. That never works.
2. There was no baseline measurement. Afterward nobody can show that anything improved, so the attention drops away and the budget with it.
3. Too much at once. Five half finished applications deliver less than one that is finished.
4. No owner. An AI application without someone accountable at the client is a demo, not a change.
5. Nobody trusted the outcome. If a figure cannot be traced back to a source, it gets waved away in the meeting. Rightly so.

Notably, data quality is not on this list, while it is the objection raised most often up front. We come to that shortly.

When should you not start with AI?

This question is rarely asked and is perhaps the most useful one on this page. There are four situations in which we ourselves advise against it.

• When the real problem is a process. If additional work is not invoiced because nobody writes it down, no agent solves that. Make the agreement first, then the measurement.
• When there is no decision attached to it. Interesting insights without a decision change nothing and drop out of view after two months.
• When nobody has time. Two to three hours per week from one involved employee is the floor. If that is not possible, wait until it is.
• When the organization is in the middle of something big. An acquisition, a system migration or a reorganization takes all the attention. A pilot alongside it is the first thing to fall over.

Does my data have to be in order first?

No, and this is probably the most expensive misunderstanding in the market. Companies postpone AI for years because the data has to be in order first, and that job is never done.

An agent reads your sources as they are today. Messy data does lead to weaker signals, so in the first weeks you measure how complete and how current your sources are. That insight is often valuable in itself: it is usually the first time anyone says out loud that time tracking is three weeks behind.

What does really have to be true: you have to be able to get at it. A system for which the vendor supplies no export or connection is a system no application can reach. That is a vendor problem, not a data quality problem, and it is worth checking up front.

What does the law say about this?

Two frameworks are relevant, and both are manageable.

The GDPR applies as soon as you process personal data. For most management applications it is about company data, not about people, and then the question is limited. As soon as personal data is involved, a data processing agreement should be in place with everyone who processes it, including the party that supplies the model.

The European AI Act classifies applications by risk. An agent that informs and advises management falls into the lightest categories. It gets heavier as soon as a system takes decisions about people on its own, for example in recruitment or appraisal. The practical translation: record per application what the agent may do itself, where a person approves up front, what is logged and who the owner is. How you set that up is in AI at the table, people at the helm. We record it in a governance annex to every contract, and that is not a formality but exactly what you need when someone asks six months later how a figure came about.

How we approach this

The Sixth Seat is a Dutch AI company that builds AI agents and Management Intelligence for growing businesses. The image our name comes from: at a table with five chairs, a sixth one is added. That sixth seat has no vote, but it does have the best overview.

The order is always the same. First the decision scan, ten questions that show where your decisions get stuck. Then one application in 45 days, with a baseline measurement up front and a factual decision afterward. If it does not work, we stop. If it does, the next application comes on the same foundation, and from about five applications on a coherent management layer across the whole company takes shape, a Digital Business Twin.

People decide. The agent advises, flags and backs it up, but signs nothing.

Frequently asked questions

How long before I see something working?

With a bounded challenge, something you can look at is standing halfway through the third week. You take a decision about continuing or stopping on day 45.

Do my systems have to be changed?

No. We read out your sources, we write nothing back into them. That is a deliberate choice: it keeps the risk low and the way back open.

What if the outcome is wrong?

Then you have to be able to see it. Every figure is traceable to the source it came from. Findings that do not survive the test come off the list before they reach the leadership team.

Can an agent send an invoice or place an order by itself?

Technically it can. We do not do it without explicit approval up front. Flagging, advising and deciding stay separate.

Is this only for large companies?

No. It works from about twenty employees, because that is where you first notice that information starts to drift apart. Above a hundred the pain is usually bigger and so is the gain.

What does the decision scan cost?

Nothing. Ten questions, three minutes, and you get a report with your three biggest decision bottlenecks, what they cost per year and which first application proves fastest what AI delivers.