Short answer
An AI consultant is an adviser who helps an organisation work out where AI is worth using, sets it up, and gets the staff to use it.
In practice, an AI consultant does four things: works out where the time goes, picks out what is worth doing, builds or sets up the solution, and gets the staff to actually use it.
In a small company the emphasis is almost never on the technology. It lies in the selection and in the rollout — choosing the right thing and then getting it to survive contact with an ordinary working week.
What is an AI consultant?
An AI consultant is an adviser who helps an organisation work out where AI is worth using, sets it up, and gets the staff to use it. Anyone can use the title, so it covers very different jobs.
In a small or medium-sized company the work is mostly practical: sit with the people who do the work, find the tasks that eat time, pick the few worth automating, build or configure the first one, and train the staff. The rest of this article goes through it step by step.
What it is not
The title “AI consultant” says nothing, because everyone uses it. Under the same sign you will find researchers who train models, developers who build integrations, salespeople who repackage someone else's product, and people who were doing digitalisation last year and have changed the words.
So let me first say what I do not do, because that clears up the most common misunderstandings in the first call.
I don't train models. Hardly anyone does, and those who do work at companies with research departments. What a small or medium-sized company needs is already ready to use; the question is what you are going to use it for.
I rarely install anything new. Most organisations I come to already pay for tools with AI features built in. The cheapest improvement available is for someone to show what the thing you already pay for actually does.
I don't deliver a forecast of the future. The presentation about what will be possible in three years is free to make and worthless to own. What is worth money is a list of what you should do, and in what order, with the tedious parts left in.
Four weeks in an engagement
Here is what a fairly typical engagement looks like at a company with twenty to fifty employees. The timings vary, but the order rarely does.
Week one: find out where the time goes. I sit in, ask questions and ask people to describe their week. Not the managers first — the people who do the work. The most important question is seldom “what would you like to automate?” but “what did you do the day before yesterday that you have done fifty times before?” There is almost always at least one thing that nobody has thought of as a problem, because it has always been that way.
Week two: choose what to leave out. This is the part that decides whether the engagement is good. After the mapping there are always ten to twenty things that could be done. Doing all of them is impossible, and doing them in the wrong order is expensive. The selection comes down to three questions: how much time it saves, how often it happens, and how badly it turns out if the solution gets something wrong. Things that save a lot of time but do serious harm when they go wrong end up lower on the list than you might expect. The result is an AI strategy — which is a grand word for a prioritised list with costs and a decision.
Week three: build the first one. One bounded build, not a platform. Often it is something that could be described in a single sentence: “all incoming enquiries are summarised and filed in the right folder with a draft reply”. It is seldom technically difficult. What takes time is getting it to work with the systems you already have and with the exceptions your business is full of.
Week four: training and handover. This is where most engagements fail, and that is why this part is the one I take the most care over. A solution nobody understands the reason for gets used for three weeks. The training is less about button presses than about what the technology is bad at, how you can see that it has got something wrong, and what you do then. More on how such a half-day should be set up.
And then, a month later. Follow-up is always included, because that is the only way to find out what is actually being used. It is almost never exactly what you expected, and that information is often worth more than the whole mapping.
What you should have in hand when it is done
An engagement should leave behind things that work without the consultant. Concretely, that means four things:
A decision document, not a presentation. A document that says what is to be done, in what order, what it costs and what you get out of it. The sort of thing you can put in front of a board, or pick up again in six months and still understand.
At least one thing that works. Something you use for real, not a demo. If the engagement ends without anything having changed in everyday work, it has failed, however good the document was.
Staff who can use it. And who know where the line is for what they are allowed to put into the tool.
Written rules. A short AI policy that says what may be entered where, who is responsible and what applies under GDPR and the EU AI Act. Two pages are almost always enough.
What separates a good consultant from a bad one
Having sat on both sides of the table myself, these are the signs I trust.
The good one asks more than they tell in the first meeting. The bad one arrives with a solution before they have heard the problem.
The good one sometimes says no. Roughly one in five of my first calls ends with me saying that it isn't worth the money for you. If a consultant never advises against anything, they are not an adviser but a salesperson.
The good one talks about what happens when things go wrong. All of this technology gets things wrong sometimes, and it does so with great confidence in its voice. A consultant who doesn't raise it on their own initiative has either not thought about it or chosen not to mention it, and both are bad.
The good one wants to make themselves redundant. The goal should be for you to manage on your own afterwards. The agency model rewards the opposite, and it is worth being aware of which of the two business models you are buying from.
When you don't need one
Four fairly clear signs that you should keep your money:
You have someone internally who is interested, has the time and gets to decide. Then a half-day training session for the whole group is a far better deal than an engagement. I say this often, and it is still true.
The business is too small for it to pay for itself. If an automation saves twenty minutes a week, it takes years to earn back SEK 15,000. Do the sums before you buy — it takes five minutes.
The underlying problem is not technical. Unclear processes, an ongoing reorganisation or nobody having the mandate to decide. AI on top of an unclear way of working makes the unclear faster, not better.
You don't know why you are asking. If the reason is that a customer or a board member has said you “should do something with AI”, the first step is to find out what they mean, not to buy an engagement.
The most important thing
If you take just one thing away: the work is less technical than it sounds and more organisational than one would like.
The hardest part of my job is not building something. It is choosing to leave out nine things in ten, and then getting the tenth to survive contact with an ordinary working week. A consultant who only talks about the first has not done it often enough.
What it costs is covered in a separate text: What does an AI consultant cost?
Best wishes, Mikael GlännströmWellbeing coach, author and father of two
