Guide
AI in your business: where to start
Most owners I speak to know something has to happen with AI, but not what. This guide gives an order that works for companies between ten and two hundred people: where to start, what to skip, roughly what it costs, and where it usually goes wrong.
Do not start with AI
The question "what could we do with AI" rarely leads anywhere. It is too open, and the answer is always "all sorts of things", after which nothing happens.
The question that does work is: where does most of our time go on work nobody enjoys. Your colleague can answer that within five minutes, and the answer nearly always points at something automatable. Sometimes with AI, often with ordinary automation that is cheaper and more reliable.
That order is the single most important advice in this guide: the problem first, the technology second. The other way round you build a solution looking for a problem, which is exactly where most stalled AI projects come from.
What you do in the first week
Buy no software, invite no supplier. Do this instead:
- Ask five people from different corners which recurring work costs them the most time
- Ask roughly how many hours a week that is; a rough estimate is enough
- Note which system is involved per task and whether it allows an integration
- Strike out everything costing less than two hours a week
- What remains is your list, sorted by hours
Which one to do first
Do not pick the biggest. Pick the first that meets three conditions: it demonstrably costs time, the systems involved allow integration, and one person is willing to own it.
That third condition gets skipped most often and causes the most trouble. An automation without an owner stalls within a year, usually because a supplier changed something and nobody felt responsible for fixing it.
Expect the first application not to deliver your biggest saving. That is fine. The first one exists to learn how such work runs at your company, and to show internally that it works.
What it costs
What a first, bounded application costs depends heavily on how many systems are involved. Running costs for licences and processing come on top, and are usually modest next to the build. A concrete figure is only worth quoting once it is clear what the work actually is.
The rule of thumb I hold to: a first application should pay for itself within a year, and you should be able to calculate that up front from hours saved. If you cannot do that sum, the application was chosen too vaguely. Ask any supplier for a fixed price per phase rather than open-ended billing.
What you do not need at this stage: a data warehouse, an AI strategy on paper, or a platform you pay for annually. Those come later, if they come at all.
The four most expensive mistakes
I see these most with companies that already tried once and were disappointed.
- Wanting the data in order first: that becomes a two-year programme with nothing to show, while the data is usually good enough for a first application
- Starting too big: an application touching five departments has five times the chance of stalling
- Agreeing no success criterion: without a bar set in advance, any result gets argued into a success or a failure afterwards
- Outsourcing everything with no handover: then you hire someone again every year for something you could have maintained yourself
When to bring someone in
You can do the inventory yourself, and it is better that way: you know your business. External help pays off once you have to judge which application is feasible, or once you need to integrate with systems that do not cooperate.
What to watch for when choosing help: whether they ever say no. An adviser who answers every question with "that is possible" costs you more in the end than someone who tells you your first idea is not a good one.
Frequently asked questions
How long does a first AI application take?
From first conversation to something running, six to twelve weeks is realistic for a bounded application. Of that, two weeks is inventory, two to four weeks building and testing, and the rest production setup and handover. Anyone promising a working application within two weeks is leaving out the testing.
Do we need our data in order first?
Usually not. That advice sounds sensible but leads to multi-year programmes with no visible result. For most first applications the data is good enough, and where it is not, that surfaces within two weeks. Then you know exactly which part of your data needs attention and why, which is a far better starting point than cleaning everything at once.
Is our company too small for AI?
For training your own models: yes, almost certainly. For applying existing models to your own processes: no. The threshold is not the size of your company but whether there is enough repetitive work to earn something back. From roughly ten employees upward there almost always is.
What if it does not work?
That happens, and it is containable by starting small. A proof of concept that fails costs a few weeks; a fully rolled-out project that fails costs a multiple of that. So I work with a decision point after each phase, and stopping after a failed trial is not an exception but part of the approach.
Made the list and got stuck?
Send a few sentences on what is at the top. I will tell you what I think, which approach fits, and whether it is worth doing, including when the answer is no.