What goes wrong with most AI projects
Almost all of them start backwards. Someone decides the company should do something with artificial intelligence, and then goes looking for somewhere to put it. That is how you get pilots that impress in a demo and that nobody is using three months later.
We start from the other end: which repetitive task eats your team's hours every week, how many hours it is, and what those hours cost. If that calculation does not produce a clear saving, we say so and there is no project.
What we implement
Automating repetitive tasks
Classifying incoming email and routing it, pulling data out of invoices and delivery notes into a sheet or a system, drafting first versions of documents that are written by hand every time, turning calls or meetings into client records. Language models handle these well, as long as somebody defines the rules and puts human review where it belongs.
Internal assistants over your own documentation
A conversational search that answers using the company's own manual, agreement, catalogue or project history, and cites where each answer came from. It means new staff do not depend on asking a colleague, and the answer is the same every time.
Classifying and analysing business information
Automatically tagging incidents, customer feedback, CVs or support tickets, and turning that loose text into data you can actually report on.
How we keep it from becoming a permanent experiment
- One task, not a platform. The first project is always a specific, bounded process, chosen because it can be measured.
- We measure first. How long it takes today and how many times a month it happens. Without that number there is no way to know whether it worked.
- Human review where it matters. Nothing with legal, financial or customer facing consequences goes out without a person approving it.
- Delivered working and documented. With instructions for changing it, because processes change.
Two things to settle before starting
The first is data protection. If the process touches personal data, you need to decide what information leaves the company, to which provider, and under what processing agreement. It is possible to work with models hosted in Europe, or on your own infrastructure when the data is sensitive.
The second is expectations. Language models get things wrong, and a serious project is designed accordingly: bounding the scope, forcing citations, and keeping a person at the point where a mistake would cost money. Anyone promising you it never fails has never put one into production.
The diagnostic: where to start
Before building anything we run a diagnostic. It is half a day with the people who do the work, not with management, because the person who knows how long something takes is the person doing it. It produces a report with five things:
- The candidate tasks, with how long each takes today and how often it repeats. Measured, not guessed.
- The estimated annual saving for each one, in hours and in euros, so they can be ranked by what they are worth.
- What we rule out and why. Usually the most appreciated part, because it avoids spending on what will not work.
- A recommendation of where to start, with its cost and its timeline.
- The data protection warnings that apply to those specific processes, before they become a problem.
The report is yours and it is actionable with or without us: it is written so another supplier could execute it if you prefer.
What it costs
Half a day with your team and the report described above. If we then do the project, the €490 is deducted from the quote.
One specific process running, measured against what it used to cost in time, documented and handed over.
Frequently asked questions
Do my data need to be tidy before we start?
To automate a task, no. For an internal assistant, the better the documentation the better it answers, but you can start with whatever exists and improve it later.
Will my data end up training somebody else's model?
Not with the setup we use. We work with enterprise modes that exclude training, and when the data is sensitive we can use models hosted in Europe or on your own infrastructure.
Does this replace someone on my team?
In the cases we have seen, no: it removes the part of their job nobody wants to do. If your goal is to cut headcount we are probably not the right supplier, and it is better to say so on day one.
How long does a first project take?
Between two and four weeks once the process is defined. The slow part is not the technology, it is agreeing precisely what should happen in every awkward case.
Why is the diagnostic paid when the website audit is free?
Because we do the website audit on our own, from the outside, using public information, without asking you for a minute of your time. The AI diagnostic needs us to sit with the people doing the work. That is hours of yours and ours. If the project goes ahead, the €490 is deducted.