Hardly any artificial intelligence project fails because of the technology. It fails because it started without being able to answer a few questions: what problem it solves, with what data, who will use it and how anyone will know whether it worked.
This guide sorts those questions into four groups. You do not need every answer to get started; you do need to know which ones are missing. With that, any proposal that reaches you —ours or anyone else’s— can be assessed properly.
The problem
Describe it in one sentence
“We want to use AI in customer service” is an idea. “We want questions about an order’s status to be answered on their own, at any hour” is a case. The difference is that the second can be built, tested and measured; the first can be debated for months.
If the task does not fit in one sentence, the first step is breaking the idea down until you find a concrete case.
Measure what it costs today
Before changing anything, write down what it costs as it is: hours per week, response time, errors per month, whatever the task consumes. That is the baseline.
Without it, three months later nobody can say whether the AI delivered. With it, the conversation changes: “we used to reply in a day and now in a minute” is a result; “people are happy” is an impression.
Separate the core from the friction
Every company has work that makes customers choose it —sharp quoting, design, negotiation— and plenty of friction around it: repeated questions, incomplete requests, data that has to be fetched from three systems.
AI pays off on the friction. Aiming it at the core of the business does not save money: it makes you just like everyone else in the one thing that set you apart. We explain it with a real case in the note AI does not belong where your value is.
Ask whether a rule would do
If the decision can be written as a table of conditions —“if the amount is above a certain figure, the manager approves it”— you do not need an AI model: you need to write the conditions down. It is cheaper, faster and always gives the same answer.
AI makes sense when the task has real variation: different texts every time, requests written a thousand ways, information that has to be interpreted.
The data
Know where it lives
Half of an AI project is getting to the data. List what the task needs and where each piece comes from: the management system, a spreadsheet, emails, someone’s head. Whatever only lives in someone’s head has to be written down first.
Check what state it is in
A model fed old or contradictory data very confidently answers things that are no longer true. Everything does not need to be spotless, but you do need to know what is up to date, what is duplicated and what is missing.
Decide what must not leave the company
Customer data, contracts, salaries, health information: before starting, decide what the AI may see and what it may not, and where the data has to stay. It is the company’s decision, not the provider’s, and it is best made before someone pastes a customer spreadsheet into a free tool.
Gather real examples, good and bad
Ten or twenty real examples of the task —done well and done badly— are worth more than any specification. They are how you test whether it works, and the yardstick you measure it against.
The people
Who knows the process end to end
There is always someone who knows how it is really done, shortcuts and exceptions included. That person has to be in the project from the start: they know where the AI will get it wrong before it does.
Who reviews what the AI produces
AI makes mistakes, and it will keep making them now and then. Decide who reviews what it produces, how often, and what happens when they find an error. A system nobody reviews will, sooner or later, send a customer something wrong.
Who will use it
A tool nobody asked for is a tool nobody uses. Bring in the team who will use it from the idea stage: they will find problems that never come up in a management meeting, and they will want it to work.
The provider
These are the questions worth asking any provider, us included.
Which case do they suggest starting with, and how will it be measured?
A good provider suggests starting with a small case, in real use and with a result measurable against the baseline. A first working case teaches more than a whole transformation plan.
Which AI do they use, and why that one?
There are models from the main providers and open-source ones, and each changes the quality, the cost and where the data stays. The answer should explain the choice for your case, not repeat a brand name.
Where does the data stay, and is it used for training?
There has to be a clear answer before signing: which service it stays in, in which country, who can access it and whether the terms allow it to be used to train models.
Are scope and price closed?
If the provider bills by the hour until it works, nobody benefits from finishing quickly. Scope and price closed before starting align the interests: both sides want to finish well and soon.
Warning signs
- An impressive demo built on data that is not yours.
- A promise to “transform” the whole company in the first project.
- Nobody asks how the result will be measured.
- Nobody talks about what happens to your data.
The checklist, on one page
The problem
- The task to improve, described in one sentence.
- What it costs today: hours, response time, errors.
- It is friction around the work, not the core of the business.
- A written rule would not be enough.
The data
- Where each piece of data the task needs lives.
- What state it is in: up to date, duplicated, missing.
- Which information must not leave the company.
- Ten or twenty real examples, good and bad.
The people
- Who knows the process end to end.
- Who reviews what the AI produces, and what happens when it gets it wrong.
- The team who will use it, involved from the idea.
The provider
- A small, measurable first case.
- Which AI they use, and why that one.
- Where the data stays, and whether it is used for training.
- Scope and price closed before starting.
If you can answer most of these, you are ready for a first case. If not, the next step is sorting out the problem. That is what the express assessment is for.