AI project in the company: how to really get started

The question ‘how do we start with AI’ has an answer that disappoints those who expect it to be technological: you start by counting. Counting how much a repetitive process costs today, how many times it happens, how many errors it produces, how long it wastes.

The reason is practical. Without that number you don't know what project is worth doing, you don't know how much you can spend, and at the end of the job you won't know if it went well. All three problems occur together, and result from the same skipped passage.

The six steps, in order

1. Subscriptions for three months. A few tens of euros per person. It serves to find out where AI affects real work, not in theory.
2. Census of repetitive processes. One line per process: how many times a month, how many minutes, who does it.
3. Candidate's choice. High annual cost, low exceptions. Not the one that bothers the most.
4. Starting point measurement. One month of real data before you touch anything.
5. Realization with narrow perimeter and an internal contact.
6. Verification of numbers after a few months, and only then the decision to extend.

Step 4 is the one that is skipped practically always, and it is the one that makes everything else evaluable. The calculation method is in how much it costs to automate a process.

How to choose the first project

The criterion is a relationship between two quantities, and it is the same in any business function.

The last line is not about the process but about the organization, and it's still decisive. A technically perfect candidate without an internal referent produces a system that no one adopts.

Why not write a strategy first

It is a countercurrent position and it is worth arguing.

An AI strategy written before you realize something is necessarily based on assumptions: what AI can do for you, how much it costs, how long it takes, how the team will react. These are all things that no one can know without having tried them, and that change after the first project.

The typical result is a document with an eighteen-month roadmap that after the first real work turns out to be wrong in three out of four points. In the meantime, months have passed in analysis.

The most effective order is the opposite: a small, measured and concluded project, and then the strategy built on what you learned. At that point you know how much it really costs, how long it takes, who in the company can take care of it and where the bottlenecks are.

The internal contact person

It is the variable that predicts success better than any other, and is treated as an organizational detail.

• He knows the process, not necessarily the technology
• Ha allotted time, not just personal interest
• It can decide on exceptions without going up a chain
• It will also be there in a year, for maintenance
• He has the authority to say that a thing is not done

The last point is the least obvious and the most useful: without someone who can say "this exception we do not handle", the perimeter widens during the work and the project never ends.

What to expect from the first project

It is worth setting the right expectations, because disappointment almost always arises from wrong expectations rather than poor results.

Practical examples

Professional studio, 10 people. Three months of subscriptions revealed that the best use case was not the one expected: not the search in the rules, but the preparation of recurring responses to customers. The project started there and cost less than budgeted for the initial idea.

Commercial company, 30 people. The census showed that the most expensive process was entering invoices, which no one had ever reported as a problem. The process that everyone hated —the handling of returns —had too many exceptions and was postponed.

Craftsmanship, 6 people. The count said that no trial justified a project. They stopped at subscriptions, which however changed the way they prepared quotes and responded to emails. It is a legitimate and more common outcome than is said.

The three questions to ask a supplier

When you get to engage someone from the outside, three questions distinguish who understood your problem from who is selling a pre-packaged solution.

"What happens when the system goes wrong?" A supplier who answers by talking about the accuracy of the model has answered another question. The useful answer concerns the process: who notices it, where the error ends, how it is corrected.

"What exceptions won't you handle?" A quote that promises to cover all cases is a quote that will find out the exceptions during development, and will remake them for you as variants. Those who work well tell you in advance where it stops.

"Who is to follow him from us, and how long does it take?" If the answer is ‘no one, it is turnkey’, the supplier either has not understood or prefers not to say it. Each integrated system takes internal time, at the project stage and after.

The answers to these three questions predict the outcome better than the references and portfolio. The theme is deepened in how to choose an AI provider and in contract with an AI provider.

What to do while waiting for the three months

The trial period with subscriptions is not dead time: it is the time when the ground is prepared, and companies that use it well start advantaged.

1. Watch what people use. Tools will emerge that no one has authorized: it is valuable information, not a disciplinary problem.
2. Write two policy pages What doesn't have to come out of the company. It still needs to be done, and it needs to be done before someone gets the wrong file.
3. Place order in the data of the candidate process: records, documents, nomenclatures. It is the work that no supplier can do for you and that determines the cost of the project.
4. Locate the referent and free him time before the project starts, not during.

Point 3 is the one with the greatest impact on cost. A project on clean data costs a fraction of one on messy data, and the remediation done before calmly costs less than that done in the race during development.

Point 2 is dealt with in corporate policy on the use of AI: two well-written pages, not a regulation.

Mistakes to Avoid

• Starting with technology. It leads to looking for a problem that justifies the purchase.
• Skipping the initial measurement. It makes it impossible to evaluate the result.
• Choosing the most hated process. It is usually the least suitable.
• Starting multiple projects together. No one is followed and nothing is learned.
• Not assigning a contact who has time. It is the most frequent cause of failure.
• Writing the strategy first. It is based on assumptions that the first project will disprove.

How to apply it in the company

This week: Activate subscriptions for people who can take advantage of them and let three months go by without deciding anything.

In the meantime, have a table filled in with one line per repetitive process: how many times a month, how many minutes, who does, how many exceptions. Ask who works, not who coordinates.

At the end of the three months you will have two things that you do not have today: a concrete idea of where AI affects your work, and an orderly list of candidates. From there the choice is made almost alone. The next step is the pilot project, and the rating meter is in ROI of an AI project.

When to stop

A good method of getting started includes the criterion of not getting started. There are situations where the correct answer is to postpone, and recognizing them avoids burning the first attempt — which in SMEs is often also the last, because a project gone wrong closes the topic for years.

• No one has time to follow the project: postpone until there is
• The Candidate Process is changing: wait for it to stabilize
• The Data is unreliable and no one wants to fix them
• The company is going through other changes important
• The push comes from outside and no one internally wants it

The last point is the most insidious. Projects started because "you have to do something with AI", without an internal problem to solve, produce technically working systems that no one uses. The cost isn't just cheap: it's that the company concludes that AI isn't needed, when it really wasn't needed that project.

The second project

It is worth anticipating what happens next, because it changes the way you set the first.

The second project costs less than the first for a structural reason: basic integrations already exist. If you have connected the management system to automate invoices, that link is also used for purchases, for the warehouse, for reporting. The most expensive part is already paid for.

This suggests a criterion in choosing the first: for equal convenience, choose the one that opens more streets. A project involving the business management software is worth more than an isolated one, even if the direct return is similar, because the second and third projects become cheaper.

Conclusion

An AI project starts with a sheet and a count, not a technology. The right process is the frequent, regular and measured one; the right person is the one who knows the job and has time; the right time for strategy is after the first project, not before.

The thing that really makes a difference is the least technical of all: Measure how it is now. It is what makes spending defensible first and the result after, and in most companies it is also the only thing that is missing.

If you want to understand what processes in your company can be automated with AI, Giallo Studio can help you turn the problem into a real workflow. You can see systems working in The Lab or starting from AI for Companies: Complete Guide.