An artificial intelligence project depends less on the choice of model than on three decisions made upfront: the intended use, the data available and how the result will be measured. This guide follows the steps in order, from idea to go-live.
What is an AI project in a company?
An AI project in a company means giving software a specific task that used to require a person to read or write: finding information, summarising a file, classifying requests, extracting data from a document. In most cases, the point is not to create a model but to connect an existing language model to your documents and tools, then check that its answers are reliable. It is a software project in its own right, with a scope, tests and maintenance.
1. Choose a precise use case
A good first use case meets four conditions:
- a frequent task that takes up qualified people’s time;
- data that already exists, as documents or databases;
- a result a user can check, while keeping the decision;
- limited risk if an error occurs.
| Type of use | Example | What to check |
|---|---|---|
| Document search | Finding a clause in contracts | Are the documents up to date and accessible? |
| Data extraction | Reading invoices or purchase orders | Who checks the extracted values? |
| Sorting and routing | Classifying incoming requests | What happens to a misclassified request? |
| Writing assistance | Preparing a reply or a meeting report | Who reviews it before it is sent? |
Conversely, it is better to rule out at the start any decision that commits the company without human review, and tasks for which no reliable data exists.
2. Check the data
AI answers from what it is given to read. Before development, check:
- where the documents are, and in which formats;
- whether they are up to date, and who maintains them;
- who is allowed to read them: the assistant must not show a user a document they cannot access;
- whether some contain personal or confidential data.
Scattered, contradictory or outdated documents produce answers of the same quality. Sorting them first often does more for quality than the choice of model.
3. Decide where the models run
There are three options, to compare according to how sensitive the data is.
| Option | Principle | Choose it when |
|---|---|---|
| A provider’s cloud service | The model is called through an API, under terms set by contract | Data may leave the company under conditions |
| Cloud in a chosen region | The model runs in an environment reserved for you | Data location must be controlled |
| On-premises hosting | The model runs on your own servers | Data must not leave your premises |
On-premises hosting requires properly sized compute resources: see our IT infrastructure services. In every case, read the provider’s terms: how long data is kept, whether it is reused for training, and the hosting country.
4. Frame confidentiality and obligations
As soon as personal data is processed, the usual rules apply: law 09-08 in Morocco, under the supervision of the CNDP, and the GDPR for people located in the European Union. The EU Artificial Intelligence Act, which entered into force in August 2024, adds obligations that apply in stages according to the risk level of each use. Have these points validated by your legal adviser. On the technical side, access management and logging are part of cybersecurity.
5. Prototype on a real case
A prototype limited to one use case and a set of real documents lets you decide on results. Before starting, set:
- a list of test questions or documents, with the expected answers;
- the success criteria: accuracy, cited sources, time saved;
- the users who will test, and how long the test lasts.
6. Measure quality before deploying
AI can produce a wrong answer with confidence. Quality is therefore measured on examples, not on a general impression: the share of correct answers, the share of answers citing the right source, and the cases where the assistant should have said it did not know. Keep this test set: it will be reused each time the model or the documents change.
7. Integrate into existing tools
A useful assistant sits where teams already work: in the business application, the intranet or the support tool. Integration is custom software development work: authentication, rights, interface, connection to data.
8. Track costs and quality over time
Once the solution is in service, three things need tracking: answer quality, based on user feedback; costs, which depend on the volume processed; and changes to documents and models. This follow-up can be part of an application maintenance contract.
Common mistakes
- Starting from the technology rather than from a use.
- Choosing a use case without reliable data.
- Not setting success criteria before the prototype.
- Forgetting document access rights.
- Removing human review from important decisions.
- Not planning a running budget after launch.
The checklist in short
- A precise, frequent and verifiable use case has been chosen.
- The necessary documents are accessible, up to date and sorted.
- The assistant applies the documents’ access rights.
- The hosting mode for the models is decided and documented.
- Personal data obligations have been checked.
- A test set and success criteria are set.
- Human review is planned for important decisions.
- Quality and cost tracking is assigned.
To choose your use cases before investing, our IT consulting team can run the scoping. For the project itself, see our AI solutions, or contact us.