How to introduce AI in a department
Pick a task that already exists, test it on real cases, and write down when a person has to review the result.
An AI project with no task
'We want to use AI' does not say what will stop being done by hand. Without that sentence, the pilot shows a chat and changes nobody's week.
The opposite risk is letting the model answer alone where a mistake has a cost: a price, a promise to a customer, an employment record.
A custom model is not required on day one. ChatGPT, Claude or another tool can serve if the data it sees is limited and a person reviews what is not routine.
What a small implementation is
An assistant, a search over the department's documents, or support for one step of a process. One of those.
A test on this week's cases, not on demo examples.
A rule: what can go out directly and what returns to a person. Training teaches that rule. It is not a generic AI course.
When it is worth it
- Someone searches or summarises text or documents every day.
- The answer can be checked against an internal source.
- The department has a person who can say whether the result is usable.
- The volume justifies stopping the all-by-eye version, not a pilot out of curiosity.
What can be left running
An assistant for one task
It summarises, classifies or drafts from material the company already has.
Internal search
A question over a bounded folder or base, with the source in view.
Support for a process
One step of the flow, not a virtual employee that does everything.
How it is bounded
Possible means
- Generative AI tools
- Documents and bases you already have
- Python
- APIs
- Human review
A typical situation
Administration searching for a clause
Each contract is opened by hand to find a condition. A search over those documents, with the page cited, changes the task. A chat that invents the clause does not.
Common mistakes
- Starting with a customer-facing chatbot.
- Not saying which source is valid.
- Measuring the project in number of prompts.
- Giving access to data that department should not see.
Frequently asked questions
Do we need a data scientist?
Not for this cut. You need the task, the examples, and someone to implement and teach.
Does it replace a person?
That is not the promise. It removes a repeated slice and leaves judgement with the team.
What about privacy?
You bound which documents go in and who can ask. If that is unclear, it stays off.
Where is the rest of the AI detail?
In the AI for business hub. This guide sets how to enter through one department.
Is a result guaranteed?
No. You agree what will be observed: less time searching, reviewable drafts, or it is switched off.
Related guides
Is there a concrete task for AI?
Describe it and say which mistake you cannot afford. If there is no task, there is no project.