Where do I start with an AI assistant for the team?

Many teams already use AI tools on their own initiative: one writes emails, another summarises meetings, another asks for ideas. Moving from this to a company assistant, shared and with rules, is the next step. But starting at the end, with a tool that does everything, almost never works.

Step 1: choose a task

Just one, to start. Good candidates are usually:

• Answering internal questions about procedures (how X is done, where Y is).
• Preparing drafts of emails, proposals or posts, which a person reviews.
• Summarising meetings and listing the next steps.
• Searching for information in company documents.

If you do not know where to start, read which tasks to automate first and when not to automate.

Step 2: choose the tool and the account

Use a company account, not personal accounts. Check in the settings and the plan's terms what happens to the data: whether it is used to improve the model, who can see it, how it is deleted. Choose the plan according to what will pass through it.

Step 3: give context

An assistant without context answers generically. Prepare a short document with what it needs to know: what the company does, the services, the tone, the main procedures, what is never said. The more precise, the better the answers. Add only information that can be shared with the tool, without customers' personal data.

Step 4: write the rules

One page for the team, with:

• what the assistant is for and not for;
• what is never put in: customers' personal data, health data, sensitive financial information;
• how an answer is reviewed before it is used;
• who decides in case of doubt.

Read ChatGPT and customer data for the care needed with personal data.

Step 5: test with real questions

Before putting it in front of the whole team, ask it twenty real questions and see where it goes wrong. Assistants can make up answers with great confidence. When the answer matters, ask it to say where the information came from, and always check.

Step 6: have an owner

One person looks after the assistant: updates the context, collects complaints, adjusts the rules. Without an owner, the assistant goes out of date and stops being used.

Step 7: measure honestly

After a month, ask the team: how much time does it really save? Where does it go wrong? Where is it not used? If the saving is small, simplify or change task. If it is large, move on to the next.

Common mistakes

• Putting in customers' personal data without thinking.
• Trusting answers without review.
• Starting with many use cases at once.
• Not explaining to the team what it is for.

In process automation, an assistant starts with a concrete task and is tested before it is switched on.

Twenty test questions

Before giving the assistant to the team, write twenty real questions the team asks often, and the right answers according to the company. Ask the assistant the questions, one by one, and mark each answer as right, partial or wrong. For the wrong ones, see whether context is missing (add it) or whether the question is not for the assistant (take it out of scope). When it gets most right and errs without causing problems, show it to the team with the rules. Keep the list: repeat the test whenever you change the context or the tool.

In short

• Start with a single task and a company account.
• Give the assistant context: what the company does, the tone and the procedures.
• Write rules on one page and name an owner.
• Test with real questions and always review the answers that matter.

Quick questions

Do I need a paid tool?

Often yes, above all to have data rules better suited to a company. Check the plan's terms before deciding.

Can the assistant answer customers?

That is another step, with more care: it must say it is an AI and have a way out to a person. Start with internal use.

What if a team member puts in data they should not have?

The written rule is for that. Correct without blaming, see what was shared and reinforce training. If sensitive data is involved, talk to someone who understands data protection.