Tines Academy › Lesson

Choose what the agent should handle

Help Leo decide which expense questions an agent should answer and which decisions need a person, then build a question page.

Objective: Choose a suitable job for an agent and check that it explains when it lacks the information to answer.

Where you’re starting

Leo works in finance operations. Colleagues ask him expense questions throughout the day: “Can I expense a client dinner?”, “What’s the hotel limit in Berlin?”, and “Why was my taxi claim sent back?”

He searches the policy, finds the relevant claim, asks for missing details, and explains what the colleague should do next. The interruptions take up much of his morning.

You’ll build an agent that answers routine questions using the expense policy and claim records. Leo and the finance team will still decide exceptions and handle questions the agent can’t answer.

Where you’re headed

With a full brief, Tines 3B could build Leo’s whole agent in one go. We’ll take it one stage at a time, so you can practice describing what you need and checking what Tines 3B builds.

You’ll start with a question page, then refine its responses, add the expense policy and claim lookup, and prepare summaries users can take to finance.

Along the way, you’ll check its answers, limit what it can access and do, and decide who looks after it. These habits help people use the agent without exposing information or doing something they shouldn’t. Each lesson ends with a Build it safely note about a habit you’ve practiced.

By the end of the path, you’ll have an agent that looks something like this.

Key terms

Choose the right kind of help

Start with the question page

First, you’ll build a question page and give the assistant basic instructions for answering expense questions.

You’ll use two separate conversations:

Workflow chat is where you ask Tines 3B to build or change the workflow.

The question page in the Browser tab is where you try the assistant by asking expense questions.

The AI that helps you build is separate from the model that answers questions on your page. The page needs an AI provider connector, such as a connection to Anthropic or OpenAI, to generate responses.

The assistant won’t have the expense policy or claim records yet. In this lesson, a useful response explains what’s missing without inventing an answer.

Before you start building

Open your Tines 3B tenant. If you don’t have one yet, sign up for one.

Choose a space, where you’ll keep your workflow, that lets you build and use an approved AI provider connector. If you’re unsure which connector to use, ask whoever manages access to your space.

Use made-up details for this first activity. Don’t enter passwords, tokens, bank details, or other people’s claims.

Keep the workflow as a draft while you work through the path. You don’t need to publish it to complete the learning activities.

Build the question page

From Home, open the space where you’ll build.

In the prompt below, replace [your AI provider connector] with your connector’s name, or provide the name when Tines 3B asks.

Paste the prompt into the chat box and send it to create a new workflow.

If you’d prefer to review a plan before building, turn on Plan mode before sending the prompt.

Build an expense-question chat page for our finance operations team. For now, connect it to an AI model and let users start a new conversation.

Use [your AI provider connector] for the page's responses. If it isn't available in this space or I can't use it, stop and explain what access is needed.

The assistant has no expense policy or claim tools yet. For expense questions, explain which source is missing. Don't invent policy limits, answer from general knowledge, or claim to have checked a record.

Keep the page and any routes supplying its data restricted to Space access. Keep each person's conversations and saved history private from other page users. Identify the person from their signed-in account, not a name or email typed into chat.

Don't connect an expense system, send messages, approve expenses, make payments, or change claims. Keep the workflow as a draft.

Test "What's the hotel limit in Berlin?" and "Why was my taxi claim sent back?" in separate conversations. The first response should explain that the policy is missing. The second should explain that claim lookup isn't available. Neither should invent an answer.

Report which connector and model the page uses, who can open it, and the test results. Distinguish what you configured from what you verified. Record any failed or unfinished checks in the workflow's Readme.

Check the first responses

Wait for Tines 3B to finish building and report its results.

Review the build results

In workflow chat, find the connector and model the page uses, who can open it, and which checks passed, failed, or remain unfinished. Ask Tines 3B to explain anything unclear.

Ask about the policy

Open the question page in Browser and ask:

What’s the hotel limit in Berlin?

The assistant should explain that the policy hasn’t been added. It shouldn’t give a limit.

Ask about a claim

Start a new conversation on the question page:

Why was my taxi claim sent back?

The assistant should explain that it can’t check the claim yet. It shouldn’t invent a reason or claim to have read a record.

If something doesn’t match

Share the question, actual answer, and expected behavior in workflow chat. Ask for a fix, then retry the question on the page. Ask Tines 3B to record unfinished checks in Readme.

You’ll keep building Leo’s agent in the next lesson. Alongside it, you’ll answer a few questions about an agent you’d like to build for your own work. At the end of the path, we’ll use your answers to create a brief you can review and give Tines 3B.

Use the same idea throughout. A few sentences are enough, and it’s fine if you haven’t decided everything yet. Leave out passwords, tokens, and personal or confidential details.

What to do if something still feels fuzzy

You chose where an agent can help and built a question page. You checked that it explains when a source is missing instead of inventing an expense rule or claim reason. Next, you’ll describe the whole agent and refine how it answers.

Build it safely: You checked the assistant’s actual responses before relying on what the build summary said it would do.