BYD TechThe short answer
To build an AI agent for a business, start with one narrow, repetitive job, collect real examples of it, and choose the simplest design that works, which is often a fixed workflow and not a free-roaming agent. Connect only the tools it needs, with limits enforced in code. Test it on past cases before it meets a customer, and launch with a person approving its work until the results earn more trust.
Key points
- Pick one job. An agent that does everything does nothing reliably.
- Real past examples are the raw material. Collect them before choosing a model.
- The simplest design that works is the right one. Many jobs need a workflow, not an agent.
- Limits belong in code, where a clever prompt cannot talk past them.
- Measure on real history, then launch with a human approving.
Step 1: Pick one job worth automating
Good first jobs share three traits: they happen often, they follow a pattern, and a mistake can be caught before it does harm. Triaging support tickets, drafting standard replies, and copying data from documents into a system all qualify. Negotiating a contract does not.
Write the job down as a sentence a new employee could follow, including what a good result looks like. If you cannot describe the job clearly to a person, a model will not do it well either.
Step 2: Collect real examples
Before any technology decision, gather a few hundred real past cases with the correct outcome for each: tickets and the reply that was sent, invoices and the fields that were entered. This set does three jobs. It shows how messy the real inputs are, it becomes your test, and it tells you honestly whether the job is as repetitive as it feels.
Step 3: Choose the simplest design that works
Anthropic’s guide to building agents draws a line between workflows, where the steps are fixed in code and a model handles individual steps, and agents, where the model decides its own steps. Its advice is to use the simplest option that solves the problem, and to add autonomy only when it clearly helps.
| Design | Use when |
|---|---|
| A single model call with good instructions | The job is one step: classify, summarize, extract. |
| A workflow of fixed steps | The steps are known in advance and rarely change. |
| An agent that chooses its tools | The path differs case by case and cannot be scripted. |
More autonomy means more that can go wrong and more to test. Most business jobs are well served by the first two rows.
Step 4: Connect tools, with limits in code
An agent becomes useful when it can look things up and take actions: read an order, update a record, create a ticket. Give it only the tools the job needs, and prefer read-only access wherever writing is not required.
- Enforce limits in code: a refund cap, a list of records it may touch, a rate limit.
- Describe each tool clearly. The model chooses tools by reading their descriptions.
- Log every action, with the reason the model gave.
- Ask for confirmation before anything that cannot be undone.
A standard way to expose tools is the Model Context Protocol. See What is an MCP server?.
Step 5: Test on history before launch
Run the agent against the examples from step 2 and compare its output with what a person actually did. Read every miss. Group the misses by type, because that tells you which categories of work it is ready for and which should still go to a person.
Keep this test set and rerun it whenever you change the instructions, the tools or the model. Without it, every improvement is a guess.
Step 6: Launch with a person in the loop
- Shadow mode: the agent drafts, nothing is sent, and you compare.
- Approval mode: the agent drafts, a person approves or edits, and edits are recorded.
- Partial autonomy: the agent handles the categories it has proved itself on, and hands off the rest with a summary.
Set a spending cap on the model account from the first day, and watch cost per case alongside quality. This is the order we follow in AI agent development: one job, tested on your real tickets, with a human handoff built in.
Sources
Checked when this guide was written.
