How do you create an AI agent for your business?

What job should the agent do first?

Pick the smallest job that would genuinely help. OpenAI's guide says teams typically do better with an incremental approach than by building a fully autonomous agent straight away, and advises starting small, validating with real users and growing capabilities over time. ^[1] Anthropic gives similar advice: find the simplest solution that works and add complexity only when needed. ^[2]

Good first jobs for a small business are answering frequent questions, collecting inquiry details, and recording appointment requests for staff to confirm. Poor first jobs are anything involving money, custom quotes, legal or medical advice, or commitments your team would have to honor.

What knowledge should you give the agent?

Give the agent facts, not marketing copy. A useful starting set covers:

• Business name, address or service area, phone, email and website.
• Opening hours, including holiday and seasonal changes.
• Services you offer, and just as important, services you don't offer.
• Published prices or price ranges, only if you are willing to stand behind them.
• How booking, quotes, payment, cancellations and refunds work.
• Licensing, insurance or certifications you can document.

OpenAI recommends building agent instructions from documents you already have, such as operating procedures, support scripts and policy documents. ^[1] Anthropic's documentation suggests explicitly telling the model to use only the information provided rather than its general knowledge. ^[3] Keep a note of where each fact came from and when it was last checked.

If you already keep your details in one place, such as a website or a digital card, that can serve as the agent's knowledge base. On Bryno, for example, the AI agent answers from your card's details and what you add and approve, so what you write there is what it knows.

What rules and guardrails does the agent need?

Instructions tell the agent how to behave; guardrails catch what instructions miss. OpenAI's guide advises making every instruction correspond to a specific action and anticipating edge cases, such as a customer who gives incomplete information or asks something unexpected. ^[1]

OpenAI also describes guardrails as a layered defense, with examples that include a relevance check to keep answers on topic and a filter that stops unnecessary exposure of personal information. ^[1] Anthropic recommends explicitly giving the model permission to say "I don't know". ^[3]

Which actions should require human approval?

OpenAI suggests rating each tool an agent can use as low, medium or high risk, based on factors such as read-only versus write access, whether the action can be undone, and financial impact, with those ratings triggering automated steps, such as pausing for guardrail checks before high-risk actions or escalating to a human if needed. ^[1]

If your agent connects to other software through the Model Context Protocol, the specification states that hosts must obtain explicit user consent before invoking any tool, and acknowledges that the protocol itself cannot enforce its security principles; and says the applications built on it should. ^[4]

A sensible default for a small business: let the agent read and answer freely from approved facts, let it send inquiries to your inbox, and require a person for anything that books, charges, cancels, refunds or changes a customer record.

How do you test the agent before customers see it?

Anthropic recommends extensive testing in sandboxed environments, along with appropriate guardrails, and stresses careful tool documentation and testing. ^[2] OpenAI suggests setting up evaluations to establish a performance baseline. ^[1] For a small business, a simple written test is enough to start:

1. Collect 20 to 30 real questions from emails, calls and reviews.
2. Add questions the agent should refuse or hand off, such as a request for a discount.
3. Add questions about things you don't offer, to check it doesn't invent them.
4. Try a message that tells the agent to ignore its rules, and confirm it doesn't.
5. Compare every answer with your approved facts and note each error.
6. Ask the same questions more than once, since answers can vary between runs.

Anthropic's documentation notes that inconsistencies between repeated answers to the same prompt can indicate hallucinations. ^[3] Fix the knowledge or the rules, then run the test again before launch.

How should handoffs to a human work?

OpenAI's guide calls human intervention a critical safeguard and names two common triggers: the agent exceeding a limit on retries, for example failing to understand the customer after several attempts, and high-risk actions. ^[1]

Decide who receives handoffs, through which channel, and what the agent passes along: the customer's question, their contact details and anything already answered. The agent should tell the customer what happens next in honest terms, and should not promise a response time your team can't keep. Using handoffs to capture new business is covered in how AI agents can generate leads.

How do you keep the agent current?

An agent is only as accurate as its latest update. Change its knowledge the same day you change hours, prices, staff or services, and review the full fact list on a fixed schedule, such as monthly.

OpenAI recommends adding new guardrails based on real-world edge cases and failures as you find them. ^[1] Anthropic advises always validating critical information, because hallucination-reduction techniques don't eliminate errors entirely. ^[3] Keep a short log of wrong or awkward answers and turn each into a fix.