Step by step
- Define the goal. One job, one success measure — for example “every new lead gets a qualified reply within five minutes”.
- Map the process. Inputs, decisions, rules, exceptions and who approves what.
- Choose the model. A capable LLM (OpenAI, Anthropic Claude, Google Gemini or open-source), balanced for quality, speed and cost.
- Connect tools and data. APIs for your CRM, calendar, messaging or databases, and approved knowledge through retrieval (RAG).
- Write instructions and guardrails. Role, tone, what it may and may not do, when to hand off to a person.
- Test on real scenarios. Including tricky and adversarial ones; measure accuracy before launch.
- Launch carefully. Start with approvals on important actions and close monitoring.
- Monitor and improve. Review conversations and outcomes, then refine.
Who is needed?
A typical team includes a project manager, an AI engineer and a full-stack developer, with a solution architect, data engineer, UX designer and QA tester as the project requires. See the team behind a project.
Common mistakes to avoid
- Starting with a vague goal (“an AI for everything”).
- Letting the agent act without permissions or logs.
- Skipping tests on real, messy inputs.
- Giving it unapproved or outdated knowledge.
Frequently asked questions
Do I need my own data to build an AI agent?
You need the information the agent should use — FAQs, policies, product data — and access to the systems it works with. Large training datasets are usually not needed.
Should I build in-house or with a partner?
In-house works if you have AI and integration experience; a partner is faster when you do not. Either way, own the accounts and data.