A small accounting firm asked us to build 'an AI agent'. After twenty minutes it was clear they meant three different things, and that only one of them, answering the phone during tax season, would pay for itself immediately. This is the process we used to get there, and to go from that decision to a live agent in three weeks.
Key takeaways
- Pick one conversation or process that is repeatable, high-volume, and currently costing you customers or hours.
- Gather the knowledge the agent needs before building, which means your prices, your hours, your policies and the questions you hear every day.
- Test against real calls and messages, run a shadow period, and launch with clear escalation rules.
Step 1: Choose the job by cost, not by excitement
List the conversations and tasks your team repeats. For each, estimate how many happen a week, how long they take, and what it costs when they are missed or late. The first agent should take the job with the biggest number, which is usually inbound calls or website enquiries for small businesses, and support tickets or document processing for larger ones.
Step 2: Write down what the agent needs to know
An agent is only as good as its knowledge. Collect your services and prices, hours and locations, policies, the twenty questions you hear most, and the rules for what to do when something unusual happens. Much of this exists in your team's heads; getting it on paper is half the build.
Step 3: Decide what it may do, and what it must hand over
Booking an appointment: yes. Issuing a refund over a threshold: ask a person. Giving medical or legal advice: never. Define these boundaries early. They shape the design and make the agent trustworthy to your team.
Step 4: Connect the tools that make it useful
An agent that cannot write to your calendar or CRM just creates a message for someone to act on. Connect it to the systems where the outcome needs to land, so it reaches the calendar, the CRM, the help desk and the order system. Platforms such as n8n make this practical without heavy development.
Step 5: Test on reality, then launch in stages
Feed the agent real past calls and messages and review its answers. Then run it in shadow mode, where it proposes but does not act, for a week. Launch on a single channel or location first, watch the transcripts, and expand once the edge cases stop appearing.
If you want help, this is exactly what a free plan review with us covers, from choosing the job to scoping the knowledge and estimating what the agent will save.
Common questions
How do I build a custom AI agent for my small business?
Choose one high-volume conversation or task, document the knowledge and rules the agent needs, decide what it may do alone and what it must escalate, connect it to your calendar, CRM, or help desk, and test on real conversations before launching in stages. Most businesses partner with an AI agent development company for the build.
How long does it take to build a first AI agent?
A focused agent such as an AI receptionist or website chat agent typically takes two to four weeks from discovery to launch.