How to create an AI agent for your business
Seven steps from one repeated job to an agent your team trusts. Most of the work happens before anyone writes code.
Free, with nothing to buy at the end. You leave knowing which job to give your first agent.
- 4 AI systems live
- 194 systems live in total
- First system live in 4 to 6 weeks
Published
On this page
How AI agents work
An AI agent takes an input, such as an email or a form. It follows the steps you wrote and uses only the tools you allow. It stops where a person needs to decide.
Step 1: Pick one job that repeats every week
One job with a clear start, a clear finish and a known weekly cost in hours.
What to do
List the jobs your team repeats each week and pick the one with the most hours and the fewest judgment calls.
What good looks like
Answering routine questions from your own documents, typing up call notes, drafting replies to reviews.
Common mistake
Starting with a job that changes every time. An agent needs a pattern to follow.
Tool or example
Not sure which job comes first? Read AI agents for small business.
Step 2: Write the workflow the way you would train a new hire
A written workflow a new team member could follow without asking you.
What to do
Write each step in order, what the input looks like and what done looks like.
What good looks like
Ten numbered lines or fewer, with the exceptions written next to the step they affect.
Common mistake
Keeping the rules in your head. The agent follows only what you wrote down.
Tool or example
A shared document or a flowchart is enough at this stage.
Step 3: Give it your documents and data
Every answer the agent gives comes from your own material.
What to do
Collect the price lists, policies, past replies and records the job draws on, and remove anything out of date.
What good looks like
One folder or one database the agent reads from, kept current by its owner.
Common mistake
Feeding it everything you have. Old and conflicting files produce confident wrong answers.
Tool or example
Your help desk articles, your CRM records and your last fifty good replies.
Step 4: Choose how to build it
A workflow builder for a single job, custom code when the job crosses several systems.
What to do
Count the systems the job touches. One or two usually fits a workflow builder.
What good looks like
The agent can reach every account it needs and nothing more.
Common mistake
Picking the tool first and bending the job to fit it.
Tool or example
Workflow builders such as n8n, Zapier and Make, for a job that touches one or two systems.
Step 5: Decide where a person approves
The agent drafts and a person approves anything that leaves the company.
What to do
Mark every step where money, a customer or a promise is involved and put a person there.
What good looks like
Approvals take seconds: the draft arrives complete.
Common mistake
Letting the agent send to customers before the tests are in.
Tool or example
A draft in your inbox or a task in your project tool, waiting for one click.
Step 6: Test it on last month's real cases
You know how often it gets the job right before a customer sees it.
What to do
Run the agent on twenty to fifty past cases and compare each result with what your team actually did.
What good looks like
Every miss traced to a step or a document, then fixed.
Common mistake
Testing on three easy examples and calling it ready.
Tool or example
A simple sheet with the case, the agent result, your team's result and a match column.
Step 7: Go live small and count the hours
One team, one month, the hours saved set against the time the build took.
What to do
Switch it on for one team or one inbox, keep the approvals and review the results weekly.
What good looks like
Approvals get faster each week and the team stops checking every line.
Common mistake
Rolling it out to everyone on day one.
Tool or example
Log each run with its time saved. Pick the next agent from that log.
When to hand the build to a team
We build and test the first agent for you when the job crosses more than two systems. The same holds when a wrong answer would cost you a customer.
Have the first agent built for you.
We built four autonomous AI systems for a national services business with 84 service contractors.
Skip the trial runs
Have your first AI agent picked for you.
One free 30 minute call starts the AI readiness assessment. About a week later you see your jobs ranked and the first agent to build, with nothing to buy.
- What you get
- A free AI readiness assessment with your three best AI systems ranked by payback.
- How it is priced
- The assessment is free. We quote after the first call.
- What you bring
- A list of the tools you use and read access to one or two.
Prefer to write? Send us a message.
Read next
How to create an AI agent: questions before you start
Do I need to code to build an AI agent?
No, a workflow builder covers most single jobs. Code makes sense when the agent works across several of your systems.
Why do AI agents fail?
Most fail on a vague job or out of date data. Steps 1 and 3 catch both before you build.
Can an AI agent send emails and update our CRM?
Yes, once it has access to those accounts. Keep a person approving anything that goes to a customer until the test results show it gets the job right.
What does it cost to build an AI agent?
Price follows the number of systems the job touches and the testing it needs. The free assessment sets the scope, then we quote.
How long until a first agent is live?
A first system goes live in four to six weeks. That time includes testing.
Can I make my own AI assistant the same way?
Yes, an assistant is an agent whose job is answering your team's questions from your own documents. Steps 1 to 3 do most of the work.