The short version
- Do this: pick one job, start where a mistake costs nothing, point it at work that pays back and widen the leash in stages.
- Not that: don't automate a mess, don't hand it ten jobs at once and don't aim it at customers on day one.
- A simple first two weeks: name the job, run it read-only, let it send the easy ones, then stack the next job.
- A disciplined start matters more than the model or the tools.
A business owner asked us a good question: what are the two or three things I should do first with an agent, and the two or three I should avoid? Here is what we tell them, with no upsell attached.
Most people overthink the start. They picture a giant rollout: every job automated, every tool connected, the whole business handed over at once. That is the version that goes wrong. A good start is small, deliberate and pointed at one thing that matters. Get that right and the agent pays for itself early. Get it wrong and you spend the first month cleaning up.
Do this
Four moves, in order. None of them is clever, and that is the point.
Pick one job
The thing you would do first if you sat down tomorrow. Not a category, a job. "Chase the quotes I sent last week that nobody replied to." "Draft replies to new website enquiries." If you cannot name it in a single sentence, it is not the one yet. A vague brief gives you a vague agent. A sharp brief gives you something you can judge by Friday.
Start where a mistake costs nothing
Internal work first. Point the agent at your own inbox, your own notes, your own follow-up list before it touches anyone outside the business. A wrong move on your own inbox is a shrug. The same wrong move on a customer is a refund, or worse, a relationship. You want the agent to make its early mistakes where the only person who sees them is you.
Point it at work that pays back
Money jobs, not tidy-up jobs. Sales, follow-up, the quote you forgot to chase, the lead that went cold because nobody had time. Those have a number attached. The temptation is to start with housekeeping: sorting files, cleaning up a spreadsheet, tidying the calendar. Those feel productive and were never going to earn their keep. Aim the agent at the work that already makes you money and you will feel the return inside the first month.
Keep it on a short leash, then let it out
Earn the autonomy in stages. First it just reads and tells you what it found. Then it drafts, and you sign off before anything goes out. Then it sends the easy ones on its own. Then it runs without you watching. You widen the leash one notch at a time, as the agent earns your trust on the last notch. This single habit is what separates the people who love their agent from the people who got burned.
Not that
The three starts that go wrong most often. All three are easy to avoid once you have seen them.
Don't automate a mess
A faster mess is still a mess. If your follow-up process is held together by memory and good intentions, an agent will run that broken process faster and more often. Fix the process first, even roughly, so there is something worth speeding up. The agent makes a good system better. It cannot rescue a bad one.
Don't hand it ten jobs at once
One focused agent beats ten half-built ones. When you split attention across ten jobs, you get ten things that almost work and nothing you trust. Get one clean win first, the kind you can point to and say "that saved me Thursday." Then stack the next job on top of it. Momentum compounds. A pile of half-finished automations does not.
Don't aim it straight at customers on day one
Trust is earned on low-stakes work, not granted on day one. Letting a brand-new agent talk to customers before you have watched how it behaves is the fastest way to a story you have to apologise for. Let it prove itself on internal work first. By the time it reaches a customer, you already know exactly what it will say.
Your first two weeks
If you want a concrete shape to follow, here is the one we would run for almost any business.
A simple starting plan
Name the one job and write the brief in a sentence. Connect only the tools that job needs, nothing else.
Read-only and drafts. The agent works your internal version of the job and shows you its drafts. You correct it. It learns your judgment.
Let it send the easy ones. The clear-cut cases go out on their own. The edge cases still come to you first.
Let it run, then stack the next job. Once the first job runs without you watching, you have your template. Add the second job the same way.
Why the start matters more than the agent
Almost every runaway bill, every "this thing went rogue" story, every disappointing pilot traces back to a bad start, not a bad agent. The agent did exactly what it was told. It was told to do too much, too soon, on work that was never clearly defined.
The reverse is also true. A narrow first job, aimed at real money, kept on a short leash and widened as it earns trust, tends to pay for itself well before you have finished setting up the second one. Not which model, not which tools, not how clever the setup is. Just a disciplined start.
Pick one money job. Prove it on your own work. Widen the leash as it earns trust. Everything else is detail.
If you would rather not pick that first job alone, that is what a scoping call is for.
Not sure what to start with?
We will look at your work, pick the first job that pays back fastest and tell you straight if now is the right time.
Book a 15-minute call →