A founder’s first AI agents: where to start.

Find the work that repeatedly interrupts the day

For one ordinary week, keep a short record of repeated tasks. Note what starts the work, who handles it, which tools are involved and what counts as finished. Include tasks that take only a few minutes but occur many times: checking a customer's status, collecting an onboarding detail or gathering campaign results for a decision.

Ask your team where work waits, not simply what feels boring. A report that takes twenty minutes may be less valuable to automate than a customer request that waits two days for someone to assemble the information. This gives the project a business reason beyond trying a new technology.

Score the opportunity before building

Our suggested assessment uses four questions. Does the work happen often? Is the process understood? Is the information available? Can a person quickly check whether it succeeded? A candidate that scores well on all four is usually easier to scope than a rare, ambiguous task that depends on unwritten judgement.

Then consider the cost of a mistake. Preparing an internal summary and changing a customer agreement are different responsibilities. You can often begin with preparation and recommendations, leaving the consequential action with a person. This still removes useful work while providing examples of how the system performs.

• Volume: how many times did this happen in a normal week?
• Readiness: can the team explain the process and provide examples?
• Visibility: can success be checked in the destination system?
• Impact: what happens if the action is wrong or late?

Write a one page implementation brief

Describe the trigger, the required inputs and the finished result in ordinary language. For example: when a new client accepts a specific service, collect the required setup information, create the delivery checklist and flag anything missing to the account owner. Identify the business record that proves each step is complete.

Add the permitted tools, actions requiring approval and the person responsible for exceptions. Include examples of a good result and examples that should be escalated. This document becomes the reference for the implementation and for your team's expectations. Without it, a working demonstration can drift into a different project from the one the business actually needs.

Prepare the business knowledge people already use

Gather the information a competent colleague would need for the task: current services, standard responses, operating instructions, templates and the location of the relevant records. Remove obsolete versions and identify who owns each source. A pile of documents is less useful than a smaller set with a clear purpose and review date.

Give the implementation team approved sample cases with unnecessary personal details removed. Include awkward examples as well as ordinary ones. If your people disagree about how a request should be handled, resolve that before expecting the agent to make a consistent decision. Automation will not settle an unclear business policy for you.

Measure the result against a real baseline

Record how the workflow performs before launch. Useful measures include handling time, waiting time, completion rate, staff corrections and the number of requests that need escalation. For sales or marketing work, connect the operational measure to a relevant commercial outcome, such as qualified conversations or customers reaching the next stage.

Consider a purely illustrative calculation. A team handles 100 cases each week, spending 6 minutes per case. That is 10 hours of handling. If an implementation reduces human handling to 2 minutes per case, the difference is 6 hours and 40 minutes before allowing for monitoring and maintenance. That is capacity recovered, not automatically a payroll saving or extra revenue. What the team does with the capacity matters.

Expand only after the first job is dependable

Begin with a limited set of cases and review the exceptions. Look for repeated reasons the agent stops or produces work that needs correction. Fix the source problem, instructions or connection rather than asking staff to compensate indefinitely. Keep a simple way to pause the workflow while a problem is resolved.

When the first scope produces dependable results, connect the next responsibility. Onboarding might lead into delivery preparation. Campaign reporting might lead into creative briefs. Expansion should follow a demonstrated need and a clear handoff, rather than a desire to fill every department with an agent at once.

Make the handover useful to your people

Your team should leave the handover knowing how to brief the system, review completed work, update instructions and deal with exceptions. They should also know who to contact when a tool changes or an issue needs specialist support. Ask for a walkthrough using your own workflow, with time for the team to practise.

A done for you implementation can reduce the work of setup, connection and testing. It should still leave you with understandable responsibilities and a system you can use. The founder remains responsible for business priorities; the agents take on the agreed work, with support available when the setup needs attention.