The Owner's Checklist for AI Agents: Seven Decisions for a 15 to 300 Person Firm

McKinsey recently shared a short checklist for CEOs who want to run their companies on AI agents. It is written for large enterprises, and even they are finding the shift hard. In its October 13 article on the AI operating system, McKinsey reports that 40 percent of respondents at large companies in its 2026 State of AI survey have scaled AI agents. Only 6 percent credit AI with at least 5 percent of EBIT. A July issue of its CEO Shortlist newsletter adds two findings from a recent survey. Ninety-three percent of respondents went over their AI budgets, and about 60 percent of what an agentic task costs is spent reviewing and fixing the agent's first attempt.

A firm with 15 to 300 people faces the same decisions with less room for error. The CEO is often also the person who signs payroll and knows every client by name. Here is how I would apply the checklist to an insurance agency, a CPA practice, a staffing desk or a freight brokerage of that size.

1. Pick the work before you pick the tool

McKinsey tells CEOs to decide how the business will compete before choosing its technology. For an owner-run firm that becomes a simpler question: which task eats the most staff hours every week, and how many of those hours follow a pattern a person could write down?

Count it before you buy anything. Take the requests that arrive in a normal week, multiply by the minutes each one takes, and estimate the share that is routine. At an insurance agency a likely candidate is certificate of insurance requests. At an accounting firm it may be chasing clients for missing tax documents. Then compare the value of those hours, after the time someone will spend checking the agent's work, with the cost of a build and its monthly fee. If the routine share is small, a better template may do more good than an agent.

2. Give every agent a named owner

The checklist asks CEOs to settle who is responsible for each agent capability. In a small firm the answer should be a person, and it should be the person whose team does the work today. The office manager owns the intake agent. The billing lead owns the invoice audit agent. IT, or your outside IT provider, keeps the connections running, but the business owner decides what good output looks like.

That owner approves the agent's first drafts at launch and reads its exception queue every day. They also answer the question nobody likes asking: when the agent gets something wrong, who catches it and who tells the client? Write that down before the agent goes live. If nobody wants the job, the workflow is not ready.

3. Measure staff hours and errors

Agent dashboards are full of activity numbers such as prompts sent and documents read. None of them tells you whether your team got time back. Compare weekly staff time on the task before and after, counted the same way both times, and include the time people spend checking and correcting the agent. Track errors separately, so a faster process cannot hide worse results.

Our own sprint is judged on hours. If our five-day sprint does not find 10 staff hours a week on the workflow, the fee is refunded. That is a test of the opportunity before anything is built, and the build then has its own acceptance criteria. If a vendor cannot tell you how they will count hours given back, treat their savings estimate as marketing.

4. Budget for checking, and match the model to the cost of a mistake

The newsletter's finding on review costs is the one owners should take most seriously. If most of the cost of an agent's task goes into checking its answer, put review time in the cost estimate and name the person who does it. Decide which fields a person always checks before anything goes out. Then decide which cases the agent should refuse and hand straight to a person.

McKinsey suggests treating AI capacity the way you treat capital, spending more where the return justifies it. In a small firm that means testing candidate models on real examples from the workflow before choosing one. For a certificate request, check whether each model gets the additional insured right, and require a person's approval wherever an error could affect coverage. Where fixed rules handle a step reliably, such as sorting incoming email into folders, simple automation is enough.

5. Check the software you already pay for before you build

The checklist also asks CEOs to rethink what they build themselves and what they get from vendors and partners. Owner-run firms face a narrower version of the same choice. Many agency management systems and practice management tools now include AI features. Where one of those features does the job inside one system, use it. It usually costs less to start, and the vendor maintains it.

Custom work earns its place when a task crosses systems: an email inbox, a carrier portal, a spreadsheet the team keeps on the side, and the system of record. Follow one request from the moment it arrives to the moment it is done. Note every place someone retypes information and how long each handoff takes. Those notes tell you whether a custom connection is worth paying for.

6. Keep the context you own

McKinsey's October 13 article puts the rule plainly: "Own the context; rent or replace the technology." Models change every few months. What makes your firm good lives in your templates, your past answers to clients, the exceptions your senior people know by heart, and the rules each large client has given you over the years.

Keep that material in your own Microsoft 365 or cloud tenant, in plain formats you can move. When you hire a partner to build an agent, the prompts and rules they write for your workflow should belong to you. If you switch vendors next year, you take the context with you and rent the next model.

7. Start with one workflow and a fixed price

Large companies can fund a platform first. A 40-person firm should start with a workflow whose current staff time it can measure, and prove that workflow before funding the next. At launch a person approves everything that leaves the building. Approval loosens only for a narrow, tested category, with the owner's sign-off in writing.

For owner-run firms, our offer starts with a five-day AI Workflow Sprint for $2,500. We watch how the work runs today and build a working prototype on sample documents. Then we count the hours it could give back. If we do not find 10 staff hours a week, you get a full refund, and if you go ahead, the fee comes off the build. A build starts from $7,500 per workflow, half to start and half when the acceptance criteria pass, and running it costs $1,200 a month. You can see how this looks for your industry on our industry pages, or start with the insurance agency section for a worked example built around certificate requests.

The short version for owners

Before you buy an agent, write down the task's weekly hours and who will own it. Decide how you will count hours given back and who checks the work. Keep your own context portable, and prove one workflow before you fund the next. Then ask any vendor, us included, for a scope with written acceptance criteria.

Sources

• McKinsey, "The AI operating system: A new way for CEOs to run the enterprise," October 13, 2026
• McKinsey, "The CEO Shortlist: The price of intelligence, the making of expertise," July 24, 2026
• McKinsey, "The CEO checklist for building an agentic operating model," social media post, 2026