Your next hire may be half person, half AI

I was testing a skill (think of it as a set of instructions or an SOP for an agent) this week to show my AI for Business Owners Cohort why skills are crucial. You can’t rely on generic chat prompts with AI. You need to do a little more work up front.

For this skill, you give it a role you are thinking about hiring for: bookkeeper, client coordinator, marketing person, etc.

It writes the job description, then breaks the job apart and estimates which pieces can be automated, which pieces can be AI-assisted, and which pieces still need human judgment.

The bookkeeper example made that obvious. The prompt was ordinary. A small medical consulting firm, 200 QuickBooks transactions a month, and an owner spending five hours a week doing the books. That is exactly the kind of problem owners live with for years. It is annoying enough to drain attention, but familiar enough that nobody stops to redesign it.

The old answer would be simple: hire a bookkeeper.

The better answer is slower. What is the work? Which parts are repetitive? Which parts are judgment? Which parts carry risk if they are wrong?

That is where AI starts to matter.

I used Anthropic’s skill-creator skill to make sure I got the right instructions. Skill creator allows you to test a generic prompt you might write up yourself for ChatGPT against a more formalized skill SOP.

With the skill, the outputs passed every check across three roles: bookkeeper, client coordinator, and marketing generalist. Without the skill, the average pass rate dropped to 80%. I put all the results up on the webpage so you can review (the content takes a few seconds to load).

The misses were instructive. The baseline AI could still write a reasonable job description. It could still name some AI and automation tools.

But it often skipped the useful structure. It failed to give each responsibility its own automation verdict. It missed implementation difficulty and its recommendations for tools weren’t thorough. In the marketing role, it missed the overall automation score entirely.

That is the difference between impressive output and “meh.”

An owner does not need a prettier job description. The owner needs to know whether they are hiring a person, buying software, building an agent, or redesigning the role around a smaller human workload.

A generic model will often blur them together. A skill forces the model to slow down and make the distinctions explicit.

The broader data is pointing in the same direction.

Goldman Sachs surveyed small business owners earlier this year and found that 76% are already using AI. Most say it is helping. But only 14% say AI is fully embedded in core operations. That gap feels familiar. Owners are experimenting. The business has not changed yet.

TriNet and Harvard Business Review Analytic Services found a similar gap. Most SMB respondents expect to increase AI use, but only 19% feel highly prepared to recruit or develop the AI skills they need.

That is the problem I keep seeing and I’ve been working to address it in my AI for Business Owners Cohort. I also wrote last week about my program to train college students to become AI builders.

The issue is rarely whether AI can do something useful. It can. The issue is whether the business can define the work clearly enough, deploy the right skill against it, and keep improving the output until it becomes reliable enough to trust.

BCG made the enterprise version of this point bluntly. Only 5% of companies are generating sustained P&L impact from AI, while roughly 60% have seen little or no material benefit.

That is exactly where small businesses feel pain. Bookkeeping. Client follow-up. Reporting. Scheduling. Marketing production. CRM updates. The unglamorous work that decides whether the business runs cleanly or constantly leaks owner attention.

AI does not give deterministic output.

That matters more than most owners realize. You can ask the same model a similar question twice and get answers that feel close but behave differently. Sometimes the difference is the whole decision.

In this eval, the skill did not make the model more creative. It made the model more disciplined.

It forced a repeatable frame: role summary, responsibilities, tools, automation verdicts, implementation difficulty, cost comparison, caveats, and recommendation.

That is what skills are for.

A skill is an operating wrapper. It captures how you want the AI to think, what it must not skip, what tradeoffs matter, and how the answer should be evaluated. Then you run it, inspect the output, notice what it missed, and improve the skill.

That loop is the work.

Most owners want AI to be a finished employee. It is closer to a capable employee who needs a clear role, examples, review, and feedback.

I keep coming back to the bookkeeper.

Five hours a week of owner time looks like a hiring problem. But when you break the role apart, some work is already handled by QuickBooks rules. Some can be handled by AP and invoicing tools. Some needs monthly human review.

So the question changes.

It is no longer, "Should I hire a bookkeeper?"

It becomes, "What parts of this role should exist at all, what parts should be automated, and what human judgment do I actually need to buy?"

That is a much better set of questions.

And these are the questions small business owners need to ask across the company now. Every role is becoming a bundle of tasks, tools, agents, judgment points, and feedback loops.

I am going to keep building these skills because this is where AI becomes practical for small businesses. They help you see the work more clearly, redesign it, and keep making the system smarter.

Alan

PS I’m finishing up the current AI for Business Owners Cohort in another week. I’ve been thinking about offering another Cohort in June and/or a more advanced Claude Code/Codex Cohort or an implementors class meant for the AI lead at your company. I’m trying to gage interest so let me know here.