I Stopped Building AI Tools. I'm Hiring Them.

This is what changed.

Three months ago I was building AI tools for myself. A month ago I was building them for my coworkers. Today I’m building them for my bots.

Each phase felt obvious at the time, and each one was wrong in a way I couldn’t see until I moved to the next one. This is what the progression looked like.

The world I was coming out of is the one every finance and data team still operates in. A non-technical manager wants a number. A technical person writes the query, builds the chart, exports the CSV, fields the follow-up. Every question hits a gatekeeper. Every urgent ask queues behind the last urgent ask. The bottleneck isn’t the data. It’s the person holding the key to it, and for a lot of years that person was me.

Phase 1: Building for me

The first version of my AI workflow was me at a keyboard, typing at Claude, getting responses, deciding what to do with them. Every interaction flowed through me. I wrote the prompts. I evaluated the output. I decided what to run next.

When something was missing, my instinct was let me give it another skill. An MCP server for QuickBooks. A tool for reading PDFs. A prompt for reconciling bank feeds. Every addition made the model more capable for me, specifically, sitting at my specific desk, running my specific workflow.

This is where most of the industry is. Copilots, chat windows, tool-use frameworks, “give your AI a thousand skills.” It works. It’s also a dead end, because the person doing the work is still me.

Phase 2: Building for my coworkers

A few weeks in, my non-technical boss wanted to run a recurring financial report himself, the kind of thing he’d normally wait on me for. I’d already wired him up with the MCP servers I’d custom-written for our financial systems, so his laptop could talk to the database. What he was missing was the choreography: which queries, in what order, producing what output.

So I wrote it once as a skill, checked it into a repo called finance-skills, and sent him a message that said, basically, first time: git clone …/finance-skills. Every time after that: cd finance-skills && claude, and paste this. The prompt told Claude to read the repo’s CLAUDE.md, run the right skill for the last three months, collect per-customer detail, produce a summary table with a totals row, and export everything to CSV. If a customer tripped a manual-review flag, the output explained why so he could eyeball those rows himself.

That’s the entire handoff. A guy who does not write code, pulling a multi-month financial analysis by typing claude in a terminal and pasting one prompt.

Pulling the report was only the start. In the old world, the analysis ended at the query or the BI dashboard. If he wanted to know “what does this look like if sales jump 10% next quarter?”, that was another ticket for me. Now he asked Claude the follow-up directly, got a modeled answer, asked another, asked another. I wasn’t the gatekeeper for any of it.

It isn’t the end state, though. He still has to run the analysis. I still have to maintain the data underneath it.

The AI had been democratized. I had not.

Phase 3: Building for my bots

This phase is still in progress. What follows is the direction, not the finished thing.

The shift was small and it didn’t feel like much when it happened. I stopped asking what can the AI do? and started asking what does this agent need in order to own its job?

The trigger was simple. I get bored doing the same thing twice. Same thing you’d tell a teammate: you’ve seen this one, come back to me if it’s different.

That’s a different question. “What can AI do” is open-ended and flattering to the model. “What does an agent need to own its job” is boring and operational. It’s the same question you’d ask when you hire a person. What’s the scope? What systems does it need access to? Who does it escalate to? What happens when it’s wrong?

That last question opens the next layer. Owning a job is one thing. Doing it without burning down the company is another.

The discipline of “not a thing more”

Running these agents independently is where it gets serious. They need to be secure and permissioned like any employee with access to money. I wouldn’t hire an outside accountant and give him wire authority on my operating account. He doesn’t need it for the job, and giving it to him anyway is how trouble finds you.

These things try really hard. They will sometimes find a way. I don’t need my agent reconciling equity with a plug JE. That was my job as an investment banker at 3am, and it only worked because I hoped my associate didn’t find it. An agent that’s “trying its best” is exactly the agent that will do the same thing and write a polite note explaining why it was the cleanest path.

The management piece that took me longest to internalize is that a good boss doesn’t over-resource. You don’t give a new hire the CEO’s calendar, the bank password, and the entire codebase on day one. You scope the role, provide the tools that fit the role, and leave everything else gated.

Digital employees work the same way. An AP agent needs access to the accounting system, the vendor list, and an escalation path to the controller. It does not need access to payroll, the operating account, or the CEO’s inbox. Not a thing more is how I think about it now: enough to do the job, nothing beyond it.

This is usually framed as a security principle, and it is. But it’s also a clarity principle. An agent with a scoped role does better work than one with access to everything, for the same reason a specialist outperforms a generalist on their own turf. The narrower the lane, the sharper the output.

Where this is going

You’re probably taking a run at this now. You’ll ship a skill, then a few more, and eventually notice your AI has access to things it shouldn’t.

I went from let me add a skill to this just needs to be running on its own. That sentence is the entire difference between a tool and an employee. Tools are picked up and put down. Employees show up every day, own an outcome, and answer to someone when the outcome doesn’t happen.

Three months ago I was building AI tools. Today I’m building a team. Next year the company doesn’t have more AI. It has more staff.