You have probably seen the job ads. “Head of AI”. “Agent Orchestration Engineer”. “Prompt Operations Lead”. The salaries are real. The work behind those titles is real too. And if you run a 10-person firm, understanding what those people actually do is more useful than the job title suggests, because the same problems exist at your scale, they just dont come with a dedicated hire to fix them.

Why this matters to you

Large companies are not hiring these roles because AI is fashionable. They are hiring them because someone has to own the plumbing: the connections between tools, the rules that stop an automated process doing something expensive and wrong, the checks that keep outputs accurate. Without that ownership, AI projects quietly fail or quietly cost money. At your size, you cant absorb a six-figure salary to solve this. But you can understand what the job is, so you can ask whether the problem is costing you anything, and whether it is worth fixing.

The five roles, translated

1. AI Product Manager

What they do: decide which business problems are worth solving with AI, in what order, and how to measure whether it worked. At a big company this is a full-time job because there are dozens of competing ideas and no natural filter. At your firm, this is a question you answer once every few months: “Is there a task we do repeatedly that costs us more time than it should?” The role is not technical. It is prioritisation. You probably already do it, you just dont call it that.

2. Prompt Engineer

What they do: write and maintain the instructions that tell an AI model what to do, what to ignore, and how to format its output. The word “engineer” sounds grand. The reality is closer to writing a very precise brief for a new member of staff who follows instructions literally and has no common sense. A prompt that works reliably is specific, gives examples, and tells the model what failure looks like. A prompt that doesnt work produces inconsistent output, which costs someone time to fix. At your scale, one well-written prompt on a repeated task, say summarising supplier emails or drafting a quote from a form submission, can return an hour a day to someone who needs it.

3. Agent Orchestration Engineer

What they do: build and manage systems where multiple AI steps run in sequence, each one passing its output to the next, with rules that catch errors before they compound. “Orchestration” just means conducting the steps so they happen in the right order. Think of a process where an incoming document is read, key figures are extracted, those figures are checked against a threshold, and a summary is sent to the right person. Each step is simple. Connecting them reliably, and making sure a failure in step two doesnt silently corrupt steps three and four, is the actual job. For a small firm, this is the difference between an automation that saves 30 minutes per document and one that creates a mess someone has to unpick on a Friday afternoon.

4. AI Safety and Evaluation Lead

What they do: test whether AI outputs are accurate, consistent, and safe to act on, then build checks that catch problems automatically. At a large company this is a compliance function. At your firm it is a simpler question: before you trust an automated output, what would have to go wrong for it to cost you money or a customer? Work that answer out, then build the check around it. An example from work we have done: a document-extraction build for a supply-chain operator includes a step that flags any extracted figure that falls outside an expected range, so a human reviews it before it moves downstream. Thats the whole job, scaled to fit.

5. AI Operations Manager

What they do: keep running automations running. Monitor them, fix them when an upstream tool changes its format, update prompts when outputs drift, and report on whether the time saving is still real. This is the role most small firms forget to budget for when they build something. An automation is not a one-time cost. It needs someone to notice when it stops working, because it will stop working, and the cost of not noticing is usually the problem you built it to solve coming back quietly.

What a 10-person firm actually needs

Not five people. Not one person. A clear answer to three questions:

  • Which repeated task is costing us the most time or the most errors right now?
  • If we automated it, what would have to go wrong for it to cost us more than it saves?
  • Who checks that it is still working in three months?

Work those out on your own numbers. Take the task. Count how many times it runs per week. Multiply by the time it takes and the rough hourly cost of the person doing it. That is your ceiling, the most the problem can be worth fixing. If the ceiling is meaningful, the conversation about fixing it is worth having. If it isnt, it isnt.

The companies paying six figures for these roles have problems at a scale that justifies the headcount. You probably have the same problems at a scale that justifies a build, a good prompt, and someone checking it once a month. The gap between those two things is smaller than the job ads make it look.

Want to know which of these applies to you?

We can look at one repeated process in your business, map where the time goes, and tell you honestly whether it is worth automating and what the sensible check would be. No commitment, and easy to say no to. If you want to start there, here is what we do.