What AI agents actually change about work
AI agents are not a headcount story. They are a task story. The useful question is not how many people, but which jobs to hand over.
By Mike Beasley, Founder, Mike Beasley & Associates
What this says
- Agents are good at bounded, repetitive, text-shaped work, and bad at judgement, accountability and anything ambiguous.
- The unit of change is the task, not the role.
- If you cannot describe the task as a set of rules, an agent will not do it reliably yet.
There are two ways to talk about AI at work, and both are unhelpful. The first says it will replace everyone. The second says it is a parlour trick that cannot be trusted with anything. The truth is narrower and more practical than either: agents are extremely good at a specific shape of work, and they fail in predictable ways outside it.
If you can tell which shape your work is, you can decide what to hand over. That is the whole skill.
The shape agents handle well
Look for work you would describe to a new starter in a few sentences, where the rules do not change much and the output is text, a classification, or a filled-in field. That is the shape.
- Reading something and pulling structured facts out of it, such as an enquiry email into a name, a need and a deadline.
- Sorting and routing, such as deciding which of four inboxes a message belongs in, and saying why.
- Drafting the first version of something, such as a reply, a summary, a description, or a follow-up that a person then edits.
- Checking consistency: does this record agree with that one, is anything obviously missing, does this look like the other fifty.
Notice what these have in common: they are bounded, repetitive and verifiable. Someone can look at the output and tell in seconds whether it is right. That last property matters more than it sounds.
Where they fail, and why it is predictable
Agents do not fail randomly. They fail at the edges of their instructions, and they fail confidently, which is worse than failing loudly.
- Judgement with consequences. Deciding whether a donation is restricted, whether a complaint needs escalating, whether a refund is warranted. These need someone who can be held to the answer. An agent cannot be.
- Anything genuinely ambiguous. If two people in your team would resolve the same case differently, the rule is not written down, and the agent is guessing. It will guess plausibly.
- Situations that are rare and high-stakes. The cases that matter most are usually the ones that happen least, which means they are exactly the cases with no pattern to learn from.
- Anything where being wrong is invisible. If nobody checks, an agent that is right most of the time is a slow-motion problem, because the failures accumulate quietly.
The unit of change is the task, not the role
This is where most workplace conversations go wrong. “Will AI replace my job?” is not answerable, because a job is a bundle of tasks and only some of them are the right shape.
A development officer’s week includes writing thank-you notes, updating records, preparing reports, and sitting with a donor working out what they care about. Three of those are the right shape for an agent. The fourth is the entire reason the role exists, and no agent is coming for it.
So the useful question is not which roles change. It is: which tasks in this role are bounded and verifiable, and what would the person do with the time if those went away? If the honest answer is “the same work, but we would need fewer people,” that is a real decision to make deliberately. If the answer is “they would finally do the part we hired them for,” that is the better outcome, and it is more common than the first.
A test you can apply this week
Pick one repetitive task. Write down the rules you would give a new starter. Then read them back and ask two questions.
- Would two competent people follow these rules and produce the same answer? If not, the rules are incomplete, and an agent will fill the gap silently.
- Would anyone notice if the output were subtly wrong? If not, you are not ready to automate it yet, not because the technology cannot, but because you cannot tell whether it worked.
If both answers are favourable, that task is a candidate. Start there, keep a human reviewing the output for the first weeks, and watch whether the reviewer is catching real errors or just confirming. That tells you when to stop watching.
What actually changes
Not headcount, in most organisations. What changes is where the human attention goes. The dull, bounded, repetitive middle of the work thins out, and the judgement, relationships and accountability, which were always the point, get more time than they have had in years.
That is a smaller claim than the headlines make. It is also the one that holds up when you try it.
Written by
Mike Beasley
Founder, Mike Beasley & Associates LLC
Mike has spent more than fifteen years leading complex business and technology projects, including senior roles in healthcare and education, and now builds automation, donor systems and client portals for nonprofits, associations and small businesses across Virginia, North Carolina and the rest of the US.