Deciding which work goes to AI, and who checks it

A common AI strategy right now goes like this: buy several hundred licenses, send an all-hands email encouraging everyone to experiment, and wait for productivity to arrive.

It is the gym membership in January. Everyone signs up. The membership is not the fitness.

Our founder, Latif Horst, spends much of his week with chief executives and operations leaders, and the pattern he describes is consistent. A few enthusiasts get real value. Most people use the tool to tidy up emails. And finance starts asking why the software bill went up while nothing else changed.

The license was never the hard part. The hard part is deciding which pieces of the real work go to AI, who checks the result, and what the AI is allowed to touch. It is unglamorous, and it is exactly where the money is.

Why "go and experiment" stalls

Nobody was told what the job is. A general-purpose assistant with no defined task gets used for general-purpose things: drafting, summarizing, rewording. Useful, but rarely the work that moves a number.

The work that matters crosses people and systems. A month-end close, a client onboarding or a proposal involves several people, several systems and several approvals. An individual with a chat window cannot redesign that on their own.

Nobody owns the checking. AI output is usually good and occasionally confidently wrong. If no one is responsible for catching the wrong part, people either stop trusting the tool or, worse, stop checking.

Access is either too little or too much. Either the tool cannot reach the information it needs, so it is useless, or it can reach everything, which is a security problem.

A better way to decide

We treat AI adoption as a redesign of specific pieces of work, not a software rollout. For each candidate piece of work, answer four questions.

1. What is the job, exactly? Name the task, the input, the output and what "good" looks like. "Prepare the first draft of the monthly client report from these three sources, in this format" is a job. "Help with reporting" is not.
2. Which part goes to the AI and which stays with people? Usually the AI gathers, drafts and cross-checks; people make judgments, handle exceptions and own relationships.
3. Who checks it, and how? Name the person and the check. For high-consequence work, use a second model or a checklist as well as a person.
4. What may it touch? List exactly which files, systems and actions the AI may use for this job, and nothing more.

Answer those for the five tasks where time or errors cost you most, and you have a real AI plan. Answer them for none, and you have licenses.

Who checks the AI’s work

There is a popular story that AI will remove the entry-level job. We see it differently.

AI does a lot of the work. Somebody still has to check it, catch the odd result and notice when something looks wrong. That is ongoing work, because you can never fully stop checking. It is too expensive and too repetitive for your most senior people, and it is almost exactly what junior staff used to do.

So we think the junior role is about to come back, in a new form: curious, early-career people who work alongside a handful of AI assistants, review what they produce, spot the problems and escalate what matters. It is Pam at reception in The Office, if Pam had five AI assistants and a spreadsheet.

That is a head start for those people, not a lost generation. They get their hands on powerful tools and real business data years earlier than previous generations did, and they learn how the whole business works by supervising it. Latif is looking to hire graduates and interns for exactly this kind of work.

The companies that plan for this role will get the value of AI with fewer expensive mistakes. The ones that assume the AI needs no supervision will simply make their mistakes faster.

A first 90 days

• Weeks 1 to 3: pick the work. With your operations leads, list the ten tasks where time or errors cost most. Choose three that are frequent, well understood and checkable.
• Weeks 3 to 6: redesign those three. Answer the four questions for each. Set up the access the AI needs and nothing more. Name the checker.
• Weeks 6 to 10: run them for real. Measure time taken, error rate and how often the checker had to intervene, against how the work was done before.
• Weeks 10 to 13: decide. Keep what worked, fix or drop what did not, and choose the next three. Decide whether you need dedicated people for the checking work and write that role down.

At the end, you will know what your licenses are actually for.

Last reviewed: September 25, 2026