The slowest part of AI is us.

Short version: the AI tools available today can do far more than most people are asking them to do. That gap isn’t a technology problem, it’s a learning problem, and it gets wider every time a new model comes out. Training on a specific tool goes stale within months. What lasts is a handful of skills that carry from one tool to the next: describing a job clearly, giving the right context, checking the work, and knowing where a human has to stay in charge. Fifteen minutes a week on a real task beats a two-day course once a year.

The tools moved faster than the habits

A couple of years ago, using AI at work mostly meant asking a chatbot to tidy up an email. It was handy, a bit unreliable, and easy to dismiss.

That’s not where things are now. Researchers at METR, who measure what AI systems can actually do, found that the length of task an AI agent can complete on its own (measured by how long the same task takes a skilled person) has been doubling about every seven months since 2019. Their January 2026 update found it has sped up: since 2023 the doubling time is about four months (131 days). Their tests are mostly software tasks, so it’s a rough guide rather than a law of nature. But it isn’t a gentle improvement. It’s the kind of curve where the thing you tried last year and gave up on is now routine.

We see it in our own work. When we had a stubborn box of client DVDs to recover recently, an AI session on a Linux laptop built and tested a small rescue program in about an hour, and it passed all 24 of its tests before it touched a real disc. (The full story is here.) The equivalent job would once have been a week of a person’s time. The AI wasn’t the slow part of that job. The slow part was us: deciding what we wanted, writing it down clearly, and checking the results.

That’s the pattern almost everywhere now. The technology is ready before the people are.

Why people are the bottleneck (and why that isn’t an insult)

Saying people are the weak link sounds harsh. It’s actually just how learning works.

• Software updates overnight. Habits don’t. A new model can arrive on a Tuesday. Changing how you do your work takes weeks of practice, and nobody gets those weeks handed to them.
• First impressions stick. Plenty of people tried an AI tool once, got a confident wrong answer, and quietly decided it wasn’t for them. That was a fair call at the time. It may not be a fair call now, but nobody goes back to re-test a tool they’ve written off.
• Trust is earned slowly. You don’t hand a new staff member the important jobs on day one, and you shouldn’t with AI either. But that means trust grows at human speed, while capability grows at machine speed.
• The tool keeps changing under you. Just as you learn where the buttons are, the buttons move. That’s tiring, and it makes “I’ll learn it properly when it settles down” very tempting. It isn’t going to settle down.
• Nobody owns it. In a small business, keeping up with AI is usually everyone’s job, which means it’s nobody’s job.

None of that is a character flaw. It’s the gap between how fast machines improve and how fast people adapt, and it’s the gap training needs to close.

Why one-off AI training goes stale

The obvious fix is to send everyone on a course. It helps, but less than you’d hope, for one reason: most AI training teaches a tool, and the tool changes.

A session on “which buttons to press in this app” is out of date the next time the app updates. Menus move, features merge, the model underneath is swapped for a better one, and the clever workaround you were taught is no longer needed. What people remember six months later is usually not the steps, it’s whether the session made them more or less confident to keep trying.

So the goal of AI training shouldn’t be to master a tool. It should be to make people good at picking up the next one.

The skills that last from one tool to the next

The good news is that the skills that make someone effective with AI barely change between tools. They’re mostly the same skills that make someone good at delegating to a person.

1. Describe the outcome, not just the task. “Write a reply to this customer” gets you something generic. “Write a short, friendly reply that apologises for the late delivery, offers a free replacement, and doesn’t promise a date we can’t keep” gets you something you can send. The AI can’t read your mind any better than a new staff member can.
2. Give it the context you’d give a person. Who the customer is, what happened last time, what your business sounds like, what you’ve already tried. Most disappointing AI results are a context problem, not a capability problem.
3. Break big jobs into checkable steps. Ask for a plan first, read it, then ask for the work. You’ll catch the misunderstanding at step one instead of at the end.
4. Check the work like it matters. AI is fluent, and fluent isn’t the same as right. Names, numbers, dates, prices and anything legal or medical get checked against the source, every time.
5. Know where the human stays in charge. Decide in advance which jobs the AI can finish and which ones it can only draft. Anything that goes to a customer, moves money or can’t be undone gets a person’s sign-off.
6. Re-test what you gave up on. Keep a short list of things AI couldn’t do for you last time. Try them again every few months. Some of them will have quietly become possible.

That last one is how you stop first impressions from going stale. It’s the cheapest habit on this list and probably the most valuable.

How we do it at Digital Greenhouse

We run a lot of our own business with AI help, and the rule that makes it work is simple: the AI drafts, a check happens, a human decides. Invoices, client emails and website pages are all drafted with AI, go through a quality check, and only then reach John to approve or send. Nothing client-facing goes out on its own.

That isn’t because the AI is bad at the work. It’s usually very good at it. It’s because the human part of the chain, the judgement about what’s right for this client on this day, is the part that actually matters, and it’s the part worth protecting.

A four-week plan for a small team

You don’t need a strategy document. You need a few weeks of deliberate practice on real work.

Week 1: pick one real, boring task. Something everyone does every week and nobody enjoys: writing up meeting notes, answering the same five customer questions, turning a messy spreadsheet into a summary. Do it with AI, side by side with the old way, and compare.

Week 2: write the brief down. Whatever instructions worked in week one, save them as a reusable template: the outcome, the context, the format you want, the things to avoid. Share it with the team. This is where individual tricks become a business habit.

Week 3: decide the rules. Which jobs can AI finish, which can it only draft, and what must never go into it (client details, passwords, anything confidential, unless you’ve checked how the tool handles your data). Write it on one page.

Week 4: run a short show-and-tell. Fifteen minutes. Everyone shows one thing that worked and one thing that didn’t. You’ll learn more from each other’s failures than from any course.

Then keep going, lightly: fifteen minutes a week, one person on the team responsible for trying what’s new, and the “things it couldn’t do last time” list re-tested every quarter.

How to keep up without reading every announcement

There’s a new AI announcement roughly every hour, and following all of them is a job in itself. You don’t need to. A few filters help:

• Wait for the tools you already use to change. Most of the useful new capability turns up inside software you already have: your email, your accounts package, your website builder. When those add something, that’s your cue to try it.
• Ignore the benchmarks, test your own work. A new model scoring higher on a test means very little until it does a better job on your task. Keep one or two of your own standard tasks to try on anything new.
• Give one person the job of noticing. Not an expert, just someone curious, with half an hour a fortnight to try new things and report back.
• Be suspicious of urgency. “Adopt this now or get left behind” is a sales line. Steady, regular practice wins over panic every time.

Where it goes wrong

A few mistakes come up again and again:

• Pasting in what shouldn’t be pasted. Client records, staff details and passwords don’t go into a free AI tool until you know where that data goes and who can see it. Business versions of most tools have stricter settings, so check them.
• Trusting fluent answers. AI can be wrong with complete confidence. If a fact matters, check it.
• Letting it send things. Drafting is safe. Sending, paying, deleting and publishing need a human hand on the button.
• Training once and stopping. The tools will keep moving. If the learning stops, the gap reopens within months.

How we can help

We run hands-on AI workshops for teams, in Microsoft 365 Copilot, Claude or ChatGPT, built around your real work rather than generic examples. They’re designed around the skills on this page, so they don’t go stale when the tools change: writing a good brief, checking the work, and setting the rules for your team. If you’re writing those rules, our guide to writing a team AI usage policy has a free template.

Common questions

Sources

All checked 13/10/2026.

• METR, Measuring AI Ability to Complete Long Tasks (19/03/2025). metr.org
• METR, Time Horizon 1.1 (29/01/2026): post-2023 doubling time 131 days; 2019 onward 196 days. metr.org