AI saved you an hour. How long did you spend checking it?

AI can produce in seconds what might have taken a person an hour. That's extraordinary. But the first output isn't always the end of the work. If you're measuring AI productivity, measure what happens afterwards too.

Here's an easy AI productivity calculation. A task normally takes an hour. AI does it in five minutes.

We've saved 55 minutes. Lovely. Except perhaps it actually went like this.

You spent five minutes explaining the task. The AI produced something in two minutes. You spent ten minutes reading it.

Found something wrong. Asked it to change it. Checked the new version.

Looked up two claims because you weren't sure they were right. Changed the wording yourself. Then discovered it had misunderstood one of the original instructions.

By the time you finished, perhaps the task took 35 minutes. That's still an improvement. A good one.

But it isn't a 55-minute saving. And if businesses want to understand what AI is actually doing to productivity, that difference matters.

The generation time is the easy bit to see

AI is incredibly fast at producing things. An article. A report.

A summary. Some code. An email.

A proposal. An analysis. A spreadsheet formula.

A customer response. A document. The speed is immediately impressive because we can compare it with how long creating the same thing manually might have taken.

But generation is only one part of the process. There's also: BRIEFING

What did we have to give the AI? CONTEXT What information did we need to find first?

GENERATION How long did the AI take? REVIEW

Did somebody need to read or inspect the output? CORRECTION What needed changing?

VERIFICATION What needed independently checking? INTEGRATION

Did somebody still need to put the result somewhere else? RECOVERY What happens when the AI gets something wrong?

That's the whole job.

AI can move the work rather than remove it

This is one of the most useful things to look for. Suppose somebody used to spend 45 minutes writing a report. Now AI drafts it in three minutes.

Brilliant. But perhaps they spend: 15 minutes checking the figures,

10 minutes correcting the interpretation, five minutes fixing the format, and another five putting the result into the correct system.

The work hasn't disappeared. Part of it has moved from creation to verification. That can still be enormously worthwhile.

But you need to know that's what happened.

Checking isn't free

Checking work requires attention. And attention is expensive. Sometimes reviewing an AI-generated answer is much quicker than creating it yourself.

Sometimes it isn't. There are tasks where the only reliable way to know whether AI did the work correctly is to effectively do the work again. That's a poor automation candidate.

Imagine asking AI to compare two complicated documents. If a specialist then needs to read every line of both documents to verify the comparison, what did we save? Perhaps something.

Perhaps not much. The important thing is to measure it.

The ideal AI task is cheap to verify

This is a useful way to think about automation. Some outputs are very easy to check. Ask AI to turn a meeting transcript into five action points.

A person who attended the meeting can probably scan those quickly and recognise whether they're right. Cheap verification. Ask AI to produce a complex technical analysis in an area nobody on the team understands.

Now verification becomes much harder. Who knows whether it's right? How long will they need?

What happens if everyone assumes it is? That's a very different risk.

Output difficulty and verification difficulty aren't the same

AI has created an unusual situation. Something can be difficult for a human to produce but easy for a human to recognise as correct. That's an excellent place for AI.

For example, turning rough notes into a well-structured first draft may take AI seconds. The person who supplied the notes can often tell very quickly whether the draft represents what they meant. Other tasks are the opposite.

AI can produce an answer quickly, but determining whether the answer is correct requires specialist knowledge or substantial investigation. The first category is usually much easier to deploy safely.

This is why first drafts are so powerful

One of the strongest uses of AI isn't necessarily: Do the entire job. It's:

Get me from nothing to something. A blank page is expensive. A first draft gives you something to react to.

An email. Proposal. Report.

Document. Outline. Piece of code.

Research structure. You can then apply judgement to something that already exists. In those situations, checking and editing aren't unfortunate leftovers from failed automation.

They're part of a sensible division of work.

But sometimes "human in the loop" hides the cost

We use this phrase constantly in AI: human in the loop. It sounds reassuring.

And often it is exactly what we want. But businesses should ask what the human is actually doing in that loop. Are they:

glancing at an obvious output for 30 seconds? approving an exception? checking a high-value decision?

Or reading every word because they don't trust the system? Those are completely different operating models. Simply saying "a human checks it" doesn't tell us whether the workflow is efficient.

Approval and verification are different

This is another distinction worth making. Approval can be quick. Yes, send that.

Verification can require work. Are all 47 figures in this report correct? If AI prepares a refund and the employee simply confirms that the customer is eligible, that's one thing.

If the employee has to reconstruct the entire case to discover whether the AI's recommendation makes sense, that's another. When designing a workflow, ask which one you're asking the person to do.

The cost of an error matters too

Suppose AI is right 99 times out of 100. Sounds excellent. But what's the hundredth case?

A slightly awkward internal summary? Probably manageable. A £20,000 payment?

Different problem. An incorrect medical recommendation? Very different problem.

A contractual commitment sent to a customer? Again, different. Accuracy on its own doesn't tell you whether a workflow should be automated.

You also need to know: How expensive is being wrong?

Some mistakes are obvious

This makes automation much easier. Imagine AI extracts an order number and the system can't find that order. Something is clearly wrong.

Stop. Escalate. Now imagine AI writes a plausible but incorrect explanation in a report.

Nothing necessarily breaks. The output looks fine. The mistake can travel much further before anyone notices.

Those are the errors I worry about more. Not because they're dramatic. Because they're quiet.

AI errors can look professional

Humans have useful warning signals. A broken spreadsheet may show an error. A failed integration may return a failure message.

AI can produce something wrong in beautifully structured English. That's one of the reasons checking matters. Fluency and correctness aren't the same thing.

An answer can: sound confident, look complete,

follow the requested format, and still contain an important mistake. The more polished the output, the easier it can be to forget that distinction.

Expertise changes checking cost

Imagine AI writes something in a subject you know extremely well. You can probably scan it and quickly notice: That's wrong.

That's misleading. That's missing the point. Now give the same output to somebody unfamiliar with the subject.

They may need to research every important claim. Same AI. Same output.

Completely different verification cost. This is one reason experienced people can sometimes gain disproportionately from AI. They can recognise good work faster.

AI can help with the checking too

There is an obvious response to all this. Use AI to verify AI. And sometimes that's useful.

A second model can: review an answer, look for contradictions,

run tests, compare against source material, check required fields,

challenge assumptions, or independently solve a problem. In software development, AI can write code and then help generate tests for it.

In documents, one AI step might extract information while another checks it against the source. That's useful. But we shouldn't turn:

AI said it was correct into proof that it was correct. The verification method should match the consequence.

Grounding makes verification easier

If AI is working from known business information, make the source visible where possible. Suppose it answers an employee's policy question. It's much easier to trust and verify the answer if it also shows:

which policy it used, which section, and when that document was last updated.

The same applies to research. A summary with traceable sources is much more useful than a beautifully written answer whose origins are unclear. Designing for verification can be as important as designing for generation.

Structure the output so it can be checked

This is a practical trick. Don't always ask AI for one enormous finished answer. Ask it to show the pieces you need to verify.

For example: Decision What does it recommend?

Evidence Which information did it use? Missing information

What doesn't it know? Exceptions What looks unusual?

Next action What does it propose doing? Now the person reviewing the work has something structured to inspect.

That can reduce verification time substantially.

Confidence isn't proof

You can also ask AI to identify uncertainty. That's useful. But an AI saying:

I'm 95% confident. doesn't automatically mean it has a calibrated 95% probability of being correct. Treat confidence indicators as one signal, not a guarantee.

What matters operationally is whether the workflow knows when to stop, ask for more information or escalate.

The best system reduces what needs checking

This is where implementation becomes interesting. The first version of an AI workflow may require a person to review everything. That's sensible.

Then you collect evidence. Which cases are almost always correct? Which cases produce errors?

Which categories are risky? Which signals predict trouble? Perhaps eventually:

routine cases proceed, unusual cases are flagged, high-value cases require approval,

low-confidence cases escalate, and only a small proportion needs detailed human review. Now AI has genuinely reduced the work.

The objective shouldn't necessarily be: remove the human. It should be:

use the human where their attention is valuable.

Don't make your best people professional AI checkers

This is a risk I think businesses should watch. Imagine AI increases the volume of work a team can produce fivefold. Great.

But now your most experienced employee spends all day checking AI output. Have you increased their capability? Or turned them into quality control for a machine?

Sometimes review is exactly where their expertise belongs. But ideally AI should remove low-value work so skilled people can spend more time on things requiring their judgement. If the new workflow consumes more of that judgement than the old one, examine it carefully.

More output isn't necessarily more productivity

This matters particularly in areas like marketing and software development. AI can generate: ten articles instead of two,

five campaign concepts instead of one, thousands of lines of code, twenty design variations,

or enormous amounts of research. But somebody may now need to: read,

choose, test, edit,

maintain, approve, and understand

all of that output. Production capacity can increase faster than review capacity. At that point, more output becomes a new bottleneck.

Sometimes the answer is to generate less

This feels almost heretical in an AI productivity conversation. But perhaps you don't need 20 options. You need three good ones.

Perhaps you don't need a 5,000-word report. You need the five things that changed. Perhaps you don't need AI to build six variations.

You need one clearly specified implementation. The cheapest thing to verify is often the thing you never generated.

Measure the workflow before and after

If you want to know whether AI actually improved a process, I'd measure something like this. BEFORE AI How long does the whole task take?

How much waiting is involved? How often does it happen? How often are there mistakes?

Who needs to be involved? AFTER AI How long does briefing take?

How long does AI processing take? How long does review take? How much correction is required?

How often does the work need to be redone? How often are errors caught? How often aren't they?

Has the number of people involved changed? Has the customer outcome improved? Now you're measuring productivity.

Not just generation speed.

I'd add one more metric: trust

Not vague "Do employees like AI?" Something more useful. How often does the reviewer accept the output without substantial change?

If people rewrite almost everything, investigate why. Perhaps: the AI needs better context,

the task is badly defined, the output format is wrong, the underlying information is poor,

the model isn't appropriate, or the task simply isn't a good AI task. Review behaviour tells you a lot.

The checking burden should fall over time

For a well-designed AI workflow, I'd generally hope to see this happen: EARLY AI does work → human checks most of it.

LEARNING We identify common errors and improve the workflow. MATURE

Routine work needs little intervention → exceptions get attention. If six months later somebody is still manually reconstructing every AI answer to make sure it isn't wrong, I'd question what we've actually automated.

There are also benefits that aren't time savings

We shouldn't become too obsessed with minutes. AI may make a process better because: responses are more consistent,

customers receive answers sooner, more information is considered, fewer tasks are forgotten,

employees can handle more work, the business can offer something new, or people spend less time on work they hate.

Those things have value. The point is simply to measure the real outcome, not the impressive part of the demo.

AI productivity should survive contact with the accounts

This is where business implementation ultimately gets quite simple. If AI genuinely improves productivity, something should change. Capacity increases.

Cost falls. Output improves. Response time falls.

Revenue increases. Quality improves. People spend time on more valuable work.

Customers get a better experience. Something. If the only evidence is:

Look how quickly it generated this. we haven't finished measuring.

A simple calculation

When evaluating an AI task, I'd think about: TIME SAVED CREATING minus

TIME SPENT BRIEFING minus TIME SPENT CHECKING

minus TIME SPENT CORRECTING minus

EXPECTED COST OF ERRORS equals something much closer to: REAL VALUE

It doesn't need to become a complicated financial model for every email. It's simply a better way to think.

This doesn't make me less excited about AI

Quite the opposite. I use AI because the productivity gains can be enormous. The point is to find the places where those gains survive the entire workflow.

Tasks where: AI is genuinely faster, the output is useful,

verification is cheap, mistakes are detectable, and the person's attention moves somewhere more valuable.

That's where it gets interesting.

AI saved you an hour. Did it?

Maybe it did. Maybe it saved 25 minutes. Maybe it saved three hours somewhere else in the process you hadn't considered.

Maybe it didn't save time at all but produced a better outcome. All of those are useful things to know. What matters is that we stop measuring the moment the AI finishes generating.

Follow the work to the end. Count the checking. Count the corrections.

Count the failures. Count the improvement. Then decide whether the process is actually better.

Because AI productivity isn't how quickly the AI works. It's how much better the whole system works with AI inside it.

Where to go next

• Why your AI demo worked and your AI implementation didn't The impressive output is only the beginning. Real implementation has to survive the whole workflow.
• What should businesses actually automate with AI in 2026? Find the tasks where AI creates real value rather than simply faster output.
• Don't start with AI. Start with the work. Measure the problem before choosing the technology.
• The important question isn't what AI can do. It's what you should let it do. The harder an AI action is to reverse, the more carefully its authority should be designed.

Related: Why your AI demo worked and your AI implementation didn't.