Accuracy Gap

What the Accuracy Gap means

The accuracy gap is the difference between how accurately AI and experienced people make the same call. It is a comparison, and it has two parts that the site’s task and decision article reports side by side. In ambiguous, judgement-heavy situations the gap is wide. On structured, data-rich predictions it narrows, but it does not close.

The term matters because it changes how much authority a business should hand to a tool. The article calls the gap the reason for its whole approach and not an afterthought. If AI were equally accurate everywhere, a business could treat automation as one dial and turn it up. Because the accuracy differs between the two kinds of work, the sensible response is to treat them differently.

This page defines the term, sets out the two comparisons, and explains what they do and do not show. It also separates the accuracy gap from a second pair of figures in the same article that is easy to confuse with it. It repeats the article’s numbers as the article states them, and it promises no result.

What are the two comparisons?

The article reports two, each from research it cites. This table sets them side by side.

The first comes from a meta-analysis the article cites. The second is for calls such as candidate shortlisting, where the inputs are structured and plentiful. In both, the comparison is with experienced humans working the same call, which matters: it is not a comparison with a beginner, and it is not a comparison with a perfect answer.

The article’s own summary of the pair is a single sentence. A system 20 points behind a person on the hardest calls and 10 points behind on the easiest ones is not one dial. It is two different tools doing two different jobs.

Why is the Gap widest on ambiguous calls?

Because that is where the work asks the most of a system. The article says the gap sits precisely in ambiguous, judgement-heavy territory, which is exactly where a decision, rather than a task, tends to live. A task rarely asks a system to weigh conflicting factors with no fixed answer. A decision almost always does.

That is the link between the accuracy gap and the distinction the article is about. A task is checkable against a right answer: formatting a report, pulling figures into a template, routing a ticket to the right queue. A decision is a call such as which candidate gets a callback, which claim gets flagged as fraud, or which client gets an exception to a policy. Getting one wrong costs something a checked task never costs, because a person somewhere has to explain the outcome.

So the wide end of the gap is not a random weakness. It lines up with the kind of work where a wrong answer costs the most. The decision automation page covers handing such a call to a system, and the task automation page covers the steps the gap does not hold back.

What does 75 to 80% Accuracy mean in practice?

The article is specific. At 75 to 80%, AI output is close enough to be genuinely useful as a second opinion, ranking or flagging cases for a person to look at first. It is not close enough to hand the call over and stop checking.

It then names the trap. Treating a 75 to 80% system as though it clears the same bar as an 85 to 90% one is where confidence outruns the evidence.

A piece of plain arithmetic puts the lower figure in context. Accuracy of around 65% means that roughly one call in three comes out wrong. Experienced people are not perfect either, since 85% means roughly one in seven goes wrong for them too. The gap is the difference between those error rates. It is not a claim that people are right every time, and it is not a claim that a person never needs checking.

Is the Gap a reason to distrust Automation?

No, and the article is explicit about it. It says neither number is a verdict on the technology overall. It is a description of where AI’s own error rate sits relative to a human’s, at each end of the ambiguity scale, and that description is the information a business needs before deciding how much authority to hand over. The article calls the numbers the evidence for treating the two kinds of work differently, and not a reason to distrust automation across the board.

This is also the sense of the term as the site uses it: the gap is a reason to sort each step by how much authority to hand over. A password reset is a task with a rule that has no ambiguity in it, and it can be handed over completely. A candidate shortlist is a structured prediction where a person needs to stay accountable. The same technology sits at different levels of trust for good reasons.

The well-scoped workflow page covers the practical result. A workflow where every step has been sorted onto the right level can deliver a fast, measurable return on the steps that were correctly sorted to run unsupervised, while the steps that need a person return a better-informed person making the call faster.

How is it Different from the 45% and 5% figures?

The same article reports a second pair of numbers, and they measure something else. Industry-wide, AI can automate up to 45% of work activities, but only 5% of jobs are fully automatable, according to an analysis the article cites.

The article reads the 45% as how much of the average job is made of steps that can be checked against a fixed standard. It reads the 5% as how rarely an entire role, decisions included, can be handed over without a person still answering for the result. Most of the distance between those two numbers, it says, is the task and decision distinction, and not a measurement error or a gap that will simply close as the models improve.

The two pairs support each other, but they are not the same claim. One is about accuracy on a given call. The other is about how much of a job is made of calls at all.

How does the Gap sort steps onto the Delegation Ladder?

The delegation ladder is where the gap is applied. It sorts each step of a workflow into one of four levels:

• Assist. AI drafts, gathers and compares options, and the person decides. It fits ambiguous, strategic calls.
• Recommend. AI ranks or suggests, and a person stays accountable. It fits structured, data-rich calls such as fraud alerts or shortlisting.
• Execute. AI acts inside clear rules, unsupervised. It fits high-volume, low-consequence, reversible steps.
• Never delegate. Hiring, dismissal, and anything with legal, safety or rights implications. These stay human regardless of capability.

The rungs are sorted by five questions and not by what the technology can attempt: how often the work happens, what a wrong outcome costs, whether it can be undone, how clear the governing rules already are, and whether the outcome needs a kind of legitimacy a model cannot hold. The accuracy gap is the evidence behind why ambiguous work sits low on the ladder. The reversibility page covers one of the five questions.

The article’s steer for a real process is to sort each step and not the chain. A chain approved as one block can contain an ambiguous decision that was never scoped as one, and the result is a system quietly making calls nobody agreed it should make.

Why does candidate shortlisting land at Recommend?

The article uses it as its main example, with a password reset as the contrast. Both, in a narrow sense, decide something about a request. They land on different rungs for reasons unrelated to which one AI can technically attempt.

A password reset is reversible in seconds, happens thousands of times a day, and is governed by a rule with no ambiguity: the request either matches the account or it does not. Every question the ladder asks about it comes back low-risk, so it climbs to Execute.

Candidate shortlisting happens far less often, is much harder to undo once a candidate has been told no, and is judged against criteria that are never perfectly fixed in advance. It scores high enough on consequence and ambiguity that a person has to stay accountable for the final call, however good the AI’s ranking is underneath. That is why it stops at Recommend. The candidate shortlisting page covers the example in detail, and the 75 to 80% figure is the accuracy that goes with it.

Hiring itself, as opposed to shortlisting, sits at Never delegate. The recruitment AI page covers how the EU AI Act treats hiring uses.

What does the Gap not say?

It has limits, and the article’s own wording sets some of them. The page adds one more, in plain terms.

• It is not a verdict on the technology. The article says so directly.
• It is not a forecast. The figures describe a pattern in research the article cites. They are not a prediction for a particular tool, version or task, and the site does not claim they are.
• It is not the site’s own measurement. The article reports them from a meta-analysis and an analysis it cites, and this page repeats them as stated.
• It does not say how accuracy was measured. The article does not describe how accuracy was defined in each study, so a business should not assume the same measure applies to its own work. This is the main reason to treat the numbers as a direction and not a target.
• It is not a claim that people are always right. Experienced humans sit at roughly 85% and at 85 to 90%, which is high and not perfect.

Read within those limits, the figures do one job well. They show that the distance between AI and a person is not constant, and that it is widest where a decision is hardest to check.

How should a Business respond?

The article’s answer is to sort before automating and to keep a named person where the gap is widest. Anything high-volume, reversible and governed by clear rules can move to Execute. Anything ambiguous or judgement-heavy stays at Assist or Recommend, with a named person accountable for the call. Anything touching hiring, dismissal or a legal or safety obligation stays off the ladder entirely.

The human in the loop page describes what keeps the middle rungs honest: a named person who checks the result, with the check naming who reviews, at what stage and against what standard. The checking article adds why the check has to sit outside the model. A wrong answer arrives in the same polished sentences as a right one, and asking a system to double-check itself is not a control, because the same system that made the mistake is marking its own work. It also advises matching the depth of review to the cost of being wrong and not to how good the draft looks.

The article warns about one more thing. A step can slide up the ladder without anyone deciding. A step scoped as Recommend can drift towards Execute because nobody revisited the original design once the tool proved reliable on the easy cases. The drift page covers how a tool slides off the level it was given, and the fix is a named owner and a regular review.

How does the Gap connect to Governance and the Law?

The article makes two links. The first is governance. A written AI governance policy is, among other things, a record of which rung each automated process sits on and who signed off on putting it there. Without that record, a business finds out a process crossed into never-delegate territory only after something goes wrong. Good workflow design already asks the five questions a governance review asks, and keeping them in separate documents is how a process ends up governed on paper and misclassified in production.

The second is the law. The EU AI Act draws its own line around what counts as a high-risk AI system, and the categories it names, employment decisions among them, map closely onto the Never delegate rung. The article treats the overlap as no coincidence. Both the law and the practical test answer the same question: which decisions carry enough consequence that no amount of model accuracy changes the answer. A business that has sorted its workflows onto the ladder is most of the way to a defensible position on the regulatory question.

What to do first

Before automating a process, list its individual steps and write down which of the five questions each one scores high on. Sort each step onto the ladder and not the chain. For any step at Recommend or Assist, name the person accountable for the call and the point at which they check.

For anything already running, look for the step that quietly changed level. Ask who signed off on it, and whether the original design still describes what the tool does now. Where nobody can say, that is the first step to review.

None of this needs a model or a measurement of your own. It needs the steps written down and a name against each one that needs a person. The accuracy figures and the ladder are set out in full in the difference between automating a task and automating a decision, and the check that sits behind the middle rungs is in how to check AI output before it goes to a customer.