Your AI feature has usage. It might not have acceptance

By Lindsey JOctober 2, 2026

The usage metric that misses the real question

Most teams watching an AI feature track whether people open it. Daily active users, session count, click-through rate. These numbers answer one question: did someone try it today. They do not answer a different question: did the AI's output survive contact with a real task.

Diagnostic question: do you know what share of your AI feature's output a user keeps without editing it?

What GitHub's own data shows about acceptance rate

In research GitHub published with Accenture in May 2024, developers accepted around 30 percent of GitHub Copilot's suggestions. The same research reported that developers retained 88 percent of Copilot-generated characters in their editor. GitHub reports those as two separate metrics. Neither is a breakdown of the other. Both describe how much of the AI's work survives contact with a real task, which is the question usage numbers alone cannot answer.

This was measured on code suggestions inside a developer's editor, for developers at Accenture, in a randomized trial and a company-wide rollout, published by GitHub in May 2024. It is a reference point for one kind of AI feature, not a universal target. Measure your own workflow before you compare yourself to it.

Aero has covered the usage half of this gap before, in why nobody is using your new AI feature. Acceptance is the number to watch once usage is no longer the problem.

Why a low acceptance rate often sits in the handoff

A low acceptance rate has two different causes. They call for different fixes. The first is quality: the suggestion is wrong, off topic, or outdated. The second is friction in the handoff: the suggestion is usable, but the way it is presented costs more trust or effort than accepting it is worth. A suggestion with no diff view, no confidence signal, or too wide a scope asks the user to take it on faith. Before retraining a model or paying for a bigger one, spot check a sample of rejected suggestions. If most of them were actually fine, look at the interface before you look at the model.

A decision rule for a low acceptance rate

Pick one AI-generated step in your product. Measure its acceptance rate, meaning the share of outputs a user keeps without editing, over two weeks. If it stays low after that ramp period, change how the suggestion is presented before you touch the model: show a diff, narrow the scope of what the AI proposes, or add a confidence signal. Re-measure for another two weeks. If acceptance still has not moved, the quality of the output is the more likely cause. That is where the next investment belongs.

A worked example

Picture a mid-market support team that ships an AI-drafted reply inside its helpdesk tool. Agents can accept the draft, edit it, or write their own from scratch. After launch the team tracks only how often agents open the draft panel. That number looks healthy. When they add acceptance tracking, they find agents keep only about one in eight drafts without heavy edits. A review of the edited drafts shows the model is usually right about the facts. Agents are rewriting the tone and cutting the draft down to two sentences. The team narrows the AI's job to drafting the resolution steps only, in the agent's own house style. It leaves the greeting and closing to a fixed template. Acceptance rises within the next sprint, with no change to the underlying model.

Before you ship the next AI feature

Do you measure acceptance rate per AI-generated step, not only overall feature usage
Do you know, from a sample, whether your low-acceptance cases are wrong outputs or just badly presented ones
Have you set your own benchmark for a healthy acceptance rate, rather than borrowing one from a different product's workflow
If acceptance stays low after you fix the presentation, do you have a plan to scale the feature back rather than let it keep eroding trust

FAQ

What counts as an accepted AI suggestion?

A suggestion a user keeps in the final output without rewriting it. Different products draw this line differently. Decide yours before you start measuring, and keep the definition the same across every reporting period so the number stays comparable.

Is a low acceptance rate always a problem?

No. A feature meant to produce several options for a person to choose from, such as a set of draft headlines, can reasonably have a lower acceptance rate per option. Judge acceptance against the job that specific feature is meant to do. A flat number across every feature will mislead you.

How is acceptance rate different from adoption?

Adoption asks whether people use the feature at all. Acceptance asks whether the people who use it trust what it gives them enough to keep it. A feature can have high adoption and low acceptance. That combination usually means people open it out of curiosity or habit, then do the work themselves anyway.