# Copilot

An AI copilot helps a person do a task while that person sets direction, checks the result, and decides what happens next.

Source: https://rodamora.com/glossary/copilot
Updated: 2026-08-09

---

An AI copilot helps a person complete part of a task while the person sets direction, checks the result, and decides what happens next. That boundary is easy to miss when a fluent draft arrives. In a study of 758 consultants, people using AI completed 12.2% more tasks and worked 25.1% faster on tasks inside the tested range. On one task outside that range, they were 19 percentage points less likely to reach the correct answer. A copilot therefore needs a [named review point](https://rodamora.com/glossary/human-in-the-loop), evidence beside the output, and someone with authority to reject or redo it. The tool can speed the draft. It cannot own the judgment, final choice, or consequence.

## What does an AI copilot do?

An AI copilot helps with the current step of a task. The person supplies the goal and useful context. The copilot may draft, summarize, analyze, or suggest. The person decides whether the result is right enough to use.

Suppose a manager is preparing a client update. The manager gives the copilot a meeting transcript, the current account record, and the approved scope. The copilot creates a draft. The manager checks each claim against the records, changes the parts that need judgment, and decides whether the message is ready to send.

The name describes who controls the work, not where the feature sits. A copilot can appear in a chat window, beside a document, inside a CRM, or as suggestions in an editor. Microsoft’s [current task guidance](https://support.microsoft.com/en-us/Microsoft-365-Copilot/decide-when-copilot-or-an-agent-is-the-right-tool-for-your-work) describes the same pattern. A person sets the direction and limits. AI drafts or analyzes. The person reviews and approves.

That makes a copilot a useful first step for a firm that is not ready to rebuild delivery. It can shorten research, drafting, and analysis while [the existing process stays intact](https://rodamora.com/blog/your-ai-is-just-a-better-google-search). Someone still starts the work, checks the result, corrects it, and carries it into the next system.

## How is a copilot different from a chatbot, workflow, and agent?

These names describe different parts of the work. One product can contain all four, so the product label does not settle the question.

| Term | Who starts the work? | Who chooses the next step? | Who approves the result? |
| --- | --- | --- | --- |
| Chatbot | A person | A person | A person |
| Copilot | A person | A person, with AI suggestions | A person |
| [AI workflow](/glossary/ai-workflow) | A person or a defined trigger | A path written in advance | A person at a named step, or an automated check |
| [AI agent](/glossary/ai-agent) | A person or a defined trigger | The model, inside set limits | Depends on the action and its controls |

A chatbot names an interface. A copilot names work where a person leads. A workflow names a sequence. An agent names software that can choose among allowed next actions while it works toward a goal.

Vendor bundles make this harder to see. Microsoft and GitHub both use Copilot as a brand across assistant and agent features. Look at the actual behavior. Ask who starts the work, who chooses each next step, what the software may change, and where a person can stop it.

## What can a copilot improve?

A copilot can make a defined task faster. The strongest evidence measures a task or a specific work setting, not a whole company.

In a controlled experiment with 95 professional developers, the group using GitHub Copilot completed one JavaScript task [55.8% faster](https://arxiv.org/abs/2302.06590) than the control group. The authors described the tool as an AI pair programmer. That result tells us about the tested task. It does not show that every developer, project, or firm becomes 55.8% more productive.

Noy and Zhang found a similar result in professional writing. Participants using ChatGPT finished assigned writing tasks [40% faster and produced work rated 18% higher in quality](https://doi.org/10.1126/science.adh2586). In customer support, Brynjolfsson, Li, and Raymond studied an AI assistant used by 5,172 agents. Their 2025 journal article reports [15% more issues resolved per hour on average](https://www.gsb.stanford.edu/faculty-research/publications/generative-ai-work), with larger gains among less experienced workers.

These studies show what happened in defined conditions. They do not show where saved time went after the task. [The extra time AI buys can disappear](https://rodamora.com/blog/the-extra-time-ai-buys-you-is-already-gone) before a firm turns it into useful capacity.

## Why can a copilot make good people more confidently wrong?

AI help varies across tasks that look alike. A reviewer may not know whether the current case sits inside the range where the tool performs well.

Dell’Acqua and colleagues tested AI assistance with [758 consultants](https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838). Across 18 tasks inside the tested range, people using AI completed 12.2% more tasks and worked 25.1% faster, with higher quality. On one task outside that range, AI users were 19 percentage points less likely to reach the correct solution.

The outside task still produced answers that sounded persuasive. That creates a review problem for service firms. A rough error invites a second look. A polished recommendation can pass through a busy review when its weak point sits inside an assumption, a source choice, or one missing exception.

A weak review design gives someone a smooth draft without its sources, a written standard, or enough time to test the claim. The person has no practical way to exercise control. Saying that “a human approved it” tells us that someone clicked. It does not tell us whether that person could find the error.

## What does meaningful human control require?

Human control needs a real decision point. The reviewer needs the source material, a written standard, enough time to compare the output with both, and authority to reject or redo the work.

A proposal copilot shows what this looks like. It drafts from an approved transcript and account record. The reviewer sees those sources beside the draft, along with the current rate card and a checklist for scope, price, and promises. One named person can stop the proposal before it reaches the client.

A weak version posts a fluent proposal in a chat channel and asks, “Looks good?” Nobody can see which record supplied the price or whether the scope matches the call. Several people react with approval. Responsibility spreads until no one owns the decision.

Review should match the consequence. The test in [Delegate the Inputs, Own the Outputs](https://rodamora.com/blog/delegate-the-inputs-own-the-outputs) asks whether the result can be checked and the action undone. A draft that has not left the firm is easy to reject. A signed quote with the wrong price is harder to recover. The first can use a lighter review. The second stays under named human approval.

## How should a service firm measure a copilot?

Measure the whole task, including review. A copilot that saves ten minutes of drafting and adds fifteen minutes of checking made the task slower.

Choose one frequent, clearly defined task and record its current completion time. Run representative cases through the copilot. Track total time, review time, how often the output is accepted without change, how often it is edited, how often it is rejected, and which errors recur.

Completion time shows whether the task moved faster. Review time shows how much of the gain moved downstream. Correction and rejection rates show whether the copilot stays inside the range where it works reliably. Repeated error types can become instructions, checks, or reasons to keep part of the task human-led.

Prompt counts, active seats, and draft volume only show that people used the tool. They do not show that delivery became faster or better. If output grows while the review queue grows faster, the copilot has created more material to sort.

## When should a copilot become a workflow or agent?

Move the work into a workflow when the same inputs, sequence, checks, and handoffs repeat. A trigger can start the path. A person can remain at the decision that needs judgment. The firm can then see and improve a process that started as one person’s habit.

Add an agent when the next action depends on evidence found during the run. The firm needs limits on its tools, a record of its actions, a test for the final state, and a way to recover a failed run. More autonomy without those controls moves the review problem later in the process.

Keep the copilot pattern when judgment changes from case to case, errors are hard to detect, or the standard still lives in one person’s head. The tool can help that person work. It cannot turn unwritten judgment into a dependable operating role.

This is the boundary between assisted work and [an AI employee](https://rodamora.com/blog/what-is-an-ai-employee). The copilot helps with the current task. A role starts to own work when a trigger starts it, ongoing access provides the context, and a check proves that the result exists in the right system.

## Where does a copilot sit on the Delivery Model Ladder?

A copilot sits at **Stage 1, Enhanced**, on the [Delivery Model Ladder](/delivery-model-ladder). A person still starts the work, carries context into the tool, checks what comes back, and moves the result into the next system. The task improves while the person carries it from start to finish.

Stage 2 begins when the firm writes the sequence, triggers, handoffs, and checks into delivery. A copilot may still help inside that path. The tool does not decide the stage. The way the work runs does.

This advice assumes someone can judge the result from available evidence and has enough time to do it. If correctness cannot be checked, keep the task human-led or create the standard and test first. A copilot does not make an uncheckable task safe.

Write down who starts the work, what the AI may produce, which evidence the reviewer receives, who approves it, and what happens after rejection. If those answers still live only in one person’s head, keep the tool at Stage 1.