AI earns its place when it drafts and a person approves

AI earns its place in a company where the work is reading, writing, sorting and checking, and where a person can catch a mistake cheaply. Start in draft mode: the AI prepares the work and a person approves it. Remove the approval step one task at a time, and only when your own data shows the drafts are reliable. Keep decisions that cannot be undone with people.

Most companies use AI, and few can see it in their profits

In McKinsey’s 2025 survey, 88% of respondents said their organizations used AI regularly in at least one business function. Only 39% reported any effect on EBIT, and most of those said AI accounted for less than 5% of it.¹ BCG’s 2024 survey of 1,000 executives found that 74% of companies had yet to see tangible value from AI. About 70% of the obstacles were people and process, against 20% technology and 10% the algorithms.²

The companies that profit most work differently. McKinsey’s high performers, who attribute more than 5% of EBIT to AI, were nearly three times as likely as everyone else to have redesigned their workflows around AI.

1. AI high performers55%
2. All other respondents20%

Companies that profit most from AI are nearly 3 times as likely to have redesigned the work

Group	Share
AI high performers	55
All other respondents	20

Source: McKinsey Global Survey on the state of AI, 2025¹

Note: High performers attribute more than 5% of EBIT and significant value to AI (n = 109). All others, n = 1,644.

The lesson holds for a small company as much as for a multinational. A model added to a broken workflow produces faster versions of the same problems. The value comes from redesigning the work around what the model does well, and around what it should never do.

The best first uses save time on a draft and fail safely

Good first candidates share two traits. A draft saves real time, and a person can check it quickly. In companies, that usually means:

• Intake. Reading incoming emails, forms and documents, and pulling out the details a person would otherwise retype.
• Drafting. Preparing replies, status updates and first versions of routine reports for a person to edit and send.
• Sorting and routing. Tagging requests by type and urgency, and sending each to the right person.
• Extraction. Turning invoices, contracts or delivery notes into structured records.
• Quality checks. Comparing a deliverable against a checklist before a person reviews it.

Each of these replaces minutes of reading or typing with seconds of checking. None of them lets the AI act alone on a client, a contract or a payment.

A person approves first because AI fails unevenly

The strongest evidence for approval comes from a professional services firm. In a randomized trial, 758 Boston Consulting Group consultants worked with or without GPT-4. On 18 realistic tasks, those with AI finished 12.2% more of the work, 25.1% faster, at higher quality. On one task chosen to sit just outside the model’s abilities, they were 19% less likely to reach the right answer.³

Task	Effect of AI access
18 tasks within AI's reach	12.2% more tasks, 25.1% faster
1 task outside AI's reach	19% less likely to be right

Source: Dell'Acqua et al., field experiment with 758 BCG consultants³

The task that went wrong looked no harder than the others. That is the practical problem. You cannot tell in advance which of your tasks sit inside the AI’s reach, so you find out by checking its drafts and counting.

Decisions you cannot undo stay with a person

Some decisions should stay with a person however good the drafts become. These include commitments on price or dates, refunds above a set amount, legal notices, anything involving safety, and anything a client would be upset to learn a machine decided.

A useful rule: if an error would cost more to fix than the AI saves in a month, a person approves it. The rule is deliberately conservative. Approval costs seconds. A wrong promise to a client can cost the account.

Approval comes off on evidence, never on confidence

Treat draft mode as a measurement period. Every approval is a data point: was the draft sent as is, lightly edited or rewritten? After about four weeks, the pattern is clear for each type of task. Track three things for each:

1. The share of drafts approved without changes.
2. The kinds of edits people make, and whether they fix style or substance.
3. Any error that would have reached a client if nobody had checked.

When a task type reaches a threshold the owner set in advance, measured on the company’s own data, it can run automatically with a person checking a regular sample. When it does not, it stays in draft mode, and the drafts still save time.

The threshold is a business decision. A company may accept a lower bar for internal summaries than for client emails. What matters is that it is written down before the trial starts, so nobody argues from anecdotes afterwards.

The rules you write down matter more than the model

AI drafts are only as good as the rules and examples behind them. Most companies keep those rules in people’s heads: how we word a delay, when we offer a discount, which clients get a call instead of an email. Writing them down is much of the work of an AI project. It is also the same work as mapping the process, which is why a company that has mapped its operations already holds much of what the AI needs.

Illustration. An agency of about 40 people receives about 60 client emails a day asking where a project stands. An AI reads each one, finds the project in the operations system and drafts a reply with the current stage and the next milestone. An account manager approves or edits each draft. After four weeks, status replies are approved unchanged almost every time, so they move to automatic sending, with a person checking a weekly sample. Replies about scope or budget stay in the approval queue, because a wrong answer there costs a client relationship.

One task and one written approval rule are enough to start

Pick a task your team does many times a week, where a draft would save time and an error is easy to catch. Write down what a good output looks like, what the AI may never do and who approves. The playbook Write an approval rule an AI can follow sets out the page. Run the task in draft mode for a month and count. You will then know, from your own data, whether AI has earned its place in that task.