Keep your team thinking, not just accepting.

Most of the AI mistakes we see in teams start out small. A summary leaves out the one condition that mattered, a reply quotes last year's fee schedule, or a report total doesn't match the spreadsheet it came from. The tool got something wrong, and the person who sent it never looked closely enough to notice.

The first draft now arrives finished.

For most of working history, a first draft was visibly rough. You could see the gaps, so you filled them, and the filling was where the thinking happened. An AI draft arrives with full sentences, clean formatting and a confident tone, and it looks like something that has already been checked.

That appearance is the problem. When a reply to a parent, a grant summary or a vendor comparison reads as finished, the natural move is to skim it and send it. Nobody decides to stop checking. It wears away one reasonable shortcut at a time, usually on a busy afternoon, and after a few months the shortcut has become the process.

The risk grows as the tools improve. Picture a tool that gets something wrong in one draft out of five: people stay alert. Make it one in fifty and people learn to trust it, so the fiftieth draft goes out with the error in it.

The signs that checking has stopped.

You rarely catch this by asking people whether they review AI output, because everyone says yes. Look at the work instead. Some signs are easy to spot once you know them:

• Documents from different people start to sound the same, down to the phrasing.
• Someone asked about a figure in their own report can't say where it came from.
• Questions get answered with "that's what it said."
• Long responses appear for small decisions, and the decision itself still isn't made.
• A person who used to push back in meetings now forwards several options without picking one.

None of these is a disaster on its own. Together they tell you the team is handing over judgment along with the typing, and judgment is what you hired them for.

Form a view before you ask.

The single most useful habit is also the plainest. Before asking AI for a draft, an analysis or a recommendation, write down your own answer in two or three lines. What do you expect the summary to say? Which vendor do you lean toward, and why? What should the total roughly be?

This takes a couple of minutes, and it changes how people read what comes back. With a view of your own, you notice where the tool agrees, where it disagrees and where it has invented something. Without one, you have nothing to compare against, so you accept the draft because it looks reasonable.

Before you ask AI for an answer, write down what you think the answer is.

It also produces better requests. People who have thought first ask narrower questions, include the detail that matters and know what to reject. The tool does better work when the person using it has already done some.

Put checks where the work leaves the building.

Asking everyone to "verify everything" fails, because nobody can and nobody does. Pick a few points in the work where an error would be expensive or embarrassing, and make checking part of the process at those points. Leave everything else alone.

In most organizations those points are predictable: anything sent to a client, patient, resident or funder, any figure that goes into a budget or a board packet, and anything that commits the organization to a date, a price or a promise. At those points, a person traces each number back to its source, confirms names and dates, and reads the whole thing at the speed the recipient will.

Managers can help with one question asked consistently: what did you change after the first draft? If the answer is "nothing," that may be fine for a meeting recap. For a proposal or a policy notice, it should lead to a conversation. Asked often enough, the question becomes the habit, and people start to arrive with an answer ready.

It also helps to keep a few tasks deliberately manual. A team that still writes its own monthly summary by hand once a quarter keeps the skill that lets it spot a bad one.

A checking habit list for each role.

Generic advice to "use judgment" doesn't stick. Specific habits tied to the work people already do are easier to follow and easier for a manager to ask about. Adapt these to your own roles.

Front desk and intake

1. Check every date, time and name in a reply against the calendar or the record, never against the draft.
2. Read policy answers against the current policy document before sending.

Finance and bookkeeping

1. Recalculate any total the tool produced, using the source spreadsheet.
2. Treat a figure with no visible source as wrong until you find one.

Program, project and operations leads

1. Write your expected answer before asking for analysis, and note where the result differs.
2. When the tool offers several options, cut them to one recommendation yourself before anyone else sees them.

Anyone writing to the public

1. Read the final version aloud once, start to finish.
2. Remove any claim, number or quote you can't point to in a source you trust.

Managers

1. Ask "what did you change?" on anything that leaves the organization.
2. Once a month, pick one AI-assisted piece of work and trace one fact in it back to the source with the person who made it.