How to Handle Staff Who Over-Rely on AI

Treat it as a standards problem before a discipline problem. Collect two or three examples, check whether you told the team what AI may and may not be used for, then have a private conversation that sets a clear standard: they own everything they send, they check facts against the source, and some tasks stay human. Review in a month.

Over-reliance is rarely laziness. It usually grows where AI was introduced without rules, where speed is praised and accuracy isn't checked, or where someone is out of their depth and AI hides it. Fix those causes and most cases resolve without any formal process. The steps below use an illustrative café, a conversation script and a written standard you can adapt.

Follow me on Instagram@sagnikteaches

What over-reliance looks like, as opposed to heavy use

Heavy use isn't the problem. Someone who drafts every email with AI and checks each one is doing well. The problem is unchecked use, and use for things AI shouldn't decide. These are the signs worth acting on:

Connect on LinkedInSagnik Bhattacharya

The third sign is often the first one a manager actually sees. In an illustrative car repair garage, the service adviser pastes each technician's note into a chat assistant to make it "customer-friendly" before sending it:

Subscribe on YouTube@codingliquids

TECHNICIAN'S NOTE
Front pads approx 40% remaining. No action needed now.
Recheck at next service.

WHAT THE CUSTOMER RECEIVED
Our technician found that your front brake pads are
significantly worn. For your safety we strongly recommend
replacing them as soon as possible. Please call us to book.

Asked why the message recommended a replacement, the adviser says, "I just asked it to make it friendlier." The AI added urgency the technician never wrote, and a customer who later gets a second opinion stops trusting the garage. The fix is the same as at the café: rewording is fine, but every fact must come from the note.

Step 1: gather specific examples (up to a week)

Don't act on a general impression. Collect two or three dated examples, each with what happened and what it cost or nearly cost. General complaints ("you use AI too much") start arguments; specific examples start conversations.

Say a café's assistant manager has taken on the menu cards, supplier orders and review replies, and uses a chat assistant for all three. In one month:

• A new menu card described a seasonal pastry as nut-free. The recipe uses almond flour. The chef caught it before the card went out.
• A supplier order, based on an AI "forecast" from a rough description of last month, doubled the milk order for a week when the café was closed for a public holiday.
• A reply to a review apologised for "the long wait for your table", although the customer had complained about the price of a sandwich and the café doesn't take table bookings.

The first one is serious; the other two are costly and embarrassing. None of them is about using AI as such. All three are about not checking it against the source: the recipe, the closure diary, the review itself. AI tools state wrong things fluently, which is why checking matters; why AI makes things up explains the mechanism in plain English if the team hasn't seen it.

Put a rough figure on each example before the conversation, because a cost keeps the discussion on consequences rather than habits. The milk is easy to price: 420 litres ordered against about 180 needed is 240 litres over, and at an illustrative $1.10 a litre that's about $260, most of it poured away. The allergen card has no price, which is the point: one customer reaction costs more than years of checking. Against both, checking an order against the delivery notes and the closure diary takes about ten minutes a week.

Write each example down the same way, so you share facts rather than impressions:

EXAMPLE 2 OF 3
Task:            weekly milk order
What went out:   420 litres, for a week with Monday closed
What the source  closure diary: Monday closed
  said:          delivery notes: 4-week average 212 litres
Cost:            about 240 litres over (roughly $260)
How it surfaced: the driver asked whether we had an event on

Step 2: check whether the rules ever existed

Before you judge anyone, ask yourself honestly: did the café ever say that allergen information must be checked against the recipe sheet by the chef? Did anyone say review replies must answer what the customer actually wrote? If AI arrived with "use it, it saves time" and nothing else, this is partly your gap, and the conversation should acknowledge that.

If there's no written guidance at all, write a short standard (step 4) before the conversation, or during it. Rolling out an AI policy so staff actually follow it covers the team-wide version.

Step 3: have the conversation

Private, unhurried, and early in the week rather than at the end of a bad shift. Aim to understand before you correct: the reason behind the over-reliance decides the fix.

CONVERSATION OUTLINE (about 20 minutes)

1. Say why you're meeting, plainly.
   "I want to talk about how we're using AI for the menu cards,
   orders and review replies. Nobody's in trouble; I want to fix
   how we do it."

2. Share the examples, facts only.
   "The pastry card said nut-free; the recipe has almond flour.
   The milk order doubled for the closed week. The review reply
   answered a different complaint."

3. Ask, and listen.
   "Talk me through how you did those. What made it hard to check?"
   (Listen for: time pressure, unclear instructions, not knowing
   where the source is, not realising AI can be wrong.)

4. Own your part, if it's yours.
   "I never told you allergen info has to go past the chef.
   That's on me, and I'm fixing it."

5. Agree the standard.
   Walk through the three rules (below) and the human-first tasks.

6. Agree the support.
   "What would make checking easier? The recipe sheets in one
   folder? Ten minutes set aside before orders go in?"

7. Set a date.
   "Let's look at the next few cards and orders together in two
   weeks, and review properly in a month."

Listen hardest at point 3, because the reason decides the fix:

One case needs extra care. Someone who writes in a second language, or who has dyslexia, may use AI to produce text they'd struggle to write quickly, and for them it's a reasonable aid rather than a crutch. The standard doesn't change (facts checked against the source, human-first tasks stay human), but the support does: keep the sources in one place, have a colleague read the high-stakes messages, and remember that polished wording isn't what you're checking. If the conversation touches on a disability, handle any adjustments through your usual process and take HR advice where you need it.

Step 4: set the standard in writing

Keep it short enough to pin on the wall. Three rules and a list:

1. You own what you send. If it goes out under your name or the café's, you've read it and you'd defend it. "The AI wrote it" is never the answer to a customer.
2. Check facts against the source. Prices against the price list, allergens against the recipe sheet, dates against the diary, the reply against the review. Every fact, not a skim.
3. Some tasks are human-first. A person decides or writes these, and may use AI only to tidy wording afterwards:
   • allergen and dietary information
   • replies to complaints and anything about illness
   • refunds and compensation
   • anything about a member of staff
   • order quantities (AI may help with the sums, from real figures)

For the most important tasks, add a "show your working" habit: a line noting which source was checked ("allergens checked against recipe sheet v4, chef initialled"). It takes ten seconds and makes the check visible. At the café, a week of those lines might read:

Menu card, pear and almond tart: allergens checked against
  recipe sheet v4 (wheat, egg, milk, almonds). Chef: initialled.
Milk order: 4-week average 212 litres; closed Monday, so
  x 5/6 = about 177; ordering 180. Source: delivery notes.
Review reply: complaint was the $9.50 sandwich price.
  Reply answers the price, nothing else. Checked: yes.

Each line names the source and the fact that was checked. If a line can't be written, the check didn't happen, and that's visible before anything goes out rather than after.

The review reply from step 1 shows the second rule at work. The customer had written: "Nice coffee, but $9.50 for a cheese sandwich is steep." Here is what went out, and the version written after checking the reply against the review:

AI DRAFT (SENT)
Thank you for your feedback. We're so sorry about the long
wait for your table. We're working hard to reduce waiting
times and hope to welcome you back soon.

CHECKED VERSION
Thanks for the kind words about the coffee. You're right that
the sandwich isn't cheap: it's made to order with bread baked
that morning. From next week there's a half sandwich at $5.50
if you'd like something lighter.

Every claim in the second version is one the café can stand behind. If the half sandwich weren't real, that sentence would have to come out, however good it sounds.

The garage from earlier needs the same rule built into the request itself, so the tool can't strengthen a recommendation. The adviser's saved prompt now reads:

Rewrite this technician's note in plain, friendly English
for the customer. Don't add, remove or strengthen any
recommendation. If the note says no action is needed, say
so in the first sentence. Keep any measurements as written.

ILLUSTRATIVE RESULT
Good news: your front brake pads have about 40% left, so
nothing needs doing now. We'll check them again at your
next service.

That result is fine to send. If a reply ever includes "we recommend" where the note didn't, the adviser deletes it and tells the manager, because it means the instruction is being ignored and the prompt needs tightening.

Step 5: rebuild the skill where it has slipped

If someone can no longer do a routine task without AI, the business is exposed on the day the tool is down or wrong. For tasks that matter, use "attempt first, then compare": the person drafts the supplier order or the review reply themselves, then asks AI for its version, and compares. At the café, that means the assistant manager works out next week's milk order from the last four weeks' delivery notes and the closure diary, then asks the assistant to check the sums and point out anything unusual. The AI becomes a second pair of eyes instead of the first and only one. The request and an illustrative reply might look like this:

PROMPT
Here are the last four weeks of milk deliveries: 205,
198, 248 and 197 litres. The café is open six days a
week and closed next Monday. My order for next week is
180 litres. Check my sums and point out anything odd.
Don't give me a number of your own.

ILLUSTRATIVE REPLY
Your four-week average is 212 litres. Reducing by one
day in six gives about 177, so 180 is consistent. Week
three (248 litres) is well above the others: was there
an event? If that week was unusual, your average without
it is 200, which would suggest about 167. You could
round up to 200 to be safe.

The first half is useful: the sums are right, and the question about week three is exactly what a second pair of eyes is for (it was a private party). The last sentence ignores the instruction and invents a safety margin, which is the habit the assistant manager is learning not to take on trust. She checks the average on a calculator once, keeps 180 and moves on.

It's slower for a few weeks and it rebuilds judgement quickly. Pair them with someone experienced for the first few rounds. Training staff to use AI in a small business covers how to build this into everyday work.

Step 6: fix what pushed them there

Look at the conditions, not just the person:

• Workload. If the assistant manager does menus, orders and reviews on top of running the floor, AI was a survival tool. Move a task or set time aside for it.
• Targets. If you praise fast replies and never check accurate ones, you get fast replies.
• Missing sources. If the recipe sheets are in three places, checking is hard. Put them in one.
• No review step. For outputs where a mistake is costly, build the check into the process instead of relying on memory. Setting up human review without slowing down shows how to match the check to the risk.

Responses that make it worse

A few reactions feel decisive and tend to backfire.

• Banning AI for everyone after one incident. Use moves to personal phones, where you can't see it and the data protections are weaker.
• Raising it in front of the team. The person gets defensive, and everyone else learns to hide their AI use rather than check it.
• Reading someone's chat history without warning. Even where a business plan lets an admin see it, doing so without telling staff what can be seen damages trust, and may raise legal questions.
• Judging people by how much they use AI. Usage counts say nothing about quality. Judge the work that goes out.

Step 7: follow up, and know when to go formal

At two weeks, look at a handful of the person's AI-assisted outputs together, informally. At a month, review properly: have the three problem areas been clean? Is the standard being followed? What's still hard?

Keep the check small and specific. At the café, the two-week look took fifteen minutes and covered six outputs, each compared with its source:

TWO-WEEK SPOT CHECK
Menu card, autumn soup   allergens vs recipe v2      match; chef initialled
Menu card, brownie       allergens vs recipe v5      match; chef initialled
Milk order, week 5       delivery notes + diary      match; working line present
Milk order, week 6       delivery notes + diary      match; working line missing
Review reply, 4 stars    against the review          match
Review reply, 2 stars    against the review          answered the point, but offered a refund

Two lines need a word. The missing working line is a habit slipping rather than an error, so a reminder is enough. The refund offer matters more, because refunds are on the human-first list and the reply offered one without asking. That's the one item to discuss, calmly, with the written standard on the table.

Most cases end here. If the same serious problem continues after a clear written standard, a fair conversation and real support, it becomes a normal performance matter: follow your usual process, and frame it around the quality of the work and following instructions, not around the tool. Take advice from an HR or employment adviser before any formal step, because the rules differ by country and by contract.

Preventing it across the team

One person's over-reliance is often an early sign of a team-wide habit. A few small moves help everyone:

• Put the three rules in the staff handbook and go through them with every new starter.
• Make it safe to say "the AI got this wrong". Thank people for catching AI errors, share the examples at team meetings, and keep a simple log. Mistakes that are hidden repeat.
• Give people a better starting point than a blank chat box. A saved set of instructions that includes your price list, opening hours, allergen rules and tone means the AI starts from your facts rather than its guesses, and far fewer errors reach the checking stage.
• Spot-check a few outputs a week, briefly and without drama, so checking is normal rather than a sign of suspicion.
• Explain what AI is bad at: current facts, your own prices and rules, maths without real figures, and anything it hasn't been given. If you sell to customers in the EU, the EU AI Act also expects businesses using AI to take measures that support their staff's AI literacy; what your staff need to know about AI literacy sets out a sensible baseline.

Harder cases

Can I ban one member of staff from using AI?

You can set different rules for different tasks, and it's usually better to restrict the specific tasks where things went wrong than to ban AI outright. A blanket ban on one person can feel like a punishment and pushes use onto personal phones where you can't see it. If you're considering a restriction as part of a formal process, take employment advice first.

Should I monitor what staff type into AI tools?

Business plans give admins some visibility, and it's reasonable to review work output, which is what matters. Reading individual chat logs is more sensitive: staff should know in advance what can be seen and why, usually through your AI policy. Rules on workplace monitoring differ by country, so check with an employment or data-protection adviser before doing it.

What if the over-reliant person is a manager?

The same steps apply, with more urgency, because managers set the standard others copy and often approve other people's work. Have the conversation privately, focus on outputs rather than the tool, and ask them to help write the team standard. Managers who help set a rule are more likely to model it.

Further reads

• When Not to Use AI in Your Business: 9 Tasks to Keep Human — Nine tasks to keep human, useful for your standard.
• AI Error Log: Track Mistakes and Stop Them Happening Again — Log AI mistakes so the same one doesn't repeat.
• How Accurate Is ChatGPT? What Owners Should Expect by Task — Realistic accuracy by task, to set expectations.
• How to Build a Shared Prompt Library for Your Team — Good shared prompts reduce careless one-off use.
• How to Onboard New Hires Onto Your AI Tools and Rules — Set the standard from a new starter's first week.
• How to Write an AI Usage Policy for Your Small Business — Write the team-wide rules the conversation points to.
• How to Check AI Is Doing Good Work, Not Just Fast Work — A rubric template, a 30-minute weekly sampling routine, a re-runnable test set and the warning signs, shown through a garden centre's plant-care replies.
• AI Change Management for Small Teams: A Practical Plan — An eight-week plan sized for five to twenty people, with a meeting script, a resistance table and a worked nursery example.
• How to Get Your Staff to Actually Use AI Tools — Diagnose low AI use, fix the six usual blockers, and see how a tutoring agency went from 3 regular users out of 12 seats to 9.
• How to Redesign Job Roles Once AI Handles Routine Tasks — A four-step way to rework roles after AI takes routine tasks: task inventory, honest freed-hours count, a decision on the time, and new role descriptions.
• Will AI Replace My Employees? An Honest Answer for Small Firms — An honest answer for owners: what the evidence shows, a task-by-task method to score each role, a bookkeeping firm example and what to tell staff.
• Pros and Cons of Using AI in Your Business, With Real Costs — Every pro and con of business AI with a price attached, a first-year ledger for a small law firm, and the costs that never reach an invoice.
• Shadow AI: Is Your Team Using AI Without Telling You? — How to find the AI tools your staff use without telling you: an amnesty survey, five afternoon checks, and a keep-move-stop rule for each thing you find.
• How to Train Front-of-House Staff to Work Alongside AI — Three short sessions, six role-play cards and a one-page counter card for teaching front-of-house teams to work with AI bookings, chat and phone assistants.
• How to Write Fairer Performance Reviews With AI — An evidence-first method for small teams: year-round notes, written criteria, an AI fairness pass and a side-by-side check before any rating is final.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: EU AI Act Article 4 as amended by the Digital Omnibus on AI (checked September 2026).