Skip to content
Limited-time offer: Free external security assessment with your Buyer's Guide. Get the Guide
Business Technologies NI
Interactive learning

Is it safe to paste this into AI?

A 10-minute story for business owners: six stages of AI arriving at your company. See how each approach plays out, and what a prepared business does instead.

How to play

  • You play the owner of Mossvale Surveying, a 20-person firm of building surveyors in Lisburn (fictional).
  • 6 stages, about 10 minutes. At each stage, choose one of four answers. You then see how all four compare, why the best one is right and what each would cost, and the story moves on.
  • At the end you get your score out of 6, the risks your choices left open, and the lessons to take away.
  • No sign-up, and nothing you choose is stored.
Interactive story · 10 minutes

You run Mossvale Surveying

A 20-person firm of building surveyors in Lisburn (fictional). See how each way of handling AI plays out for the business. At each of six stages, pick what you would do, then see how all four options compare.

What Mossvale has in place

Microsoft 365
Business accounts for everyone, with email and files in Microsoft 365.
Copilot
Two Microsoft 365 Copilot licences, bought last month for the partners.
Policy
An acceptable-use policy that every member of staff signed when they joined.
People
Experienced surveyors who use ChatGPT sensibly, they say, for emails, reports and quotes.
Files
Shared folders in SharePoint, organised by client and by team.
Payments
Supplier payments approved by a partner, by email or a quick call.
Clients
Housing associations, councils and landlords, many with data-protection questionnaires.

On paper, sensible. And AI is only a writing aid here, isn’t it?

Find out where your own gaps are

Mossvale looked sensible on paper. A free self-assessment takes a few minutes and shows where your own business’s gaps are, AI included.

Read the whole story on one page

What Mossvale has in place

Microsoft 365:
Business accounts for everyone, with email and files in Microsoft 365.
Copilot:
Two Microsoft 365 Copilot licences, bought last month for the partners.
Policy:
An acceptable-use policy that every member of staff signed when they joined.
People:
Experienced surveyors who use ChatGPT sensibly, they say, for emails, reports and quotes.
Files:
Shared folders in SharePoint, organised by client and by team.
Payments:
Supplier payments approved by a partner, by email or a quick call.
Clients:
Housing associations, councils and landlords, many with data-protection questionnaires.

On paper, sensible. And AI is only a writing aid here, isn’t it?

Stage 1 · Monday, 9:05am

AI has arrived at Mossvale

You own Mossvale Surveying: twenty building surveyors in Lisburn.

“Just so you know, half the team are already pasting things into free AI chatbots to write their emails and reports.”

Claire, your office manager

She isn’t unusual. A third of UK businesses now use AI (32% in 2026, up from 16% a year before).

And in Microsoft’s 2024 survey across 31 markets, 78% of people using AI at work brought their own tools, 80% at small and medium-sized firms.

An hour later, a big client sends over a supplier questionnaire. They want to know how you handle AI.

Question 14

“Describe your organisation’s approach to AI and how you protect our data.”

Nobody at Mossvale has an answer to that yet.

What do you do?

The best answer: Write a one-page AI policy and point everyone to the Copilot Chat in your Microsoft 365, signed in with their work account

Copilot Chat comes with your Microsoft 365 at no extra cost. Signed in with a work account, chats are protected under your Microsoft 365 terms and aren’t used to train its models.

What it costs: Some set-up time, and nothing extra to buy.

By Friday…

Question 14 gets a straight answer, in one paragraph.

Which AI the firm uses: Copilot Chat, which comes with your Microsoft 365, on work accounts. What can go in, and what never does: passwords and keys, and nothing confidential in any other AI tool.

Signed in with a work account, Copilot Chat shows a green shield. Its chats are covered by the same Microsoft 365 terms as your email and files, and aren’t used to train its models.

A personal Microsoft account doesn’t get that protection, so everyone learns how to check which account they’re in.

Then a half-hour walk-through for everyone, not an email: real examples, and what to do if something slips.

Even the right call isn’t free: some outside help to write the policy and check the set-up, and a little of everyone’s time.

Decide where company data may go, give people a safe tool they already have, and tell them face to face. That's the whole fix, started.

The traps

What good looks like

Every poor answer at that stage had the same gap: nobody decided where company data is allowed to go. The fix is four plain steps, done together.

  1. A clear AI policy. One page that says which AI tools are approved, what staff can put into them, and what never goes in: client personal data in unapproved tools, passwords and keys, unreleased financials.
  2. An approved, secure AI. Business accounts the company controls, not personal ones: Copilot Chat signed in with work accounts, AI inside your own Microsoft 365, or a private AI for the most sensitive work. Know the terms: what's kept and for how long, whether anyone can review it, where it's held and how legal holds are handled. A training opt-out covers none of these.
  3. Tidy permissions. Before any AI searches your files, remove company-wide sharing links, fix who can open which sites and folders, and label sensitive files.
  4. A trained team. Short and practical: real examples, which account to sign in with, take names and figures out before asking, and what to do if something slips, like changing a password that was pasted.

Stage 2 · Two months on

Copilot for everyone?

The partners love their paid Microsoft 365 Copilot. It drafts their reports and sums up a week of email in a minute.

“Can we get the full version for everyone? The team keep asking.”

Claire

Copilot Chat works from the web and what you give it. The paid Copilot also searches your email and files for you.

In the Government’s own trial, 20,000 civil servants reported saving 26 minutes a day, most of it on drafting.

Copilot can show a person anything they can already open. Microsoft itself warns that generative AI amplifies oversharing.

Nobody has looked at who can open what since the move to SharePoint.

How do you roll it out?

The best answer: First have IT find over-shared folders, remove ‘everyone’ links and label HR and client files. Then roll it out team by team

It’s Microsoft’s own advice: Copilot shows people what they can already open, so fix what they can open first. Then asking about pay gets nothing back.

What it costs: Some IT work up front, before anyone gets a licence.

Six weeks later…

The HR folder belongs to the partners again. Each client’s folder is open to the people working for that client.

Copilot reaches the rest of the team in two batches. The first person to ask it about pay gets nothing back.

It’s Microsoft’s own advice: find the over-shared content, remove ‘anyone’ links, restrict access and label sensitive files.

Even the right call isn’t free: some IT work, and time from each partner and team lead checking their own folders.

Tidy first, then switch on: AI should only show people what they were always meant to see.

The traps

Stage 3 · Thursday, 3:15pm

Mark’s confession

Mark knocks on your door. He looks like someone who has rehearsed this.

“Before the policy, I pasted a report into ChatGPT on my own account. The tenant’s name and address were in it. And a note about her stair-lift.”

Mark, a surveyor

Nobody else knows. He could have said nothing.

A tenant’s details in a tool you have no contract with is a personal data breach, and a note about her health makes it more serious.

You must record every breach. If it’s likely to put her at risk, you report it to the ICO within 72 hours of becoming aware.

What do you do?

The best answer: Thank him, log it, get the chat deleted, assess the risk to the tenant, report it if it’s likely, and tell the client yourself

That’s the ICO’s way: record it, assess it, report within 72 hours if she’s at risk. Her health note makes it likely, so you report, and ring the client first.

What it costs: Some of your time; reporting itself is free.

By Monday…

You thank Mark. He did the hard part.

It goes in a breach log: what went in, when, which tool, and what you did about it. He deletes the chat and switches off its history.

Her health is involved and the copy is out of your control, so you report it to the ICO inside the 72 hours, with what you know.

Then you ring the housing association yourself, before anyone else can.

A report isn’t a fine. With small organisations, the ICO says its focus is on helping you comply.

Record it, assess it, report it if it's likely, and tell the client first. An afternoon, not a crisis.

The traps

Stage 4 · The following Tuesday

The quote that wrote itself

Ciara uses the firm’s approved AI to draft a fee proposal: forty homes for a landlord in Lurgan.

It reads beautifully. She checks the wording, not the sums, and sends it.

“Great price. We accept. When can you start?”

The landlord

The AI took the rate from last year’s proposal in the same folder. The fee is thousands short.

AI writes a wrong figure in the same confident tone as a right one. And the landlord has it in writing.

What do you do?

The best answer: Ring the landlord today, own it, send the right fee, and make a rule: someone checks every figure in AI-drafted work against its source

Owning it fast keeps the client, and a goodwill discount costs far less than the job. The check stops the next confident wrong figure.

What it costs: A small goodwill discount.

Done by lunchtime.

You ring before the landlord has told anyone the price.

“Fair enough. Thanks for being straight with me.”

The landlord

You knock a little off as goodwill. The rule goes into the AI policy that afternoon.

Ciara still uses AI. She checks the sums against the fee schedule before sending, and a second person checks them too.

AI drafts, people check. Own a mistake fast and it costs you a little, not the client.

The traps

  • Sounds right: Honour the price and say nothing. It was the firm’s mistake

    It feels honourable, but the job runs at a loss, the team works evenings to fit it in, and next year the landlord expects the same price.

    What it costs: A job run at a loss, and the team’s evenings.

  • Sounds right: Correct it with an apology, and ban AI from anything that goes to a client

    The ban costs more than the mistake: reports go back to being written from scratch, and some of the team quietly return to free tools on their phones.

    What it costs: Slower work for everyone, and AI used out of sight.

  • Clearly wrong: Write to the landlord that the AI got it wrong, not you, so the quote doesn’t count

    The proposal had your name on it. “The AI did it” tells a client nobody was checking, and he gives the job to another firm.

    What it costs: The job, and the client.

Stage 5 · Friday, 4:50pm

The voice on the phone

Sarah from accounts is at your door, phone in hand.

“Brian’s just rung from Spain. His phone’s dying. The Carrick Road deposit has to go today or we lose the contractor. The bank details are in his email.”

Sarah

It was Brian’s voice: his laugh, his way of saying “grand”. The email has his name on it. The deposit is the biggest payment Sarah will make this month.

Impersonating the boss to get a payment made is old. UK Finance calls it CEO fraud, and still counts cases every year.

In the Government’s survey, 12% of UK businesses had someone impersonate them or their staff in emails or online. Now the voice can be faked too.

What do you tell her?

The best answer: Thank her, ring Brian on the number you already have, and make it the rule: no payment without a call-back and a second approval

A call-back to a number you already hold beats any voice, email or deadline. Brian made no call, and Sarah’s question saves the deposit.

What it costs: A short briefing for the team.

4:56pm…

Brian answers from a sun lounger. He made no call.

You thank Sarah. She stopped the whole deposit going to a fraudster by asking one question.

The email goes to the NCSC at report@phishing.gov.uk, and the attempt to Report Fraud, which replaced Action Fraud for Northern Ireland.

On Monday the call-back rule goes into the policy, and everyone hears it in person.

A call-back to a number you already hold beats any voice, any email and any deadline.

The traps

  • Sounds right: Have Sarah reply to Brian’s email to confirm, and pay if he says yes

    The reply goes to the look-alike address, straight to the criminal, and “Yes, pay it” comes back in seconds. Checking on the channel the request came in on checks nothing.

    What it costs: The whole deposit, gone to a fraudster.

  • Sounds right: Refuse to pay, and tell Sarah she should have known better than to even ask

    Right answer, wrong lesson. Months later an urgent call comes from “you”, and Sarah, remembering Friday, pays it without asking anyone.

    What it costs: A later payment made without anyone checking.

  • Clearly wrong: It’s Brian’s voice and Brian’s email, and he’s a partner. Tell her to pay it

    A voice can be copied from a podcast clip, and the email came from a look-alike address. By Monday the deposit has moved on.

    What it costs: The deposit, and days with the bank and the police.

Stage 6 · Six months on

The big contract

Mossvale is shortlisted for its biggest contract yet: stock-condition surveys for a housing association.

The reports will include tenants’ names and addresses.

The team wants AI to speed up the write-ups.

And the housing association’s contract asks for something in writing:

From the contract

“Explain how any AI you use will handle our tenants’ data.”

How do you handle it?

The best answer: Agree in writing which AI handles client data, set it up on business accounts limited to the project team, and walk them through it

Your answer fits in one paragraph: which AI, where tenants’ data is held, who can see it and what may go in. Their data protection officer has no further questions.

What it costs: Some IT set-up for the project team.

Contract signed.

Your answer to the housing association fits in one paragraph.

Which AI you use. Where tenants’ data is held. Who can see it. How long it’s kept. What staff may and may not put in.

“That’s clear. No further questions.”

Their data protection officer

Even the best path isn’t free: IT set-up for the project, business licences for the people on it, and some of their time to learn the rules.

Reports go out faster. The project folder is locked to the people working on it. And nobody is drafting on their phone at lunch.

Policy, an approved and secure AI, tidy permissions and a trained team: that's what lets you say yes to AI and to the client.

The traps

  • Sounds right: Tell the team to take tenants’ names out first, then use whichever AI they prefer

    A good habit, not a control. Report 37 goes into a free tool on a personal account with an address in a photo caption, and the contract is paused.

    What it costs: The contract paused, and its fees with it.

  • Sounds right: Keep AI away from this client’s work altogether

    Two hundred reports and a tight deadline. The team works evenings, and by week three some are quietly using AI anyway, on their own devices.

    What it costs: Overtime to meet the deadline, then shadow AI anyway.

  • Clearly wrong: Write “we don’t use AI” in the contract, and let the team carry on as they are

    It isn’t true, and their audit finds out. The contract is paused while they decide whether to trust anything else you’ve told them.

    What it costs: The contract paused, and the client’s trust.

How Mossvale could have been better prepared

On paper, Mossvale was sensible. Every gap the story found was in that list, and none was about how much AI the firm meant to use: the team was using it already. The cost is rarely only an invoice. It’s client data you can’t get back, a client’s trust and the contract, and questions from the regulator. Here is what would have changed.

  1. Approved AI on business accounts. Before: ChatGPT ‘used sensibly’ on personal accounts, and two Copilot licences for the partners. Better prepared: Copilot Chat on work accounts for everyone, and one approved AI for client work with its terms in writing: what’s kept, for how long, who can read it and where it’s held.
  2. Access tightened before Copilot. Before: Shared folders open to ‘everyone’, the HR folder included. Better prepared: Permissions tidied and company-wide sharing links removed before any AI searches the files, so it only shows people what they should see.
  3. Labels and data-loss rules. Before: Nothing marking which files are confidential, nothing to stop a paste, and no breach log. Better prepared: Client and HR files labelled, with rules that warn or block when confidential data heads for an unapproved AI, and a log for any slip.
  4. A human check on AI-drafted work. Before: Experienced people writing and sending their own quotes, checked by nobody else. Better prepared: A named person checks every figure in AI-drafted work against its source before it goes to a client.
  5. A call-back rule for payments. Before: Supplier payments approved by a partner, by email or a quick call. Better prepared: No payment or bank change without a call-back to a number you already hold and a second person’s approval.
  6. An AI policy and a trained team. Before: An acceptable-use policy written before AI, signed on day one, and a team trusted to be sensible. Better prepared: One page on which AI to use and what never goes in, walked through in person, and short, blame-free sessions that reward reporting a slip.

The risks, and the lessons

What is at stake at each stage

  • Client and confidential data in free AI tools: a possible reportable breach under UK GDPR, the ICO’s attention and lost client trust.
  • Copilot surfacing files people shouldn’t see: an internal data leak, from the pay review to a client’s confidential dispute.
  • A data breach hidden or handled in a panic: a breach reported late or not at all, the ICO’s attention and a client lost.
  • AI-drafted work going out unchecked: a wrong quote or wrong advice, the liability for it, and the client.
  • A cloned-voice or deepfake payment request: money sent to fraudsters, often for good.
  • Client data in AI the client never agreed to: a paused contract, an answer their audit proves false, and lost trust.
  • The best answer at stage 1: Company data stays in an AI the business controls, under its Microsoft 365 terms.
  • The best answer at stage 2: Copilot only shows people what they were always meant to see.
  • The best answer at stage 3: The breach is logged, assessed and reported in time, and the client hears it from you first.
  • The best answer at stage 4: Every AI-drafted figure is checked against its source before a client sees it.
  • The best answer at stage 5: No payment leaves without a call-back to a number you already hold and a second approval.
  • The best answer at stage 6: The client’s data goes only into the AI you agreed in writing, open to the project team alone.

Other AI risks to have a plan for

  • Shadow AI: Staff using AI apps, browser extensions or meeting note-takers nobody approved, some of which record calls and upload them.
  • Made-up answers: AI states a wrong fact or figure as confidently as a right one, and the firm is liable for what it sends.
  • Prompt injection: Hidden instructions in an email, document or web page can trick an AI assistant into leaking data. The NCSC says it can be reduced, not fully prevented.
  • AI with too much access: An assistant connected to email, files or payments can act on the wrong instruction. Give it only the access it needs.
  • Personal AI accounts: Everything ever pasted into a personal AI account is exposed if someone breaks into that account.
  • Chats kept longer than you think: Free AI tools can keep chats, have people review them and be ordered by a court to preserve them, outside your own retention rules.
  • Client contracts: Some clients forbid putting their data into AI tools at all. Check the contract before anyone does.
  • Unfair decisions about people: Using AI to screen CVs or customers can be unfair, and the ICO expects decisions about people to be fair and explainable.

Lessons learned

  1. Approve one business AI tool, such as Copilot Chat on work accounts, and say so in a one-page AI policy.
  2. Never paste client or personal data into free AI tools or personal accounts.
  3. Tidy file permissions before you switch on Copilot or any AI that searches your files.
  4. When data slips, thank whoever told you, log it, assess the risk and report it in time if someone could be harmed.
  5. A person checks every figure and claim in AI-drafted work before it reaches a client.
  6. Confirm any urgent payment or bank change by calling back on a number you already hold, with a second approval.
  7. Keep a list of approved AI tools, review it regularly, and train the team to spot prompt injection and deepfakes.

Three things to take away

  1. Doing nothing, banning AI and leaving staff to choose all end the same way: company data in AI tools you don't control, and time lost besides.
  2. AI makes old risks sharper: a confident wrong figure, a familiar voice asking for money. Check AI-drafted figures, and confirm payments on a number you already hold.
  3. Safe AI is a few things together: an approved AI on business accounts, tidy permissions and labels, a policy people know, and a team that reports slips without fear.
Spoiler: the short answer

Is it safe to paste company information into ChatGPT?

Not on a personal or free account. Public AI tools can keep what you paste, have people review it and be ordered by a court to preserve it—even when they promise not to train on it. Public or anonymised information is fine in any AI; client and confidential business information belongs only in an AI your business has approved and secured; passwords and keys never go in. The safe approach is a written AI policy, an approved, secure AI tool, tidy file permissions, a person checking what AI drafts before it goes out, and a trained team that reports slips.

Your next step to TechState™

Ready to run your business on its technology?

Book your Business Technology Review. In one focused conversation, we'll understand your business, assess your platform and show you the path to operating in TechState™.