How to let staff use AI safely: a four-week rollout for small firms
To let staff use AI safely, give them approved business accounts, a one-page policy, short training, automatic checks on what goes in, and a monthly review. A small firm can do all five in four weeks. By the end, staff use AI openly, client data stays out, and you can show a client or insurer how.
AI is welcome here. Company and client data stay out of it. That is the whole aim.
Why not just ban it
Bans rarely hold. Staff find a free tool on their phone instead. The NCSC - The hidden risks of shadow AI says organisations should give employees AI tools that meet their needs while managing cyber risk, rather than trying to ban AI outright. Its reason is plain. Shadow AI use is unlikely to disappear completely.
So the plan below assumes people will use AI. It just makes the safe way the easy way.
Week 1: choose approved AI tools
Start by finding out what people already use. Ask them. Most will tell you. Our guide on finding out which AI tools your staff are using covers quieter methods too.
Then pick one or two tools and buy business accounts. Free personal accounts are not the place for work data. A plain-English reading of The NCSC's guidance on generative AI, in plain English sorts tools into tiers. Approved tools for non-confidential material and anonymised examples. New tools cleared with the owner first. Business-tier accounts for work data.
What the main business plans cost
| Tool | Published price | Notes |
|---|---|---|
| ChatGPT Business | $25 per user per month, billed annually (around $30 monthly) | 2-seat minimum |
| Microsoft 365 Copilot Business | $21 per user per month, annual plan | Up to 300 users. On top of Microsoft 365 |
The ChatGPT figures come from Areebi - ChatGPT Enterprise Pricing in 2026. The Copilot figures, including an $18 first-year promotional rate, come from the Microsoft 365 Copilot Pricing 2026 guide. Check both before you buy. Prices move.
Rule of thumb: if a tool is not on the approved list, it does not get work data. One list, kept short.
Week 2: write a one-page AI policy
Keep it short enough to read in two minutes. It should say:
- Which tools are approved.
- What must never go in: client names, bank details, National Insurance numbers, passwords, health details.
- How to ask for a new tool.
- That a person checks AI output before it goes to a client.
- Who to tell if something goes wrong.
Our UK AI policy template gives you a starting draft.
Check your data protection duties
The ICO is clear that there is no 'AI exemption' to data protection law. Even incidental processing of personal data counts, as Hamlins LLP - Decoding the ICO's Generative AI guidelines explains. The ICO's own guide, ICO - How to use AI and personal data appropriately and lawfully, advises a data protection impact assessment (DPIA). That is a written check of the risks and how you reduce them. Under the accountability principle, your firm is responsible. Not the AI company.
Week 3: train staff
Thirty minutes is enough for most teams. Show, do not lecture. Take a real task and do it twice.
At Ashcombe, Priya wants a polite chase for an overdue invoice. The unsafe way pastes the full email with the client's name and account number. The safe way writes "Client A, invoice overdue 30 days" and asks for a firm but friendly reply. The question still goes. The answer is just as good.
Then ask each person to agree the policy in writing, with a date. Note who missed the session.
Week 4: switch on automatic checks
Training fades. People rush. A paste at 5.40pm on a Friday is the one that slips. Automatic checks catch what memory misses.
Good checks share a few traits:
- They run before anything is sent.
- They take sensitive items out rather than block the whole question.
- They record what kind of data was involved, not what was typed.
- They cover every AI site staff use, not just one.
That last point matters. You want a record of events, not a copy of every message. Our piece on monitoring AI use without spying on staff explains the difference.
If you use AI agents
Some firms go further and let AI tools act on their own, sending emails or editing files. In August 2026 the NCSC published interim guidance, "Managing the cyber risk of agentic AI". It sets out seven considerations, covering threat modelling, sandboxing, observability, human oversight and emergency shutdown, as reported by Infosecurity Magazine. It also advises temporary credentials over long-lived ones, removing elevated access once a task is done, and watching for unusual activity. For most small firms, agents can wait until month two.
Every month after: review
Set a fixed slot. Thirty minutes on the first Monday. Look at:
- How many people have agreed the current policy.
- Which kinds of data were caught, and on which sites.
- Any unapproved tools that appeared.
- Who needs a refresher.
Keep the notes. They are your answer when a client, insurer or auditor asks about AI.
What good looks like after a month: approved accounts in use, a dated policy agreement for everyone, checks running on every PC, and one short monthly record you could send to a client today.
Where Helio Guard fits
Helio Guard covers weeks three and four, and the monthly review. It is a browser extension for Chrome and Edge on Windows PCs. It checks every paste, message and file upload on listed AI sites. Sensitive items are swapped for labels, so the question still goes. Client names can become a stand-in such as "Client A", then turn back when the answer is copied. Checked on the PC with fixed rules. We never see what your team types. It also records policy agreements and produces the monthly record. It costs £1 a person a month, from 5 people, with a 14-day free trial and no card needed. The pricing page has the details.
Common questions about rolling out AI safely
Do small firms need a DPIA before using ChatGPT?
The ICO advises a data protection impact assessment where AI use involves personal data. For a small firm, a DPIA can be a short document. It should say what data goes in, which tool, what the risks are and how you reduce them. Keeping it on file also helps if a client or insurer asks.
Is a free ChatGPT account fine for work?
A free ChatGPT account is best kept to non-confidential tasks with no client details. The NCSC-based guidance points firms to business-tier accounts for work data. A business account gives the firm control over users and settings, which a personal free account does not.
How long does it take to let staff use AI safely?
Letting staff use AI safely takes about four weeks for a small firm: one week each to choose tools, write the policy, train staff and switch on automatic checks. After that, a thirty-minute monthly review keeps the rollout working as tools and staff change.
Related: How to evidence AI controls for audit, insurers and clients, Staff AI training: what to cover so people stop pasting client data