How Businesses Can Approach AI Security Without the Technical Jargon: Interview With Simon Heath of The Final Step
I’m Simon Heath, and at The Final Step, I introduce organisations to our services and help business leaders understand how technology can support their goals. I’m not a technical specialist; I focus on education, starting useful conversations, and connecting people with the right expertise on our team.
Founded in 1987, The Final Step helps small and medium-sized businesses in London fulfil their potential through better use of technology. We combine day-to-day IT support with cyber security and strategic guidance, working as an outsourced IT department or alongside internal teams. Our approach is practical and security-first: understanding what each business needs, helping leaders make informed decisions and keeping technology investment focused on business outcomes.
1) Which AI-related cyber security risks should UK small and medium-sized businesses take most seriously, and where do you think the threat is being overstated?
For most SMBs, the biggest AI risk isn’t a sophisticated new attack. It’s that the hype distracts from the basics. The NCSC expects AI mainly to make existing attacks more effective, yet only 24% of UK businesses have all five Cyber Essentials controls in place.
Three risks deserve real attention. Phishing and payment fraud are more convincing now that voices and faces can be cloned. AI assistants such as Copilot can see everything their user can see, so years of oversharing suddenly become easy to find. And staff are already using AI tools you haven’t approved.
What’s overstated? That AI is a bubble you can wait out, and that it will replace people. Banning it only hides it. AI works best when experienced people shape and check it.
Get the basics right, then give people clear guidance.
2) What would you say to a business owner who assumes their company is too small, or uses too little AI, to be at risk?
Size doesn’t protect you. AI lets criminals produce convincing, personalised scams at scale. It’s cost-effective to target individuals or small businesses with fewer protections, making them an attractive entry point, especially if they work with larger organisations.
“We barely use AI” isn’t that reassuring. AI is built into so much software, and staff are probably using free tools you don’t know about. Because AI can make things up or assume control of data, it can expose you to embarrassment and confidentiality and data protection breaches.
AI is good at discovering previously unknown vulnerabilities that affect the wider environment, such as the PHP risk that affected many websites. It had existed as a flaw for many years, and attacks began within an hour of its publication.
You don’t have to be a target to be a victim.
3) Can you share an anonymised example of an AI-related security incident or near miss you have encountered, and explain what other small businesses should learn from it?
Near misses are common, because we’re all learning a fast-moving technology. Nobody is exempt, myself included.
In one senior leadership meeting, it emerged that the finance director had put private company financial data into an AI tool while building a model and, in doing so, gave up control of it. Nobody was careless by intent. The partner responsible for IT assumed safe working practices were understood; the finance director assumed safeguards were in place.
The lesson is that assumptions undo us. Capable, senior people make these mistakes when they’re busy, so leaders need to set clear guardrails and talk about them.
Don’t assume everyone knows the rules; tell them.
4) How should staff training change to help employees recognise and respond to AI-assisted phishing, rather than relying on warning signs such as poor spelling or grammar?
AI improves phishing quality, so training needs to move from spotting bad writing to questioning the request. Is the message urging you to act quickly, keep it secret or skip the usual process? Those are the warning signs now.
Teach one habit above all: verify through a channel you choose, never the number or link in the message. Make it a rule with no exceptions, because scammers rely on you thinking it’s an exception. Provide extra training for finance, HR, and personal assistants; keep it regular and interactive, and make reporting easy and blame-free.
Train people to verify, not judge spelling.
5) What verification process should a business use for urgent payments or changes to supplier bank details, particularly when an AI-cloned voice or deepfake video could be used to impersonate someone?
Seeing and hearing is no longer believing: a cloned voice or a deepfake on a video call is convincing. Fraud works by isolating people and putting them under pressure, so protection must be a process that breaks both. Slow down and check.
Leaders should make it policy that no bank detail change, new payee, or unusual payment is approved based on an email, call, video meeting, or message alone, no matter who appears to be asking. Call back on a number you already have, require a second approver who wasn’t part of the original conversation, and send a test payment first. Then back staff who follow the process, especially when the “boss” is pushing.
If the only proof is a voice or a face, it isn’t proof.
6) Is AI changing the ransomware risk for smaller businesses, and which measures would you prioritise to reduce both the likelihood and impact of an attack?
AI makes ransomware attacks cheaper and faster to launch. Still, it hasn’t changed how attackers get in: a phished password, an account without multi-factor authentication (MFA) or an unpatched firewall.
To reduce the likelihood, enable MFA, patch internet-facing equipment quickly, and remove everyday admin rights. Cyber Essentials covers these, and the NCSC reports that certified organisations are 92% less likely to make a cyber insurance claim. To reduce the impact, keep an offline backup, test that you can restore from it, and have an incident plan.
7) What could go wrong when employees paste client information or internal documents into an AI chatbot, and how should a business decide which information is safe to share?
The risk isn’t only that AI repeats your secrets to strangers. It’s that once information goes into a public AI tool, you lose control of who can see it, how long it’s kept and what it’s used for. That can breach client contracts and data protection law.
Set your main AI tools not to train on your data. For public tools, use a simple test: if you wouldn’t want to see it in a newspaper, don’t give it the data. Definitely keep client details, personal data, financials and passwords out, and anonymise anything you need help with.
If in doubt, leave it out.
8) How can an SMB identify and manage “shadow AI”, tools employees use without company approval, without discouraging staff from being honest about how they work?
Assume it’s already happening. Microsoft found that 78% of AI users bring their own tools to work. Staff rarely do this to cause harm; they do it because AI helps and nobody has offered an approved option.
Start by asking, with an amnesty: which tools do you use, and for what? Then check what your systems show, accepting that detection tools miss things. Provide an approved tool, publish a list with a quick way to request new ones, and block only the high-risk cases. Thank people who own up, or they’ll stop telling you.
Make the safe option the easy option.
9) Before connecting an AI assistant to company emails, shared files or customer records, what access restrictions and security checks should be in place?
An AI assistant sees everything its user can see, so fix your permissions before you switch it on. Files shared with “everyone” years ago, forgotten HR folders and old client sites can become searchable in seconds.
Clean up oversharing, lock down sensitive areas and make sure every account has MFA. Stop staff granting AI apps access to mailboxes or drives without approval; we often meet capable people who click “Allow” without realising what they’ve agreed to. Start read-only and pilot with a small group.
AI can ignore instructions, or be tricked into ignoring them, so permissions are the real control.
Don’t tell AI what not to do; make sure it can’t.
10) What is “prompt injection” in plain English, and how could it cause a business’s AI assistant to disclose confidential information or take an unauthorised action?
Prompt injection is when someone hides instructions in content an AI reads, such as an email, document or web page, and the AI follows them as if they came from you. AI cannot separate what genuinely came from you and what didn’t.
The NCSC warns it may never be fully fixed. So limit what an AI can reach and do, and require a person to approve anything with real consequences.
Assume AI can be talked into things, and limit what it can reach.
11) For an SMB with a limited budget and no in-house security team, which existing protections should be strengthened before spending money on specialist AI security products?
Start with what you already pay for. Most AI-assisted attacks still need a weak password, an unpatched device or poor backups to succeed, and the fixes are largely free.
Turn on MFA everywhere; Cyber Essentials now fails any business without it on cloud services. Patch routers, firewalls and software promptly. Remove unnecessary admin rights, tidy up who can see what, and test your backups. Add payment checks, an approved AI tool and a short policy. The NCSC’s free Cyber Action Toolkit takes you through it step by step.
Spend time on the basics before you spend money on AI security.
12) What should a practical AI-use policy cover, and how can a small business make it part of everyday working habits rather than a document nobody reads?
A good AI policy fits on two pages and answers the questions staff actually ask: which tools can I use, what can I put in them, who checks the output, and what do I do if something goes wrong?
Making it stick matters more than the wording. A leader should own it, find out which tools staff already use, and launch it with a short team briefing rather than an email attachment. Name someone people can ask, and review it at least every six months. We offer a free template to start from.
A policy should enable as well as restrict. At one client, two top performers had used Copilot over 500 times each while most colleagues had used it fewer than ten times. It’s worth investigating whether Copilot is enabling their better results.
A policy nobody has discussed is a policy nobody follows.
13) Which actions should an AI assistant require human approval to perform, and how can those limits be enforced technically rather than simply written into its instructions?
Anything that moves money, sends information outside the business, deletes records or changes access should need a person’s approval. So should decisions about people, such as shortlisting job applicants: AI can repeat hidden biases, and UK GDPR requires safeguards for significant automated decisions.
The key is to enforce this in the systems the AI uses, not just in its instructions. An instruction is a request that AI can ignore or be tricked out of; in 2025, an AI coding agent deleted a live database despite being told not to. So give AI its own account with minimal permissions, set it to draft rather than send, and keep dual authorisation on payments.
An instruction is a request; a permission is a control.
14) Where can AI genuinely help a small business detect and respond to cyber threats, and where would relying on it create a false sense of security?
AI excels at what people can’t do: filtering huge volumes of email, spotting unusual logins, and sorting thousands of alerts. It doesn’t get tired, so it can watch around the clock. Good managed detection and response services have combined it with human analysis for years.
The false sense of security comes from treating “AI-powered” as a substitute for the basics, or assuming someone will act on an alert at 2am. Ask any provider what happens in the 15 minutes after a serious alert, and who does it.
AI can raise the alarm; people still need to answer it.
15) What questions should an SMB ask an AI supplier about data storage, retention, model training and security before trusting the service with business information?
Ask the questions that matter most, and get the answers in writing. Will our data be used to train your models? Where is it stored, and for how long? Can your staff see it? What certifications do you hold? Will you sign a data processing agreement, and how quickly will you tell us about a breach?
Keep it proportionate. Small businesses can’t negotiate every contract, and features change monthly. Do full checks only for tools that handle sensitive data or connect to your systems, and know how you’d leave before you need to.
If a supplier can’t answer clearly, that is your answer.
16) What should UK SMBs check about their UK GDPR responsibilities before using AI to process customer or employee information?
This is general information, not legal advice. Many of our clients have their own data protection expert, and we deliver the technical controls they specify.
Using AI doesn’t change your UK GDPR duties, whether you control the data or process it for clients. Be clear on the purpose and lawful basis, check whether you need a data protection impact assessment, and put the AI supplier under a data processing agreement. Update privacy notices, keep a person involved in significant decisions about people, and check client contracts before passing their data to a new AI service. The rules are evolving, so watch for ICO updates.
Settle the data protection questions before the data goes in.
17) When a business uses an external IT support provider, how should responsibility for approving AI tools, monitoring their use and responding to security incidents be divided?
The business owns the decisions and the risk; the IT provider owns the technical controls and the alerts. You can outsource the work, but not the accountability.
In practice, you decide which AI tools are acceptable and lead on data protection, client communication and any report to the ICO. Your provider checks suppliers, configures access, monitors use and contains incidents. Write it into the contract, including who can switch on new AI features and how quickly the provider will tell you about an incident.
Agree who does what before something goes wrong.
18) What should a business do in the first hour after discovering that sensitive information may have been exposed through an AI tool?
Stay calm, contain it and start the clock: if personal data is involved, you have 72 hours from discovery to report to the ICO.
Note the time and what happened. Disable any shared links, disconnect the AI tool from your systems, and reset exposed passwords. Take screenshots and export logs before deleting anything. Bring in your IT provider and whoever handles data protection, work out what was exposed and who is affected, and tell your insurer early.
Above all, thank the person who reported it. Blame now means silence next time.
Contain, preserve, assess, and thank whoever told you.
19) How should an SMB prepare to keep operating if an AI-related security incident disrupts its email, payment systems or a critical supplier?
You need two plans. Disaster recovery gets you fully back on your feet; business continuity keeps essential work going while that happens, which can take weeks.
Identify your critical processes and what they depend on. Keep an offline contact list, agree a backup way to communicate, and know how to make urgent payments with extra checks, because fraudsters target businesses in crisis. Ask key suppliers about their plans.
Then practise. Restore from a backup to see where it breaks, and run a tabletop exercise on who tells whom, and when. The NCSC’s free Exercise in a Box can help.
An untested plan is a hope, not a plan.
20) What three actions would you recommend a UK small-business owner take over the next 30 days to reduce AI-related security risks without losing the benefits of using AI?
First, lock down logins and payments. Turn on MFA for every staff login, and make it a rule that no payment or bank-detail change goes through without a callback to a known number.
Second, give staff a safe way to use AI. Ask what they use, provide an approved business tool and publish a short policy.
Third, check what AI can see, and be ready. Tidy up access before switching on assistants, test a backup restore and write a one-page incident plan.
Then look for opportunities: decide where AI could make the biggest difference, pilot it, and measure the return.
Secure the basics, then make AI work for you.






Comment