Skip to content
AI adoption in UK businesses is accelerating, but trust hasn’t kept pace with uptake — most people remain wary of security risks and reluctant to hand AI systems full autonomy. This piece from Duco Digital Training lays out seven practical questions organisations should answer before rolling out AI tools more widely, covering data safety, accountability, GDPR compliance, output verification, staff training, data quality, and security testing. The throughline: getting AI right isn’t about moving slower, it’s about moving with eyes open.
AI has moved fast. Barely two years after generative tools became a boardroom topic, government data now shows that 41% of UK businesses handling digitised data are using some form of AI, rising to well over half among medium and large organisations. But adoption isn’t the same as trust, and for organisations working with sensitive research, client, or operational data, trust is the harder problem to solve.
Recent industry research bears this out. EY’s 2026 AI Sentiment Index found that 73% of people are concerned about AI systems being hacked or breached, and only 14% are comfortable with AI acting fully autonomously without human oversight. Separately, Infor’s Enterprise AI Adoption Impact Index found that 45% of UK professionals rank data security among their top concerns about scaling AI. The technology is ready before the governance is.
So before your organisation rolls out its next AI tool, or discovers that teams already have, here are seven questions worth answering honestly.
1. Is it safe to put your data into AI tools?
Many AI tools are built on third-party models hosted outside your organisation, sometimes outside the UK entirely. Before approving a tool, it’s worth knowing: is data used to train the underlying model? Where is it processed and stored? What happens to the data if the contract ends?
For scientific labs, this is rarely just a data question; it’s an IP question. Unpublished results, method details, and proprietary formulations pasted into a public AI tool can end up stored, reviewed, or drawn on for model training, on a server your organisation has no visibility over or control of.
A few quick lockdown steps worth having in place before AI gets anywhere near lab work:
- Use enterprise or business-tier accounts with a contractual guarantee that inputs aren’t used for training, ChatGPT Business/Enterprise, Microsoft 365 Copilot, Claude Team/Enterprise, and Google Gemini for Workspace currently all offer this; the free or personal-account versions of the same tools generally don’t
- Keep a short, named list of approved tools; nothing else goes near research or client data
- Set clear rules on what information or data can’t be used, for example, raw datasets, unpublished manuscripts, sample or patient IDs
- Turn off chat history and data retention wherever the tool allows it
None of this is difficult to set up, but it needs to happen before the first upload, not after.
For the most sensitive material such as trial data, unpublished IP, regulated client records, some organisations go a step further and run an open-weight model on their own servers, so nothing leaves the building at all for inference. It’s a legitimate option, not an automatic upgrade: local models generally lag behind the best cloud tools on complex reasoning, and running one properly takes real infrastructure and in-house expertise to maintain. Most organisations that go this route use it for their most sensitive workflows specifically, alongside cloud tools for everyday work, rather than replacing cloud outright.
2. Who’s accountable when AI gets something wrong?
AI-generated outputs can look confident and be wrong, whether that’s a misread dataset, a flawed summary, or a decision recommendation built on incomplete context. Accountability doesn’t disappear just because a machine was involved, but it’s rarely just one person. In practice, it tends to run in layers: the individual who used or approved that specific output, whoever set the process or signed off on the tool being used for that purpose, and where something goes wrong systemically rather than as a one-off; the directors and leadership responsible for making sure adequate oversight existed in the first place. A single error by one employee is a very different conversation to a pattern of unchecked AI use that leadership never put guardrails around. What this looks like in practice varies by organisation, and by use case; there’s no single template. But most organisations that manage this well have some combination of:
- A small AI governance group, sometimes called an AI board or committee, that reviews new use cases and sets the ground rules, rather than leaving each team to decide independently
- A named owner for AI-related decisions in each area, which is often an existing manager with AI added to their remit rather than a new hire
- One or two “AI champions” per team who understand both the tools and their limits, and act as the first point of contact when something looks off
A five-person lab and a 300-person diagnostics company will land on very different versions of this. The structure matters less than the fact that one exists, and that everyone knows who it is.
3. Is your use of AI still GDPR compliant?
UK data protection law has kept moving. The Data (Use and Access) Act 2025 has already amended parts of the UK GDPR and Data Protection Act 2018 framework, and the ICO continues to issue fresh guidance on AI-specific processing. A tool that was compliant when someone last checked may not be compliant now. Treat data protection compliance for AI as a recurring review, not a one-off tick-box exercise.
4. Can you trust what AI tells you?
It helps to understand what’s happening under the hood in an AI model. At their core, large language models work by predicting the next most likely word or phrase, based on patterns learned from huge amounts of text; they aren’t reasoning the way a person does, and they don’t “know” something is true so much as generate a response that statistically resembles a good answer. That’s why they can present incorrect information with the same confidence as correct information, and it’s precisely what makes unverified AI output risky in any context involving safety, compliance, or client-facing decisions. The organisations getting the most value from AI tend to be the ones that treat it as a fast first draft or a research assistant, not a source of truth, and build in a verification step before anything goes further.
5. Do your staff need AI training?
This is the question most organisations haven’t asked yet, and it may be the most important one. The UK’s Cyber Security Breaches Survey 2025/2026 found that roughly a third of UK businesses are using, adopting, or considering AI, and of those, only 24% have any formal process in place to manage the associated cyber risk. Globally, Verizon’s 2026 Data Breach Investigations Report found unsanctioned “shadow AI” use was a factor in 45% of breaches, roughly triple the year before.
A written usage policy helps, but a policy on its own doesn’t teach anyone what a safe prompt looks like, or why a confident-sounding answer might still be wrong. The organisations getting this right treat AI literacy as a genuine training need, a professional course in AI, AI ethics, or responsible use, rather than a five-minute induction slide, so staff understand not just what they’re allowed to do, but why the boundaries exist. That understanding is what closes the gap; the policy just writes it down.
6. Why does data quality affect AI results?
AI output is only as reliable as the data behind it; feed it incomplete, outdated, or poor-quality data, and you’ll get confident-sounding nonsense back. Before trusting an AI-assisted result, it’s worth asking the same questions you’d ask of any dataset: where did this data come from, how was it collected, and is it fit for this purpose?
The same logic works in reverse. Holding onto more data than you need doesn’t make AI tools smarter, it just increases what’s exposed if something goes wrong. Organisations getting this right actively review what they hold, deleting data that’s outdated, poor quality, or no longer needed, rather than letting it accumulate indefinitely. Less data held means less risk if a tool, account, or vendor is ever compromised.
7. How do you know if your AI security works?
Trusting AI in your organisation is part of a wider security posture, and it’s worth checking the fundamentals are in place rather than assumed to be:
- Independent standards such as ISO/IEC 42001, the international AI management system standard, give a structured framework rather than an ad hoc policy
- Cyber Essentials certification covers the baseline technical controls that most breaches still exploit
- Upskilling your team in security or IT security architecture is an effective way of keeping your team aware and helping protect your organisation
- Random, unannounced checks on whether staff are following the AI usage policy, not just whether one exists on paper
- Periodic penetration testing or security testing across wider defences, not only the AI tools themselves
- Confirming IT architecture is set up correctly around any new AI tool, rather than assuming it inherits the security of the systems around it
- Keeping systems patched with trusted, vendor-issued security updates, still one of the most common gaps behind breaches
The NCSC’s guidance on AI and cyber security is a useful, non-technical starting point for board members and managers working through this.
The reward for getting this right
These questions are all about AI governance: knowing where your data goes, who’s accountable, whether you’re compliant right now, whether outputs get checked, whether your people are properly trained rather than just policed, whether the data behind it all can be trusted, and whether your wider defences would hold up under real scrutiny. Organisations that get this right don’t necessarily move slower on AI; they move with their eyes open. That’s a fair trade for the level of trust regulators, clients, and your own staff are asking for.
Darren Winter is Company Director of Duco Digital Training, a BCS-accredited provider of courses in Responsible AI, Information Security, and Data Protection for individuals and organisations across the UK. He is also a PhD candidate in AI Ethics at the Universidad de Córdoba.
