Are AI outputs too unreliable for real work?

One in five UK banking and financial services workers in a Skillcast survey received no AI guidance from their organisation in the past year. Yet 72% of those surveyed use AI every single working day. That gap, not the technology itself, is what produces the inaccurate outputs that make Operations Managers, HR Managers, and Finance Managers nervous about trusting AI with anything that matters. The evidence is clear: when organisations pair structured training with sensible governance, accuracy concerns give way to confidence in appropriate use. The technology has not changed. The context around it has.

Why the unreliability myth took hold

The concern is reasonable. A Skillcast survey of UK banking and financial services staff found that 32% report AI producing inaccurate or misleading results at least half the time. Accuracy ranked as their primary worry, above privacy and above security. When frontline staff share stories of AI confidently stating something wrong, it spreads quickly. Leaders hear those stories and make a rational judgement: if it gets things wrong that often, we cannot risk it on client-facing work or regulated tasks.

That judgement is sensible. The mistake is in the diagnosis. The survey does not show that AI is fundamentally broken. It shows that large numbers of people are using AI tools daily without any shared understanding of where those tools are reliable and where they are not. That is a training and governance problem. It has a known fix.

What the evidence actually shows

The Skillcast data tells a more nuanced story than the headline number suggests. Yes, 32% of workers report frequent inaccuracies. The same survey found that one in five respondents received zero organisational guidance on AI use in the past year. Consider what that means in practice. Teams are reaching for AI tools to speed up their work, which is a sensible instinct, but with no shared framework for which tasks suit AI well, which require careful checking, and which should stay entirely with a human.

The accuracy problem clusters around predictable failure modes. AI models are strong at drafting, summarising, spotting patterns, and generating first-pass research. They are weaker when asked to reason about very specific regulatory details, perform precise calculations without structured inputs, or retrieve facts they were not trained on. When staff do not know this distinction, they apply AI indiscriminately. Some of those applications will produce errors. The errors then confirm the belief that AI cannot be trusted at all.

This is the real finding buried in the Skillcast survey: the problem is not the tool. It is the absence of a shared map showing where the tool is useful and where it needs human oversight.

What this costs UK SMEs in practice

For an Operations Manager, a Finance Manager, or an HR Manager at a UK SME, the consequences of this myth run in two directions. Both are expensive.

The first consequence is over-caution. A team that has heard enough hallucination stories will avoid AI entirely for anything important. That means hours of manual work that structured AI use could reduce. Drafting supplier communications, summarising meeting notes, reviewing policy documents for key clauses: these are exactly the tasks where AI is reliable under human review, and exactly the tasks that eat into the working week.

The second consequence is under-caution. Without guidance, some staff will use AI for tasks it handles poorly, trusting outputs without checking them. In a regulated context, that creates real compliance exposure. The FCA and ICO both expect organisations to maintain human responsibility for final decisions and to be able to evidence their review processes. An undocumented, untrained approach to AI use does not meet that standard.

Both consequences share the same root cause: no one has sat the team down and explained how AI works, where it adds value, and what responsible use looks like for this specific organisation.

A realistic approach to building accuracy confidence

The intervention the evidence points to is straightforward. It does not require replacing existing tools or running a large-scale change programme. It requires four things.

1. Define the task categories. Separate tasks into three groups: AI-appropriate with light review (internal drafts, summaries, research starting points), AI-appropriate with thorough review (anything external or client-facing), and AI-excluded (final regulated decisions, calculations requiring legal precision). Write this down and share it.

2. Train staff on prompt design. Most inaccurate outputs follow from vague or ambiguous prompts. Teaching staff how to give AI clear instructions, provide relevant context, and specify the format they need reduces error rates significantly. This is a skill that improves with practice and structured guidance.

3. Pair AI tools with authoritative internal data. AI models produce stronger outputs when they have access to accurate, current information about your products, clients, and processes. Building that connection is a deliberate design choice, not a default setting.

4. Measure near-misses over time. Track where AI outputs needed correction and why. Over six months, this builds an evidence base showing which use cases are genuinely reliable for your team and which need more oversight. It also demonstrates to regulators that your approach is managed and documented.

This treats AI as an assistant whose work is checked, not an autonomous decision-maker. That framing matters. It aligns with how the most effective teams actually use these tools, and it matches the expectation UK regulators have set for human accountability in AI-assisted processes.

If your organisation wants structured support building this framework, gecco's Training and consultancy programme gives teams the practical grounding to use AI accurately and responsibly. The Essentials tier covers AI fundamentals, the GRAFT methodology, and hands-on assistant building across three days. The Complete tier adds autonomous tools and an applied session over five days. Delivery is online, in-person, or hybrid, and gecco has trained 1,000+ professionals with a 92% onboarding cohort completion rate.

Honest caveats worth keeping in mind

The training-and-governance fix is real, but it is not instant and it is not total. AI models do still produce errors, even when used well. The goal is not zero inaccuracy. It is a managed process where errors are caught before they cause harm, and where the reliability of AI in appropriate use cases is high enough to justify the time it saves.

The Skillcast data comes from banking and financial services, a sector with higher regulatory stakes than many SMEs face. Law firms, accountants and insurers will recognise the same pressure: regulated advice, client data and a duty to evidence review. The accuracy concerns are genuine in that context. For most SME use cases, such as internal communications, HR documentation, operations planning, and finance reporting at the draft stage, the risk profile is lower. The same principles apply: define the task, train the team, review the output. The standard of review required scales with the stakes of the decision.

There is also a meaningful difference between AI tools used without any configuration and AI assistants built around your organisation's specific context, documents, and processes. A well-built assistant with access to your internal data and a clear brief performs considerably more accurately than a generic prompt sent to a public model. That distinction matters when setting expectations with your team.

Your next step

If your team is weighing up whether AI accuracy is a technology problem or a training problem, that is exactly the question the AI Readiness survey is built to surface. It takes around ten minutes and generates a custom report showing where your organisation sits on each dimension of AI readiness, including governance and staff capability.

Take the AI Readiness survey. You will get access to 65+ free resources and a custom AI Readiness report. We then offer a free 30-minute showcase to walk through your results.

One in five UK banking staff received no AI guidance, yet 72% use AI daily. The training vacuum is the real risk. gecco's Training and consultancy programme is built to close it, with vendor-neutral guidance matched to your team's actual tools and tasks.

---

Website · LinkedIn · Case Studies · Newsletter