They Can Use the Tools. They Cannot Judge the Output.
Middle managers across most enterprises have reached a baseline level of AI tool proficiency. They can write prompts. They can generate summaries, draft emails, analyse data, and create presentations using AI assistants. The "how do I use this" phase, for the most part, is behind them.
What remains is a subtler and more consequential gap. They have not developed the judgement to know when the AI output is trustworthy and when it is not.
This gap is not hypothetical. It shows up in decisions every day. A manager accepts an AI-generated analysis at face value and presents it to leadership without verifying the underlying assumptions. A team lead uses an AI-drafted recommendation for a client without recognising that the model hallucinated a statistic. A department head approves a process change based on AI-generated efficiency projections that were plausible but inaccurate.
Each instance is individually small. Collectively, they represent a systemic risk that most organisations have not yet addressed, because the training programmes they offer are still focused on the wrong layer.
Why Prompting Is the Easy Part
Prompting is a mechanical skill. It can be taught in a half-day workshop. Write clear instructions. Provide context. Specify the output format. Iterate on the result. The learning curve is short, the feedback loop is immediate, and most knowledge workers pick it up quickly.
Judgement is a cognitive skill. It requires understanding what the AI is doing well enough to evaluate its output critically. That means knowing the tool's limitations, recognising common failure modes, understanding when the output is likely to be reliable and when it is likely to be unreliable, and having the domain expertise to catch errors that the AI presents with confidence.
According to Gartner's 2025 research on AI adoption in the enterprise, the risk of over-trust in AI outputs increases with user familiarity. As employees become more comfortable with AI tools, their tendency to critically evaluate outputs actually decreases. They develop a learned confidence in the technology that outpaces the technology's actual reliability.
For middle managers, this dynamic is particularly dangerous because they sit at the decision layer of the organisation. Their judgements cascade into team actions, resource allocations, and client commitments. A flawed AI output that a junior team member accepts and shares internally has limited blast radius. A flawed AI output that a middle manager accepts and acts on can ripple across the function.
The Three Dimensions of AI Judgement
AI judgement for middle managers operates across three dimensions.
Evaluating. Can the manager look at an AI-generated output and assess its quality? This includes checking whether the data cited is real, whether the reasoning is sound, whether the output is appropriate for the context, and whether there are obvious gaps or errors. Evaluation requires enough domain expertise to recognise when something looks right but is wrong, which is the failure mode AI is most prone to.
Validating. Can the manager confirm the output against independent sources or organisational knowledge? Validation goes beyond evaluation by adding a verification step: checking the AI's claims against real data, cross-referencing with subject matter experts, or testing the output against known baselines. This dimension catches errors that look plausible on the surface but do not hold up under scrutiny.
Deciding. Can the manager make a sound decision about whether and how to use the AI output? Deciding includes choosing to accept the output, reject it, modify it, or escalate it to someone with deeper expertise. It also includes knowing when an AI-assisted decision requires human override, when the stakes are high enough that AI output should be treated as a starting point rather than a conclusion, and when the output is reliable enough to act on directly.
Why Current Training Misses the Mark
Most AI training programmes for middle managers are structured around tool proficiency. They teach employees how to use specific platforms, how to write better prompts, and how to integrate AI tools into their workflows.
These programmes address a real need. Tool proficiency matters. But they rarely include the judgement dimension. They do not teach managers how to catch a hallucinated statistic in an AI-generated report. They do not practise the skill of evaluating whether an AI recommendation is appropriate for the specific context. They do not build the confidence to override AI output when domain expertise suggests the output is wrong.
The gap persists because judgement is harder to teach and harder to measure. A prompt engineering workshop has clear learning objectives and observable outcomes. A judgement development programme requires scenario-based learning, repeated practice with ambiguous situations, and assessment methodologies that evaluate critical thinking rather than procedural knowledge.
Our AI Judgement Programme takes a different approach. It trains managers to evaluate, validate, and act on AI outputs with confidence, using scenario-based exercises built around real decision contexts from their own organisation. The goal is not proficiency with any specific tool. It is the cognitive capability to work with AI as a decision partner rather than as an oracle.
The Consequence of Not Addressing This Gap
The risk of the judgement gap compounds over time. As AI tools become more capable and more embedded in organisational workflows, the volume of AI-influenced decisions grows. Each decision that passes through a manager who lacks the judgement to evaluate the AI's contribution is a decision that carries unquantified risk.
The organisations that address this gap early will build a management layer that is genuinely augmented by AI: faster, better-informed, and more confident. The ones that focus only on tool proficiency will build a management layer that is faster but not necessarily better, producing more AI-influenced decisions without the quality control that judgement provides.
If your organisation has trained managers on prompting and is ready to address the deeper gap, we can help.
Sources: Gartner. "How to Overcome Digital Workplace Resistance, 2025." https://www.gartner.com/en/information-technology/topics/digital-workplace McKinsey & Company. "The State of AI in 2025." https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai