I discussed in a previous blog how one can make AI safe via well-established engineering techniques. This isn’t the usual approach. A more typical way to mitigate risks of AI hallucination, bias, unsafe actions, or loss of accountability is simply to put a human in the loop. But what does this even mean? And does it work?

What does oversight really mean?

Appropriate human oversight is one of the requirements placed on high-risk systems by the EU AI Act, alongside traceability, documentation, accuracy and robustness. Responsible-AI frameworks from NIST and the European Commission similarly emphasise human agency, clearly defined responsibilities and the ability to understand, challenge and override AI-supported decisions.

However, none of this thoughtful legislation and guidance explains what humans in the loop are expected to do in practice. Check every fact? Reconstruct every inference? Notice missing evidence? Recognise when an apparently plausible answer is subtly wrong? Resist the temptation to accept a high-powered recommendation that is expressed in confident, elegant prose?

And how can they do all this without taking so much time that the efficiency saving of using AI is eliminated?

The cognitive trap

The basic challenge is that we use AI because the task is difficult, the information required is extensive, and / or the connections within the information are complex. We use AI to remove these barriers. Then we put them all back on the human in the loop.

To make matters worse, when the AI is wrong the clues may be well-hidden. False statements, inferences, and conclusions will be presented with the same fluency, structure and confidence as correct ones. The 2026 International AI Safety Report notes that although current AI systems may perform impressively on expert-level benchmarks, they remain entirely capable of factual and logical errors, inconsistency and fabricated citations.

This exposes the human reviewer to automation bias: our tendency to defer to an automated answer, particularly when the system usually performs well, the answer looks plausible, and checking it would require extra effort. The International AI Safety Report concludes that automation bias can discourage active reasoning and verification, causing people both to overlook problems and to act on incorrect automated advice. The effect varies according to the task, interface and degree of accountability, but it has been observed in settings ranging from aviation to medical diagnosis.

The reverse problem also exists. Some people distrust automated systems so thoroughly that they reject useful advice. A human–AI process can therefore fail because the human trusts the machine too much or because they trust it too little. The difficult requirement is not simply trust, but calibrated trust: knowing when this particular system is likely to be right about this particular question under these particular circumstances.

This is not something that can be solved by adding a confidence score and a brightly coloured explanation panel. A major systematic review and meta-analysis found that providing AI explanations or confidence information did not significantly improve the overall performance of human–AI combinations. The division of work between human and machine, the nature of the task, and their relative capabilities mattered more.

Disempowering humans

The next problem with a typical human-in-the-loop approach is disempowerment. The better the AI appears to work, the less cognitive work the human is likely to perform. This is known as cognitive offloading.

Cognitive offloading has been part of history since humans started to work collaboratively, build tools, and record information – in other words, forever. We use specialised skills, tools, and written knowledge to reduce the amount we must hold and process in our own heads. Offloading work frees you up to concentrate.

However, sustained reliance on AI may weaken the very skills needed to recognise when it has failed. The 2026 International AI Safety Report describes emerging evidence connecting extensive AI reliance with reduced cognitive engagement, weaker critical-thinking behaviours and professional skill decay. It cites one observational study in which clinicians’ unaided tumour-detection performance fell after working with AI support, while also cautioning that the wider evidence remains at an early stage.

The human in the loop may therefore be asked to remain alert while observing a system that is correct most of the time, to preserve expertise that the system gives them fewer opportunities to exercise, and to take control instantly when something unusual happens.

Pilots, safety engineers, and control-room operators will recognise this problem. Passive supervision is not the same activity as active performance. Situational awareness, memory of the underlying evidence, and readiness to act typically decline when your normal role is reduced to watching.

The effort nobody counts

Many years ago, I was asked to help the engineering team building an innovative air traffic control system fix what appeared to be intractable bugs. As part of that work, I did some research into the theory of debugging, which was fascinating. If you ask an engineer how much time they spend fixing problems, most will say up to 25%. If you measure it, the true figure is 84%.

Similarly, the business cases put together for AI adoption commonly calculate how much time the system saves. They do not usually calculate the time required to check its work properly. This in itself could account for why only 10% of organisations currently realise significant ROI from agentic AI.

As discussed above, a conscientious reviewer needs to do multiple pieces of highly complex analysis. Is the answer factually correct? Are the sources reliable? Are they current? Has important evidence been omitted? Do different sources contradict one another? Has correlation been mistaken for causation? Does the conclusion actually follow from the evidence? Has the system applied the right policy, rule or professional practice? Is the answer appropriate to the specific circumstances? Are uncertainty and dissenting evidence visible?

Verification is a major analytical task that may be as demanding as completing the original work – or more. Reviewing a polished AI-generated answer can be harder than producing an answer from scratch, because of the time required to discover the assumptions, sources and intermediate steps that produced it.

This can require an entire team. NIST argues that human roles and responsibilities must be clearly defined and differentiated. Organisations must decide who is using the system, who is overseeing it, who has authority to challenge it, what information each person receives, and who remains accountable for the resulting decision. Merely inserting a human somewhere between input and output does not answer any of these questions.

As well as carefully allocated human resources (and plenty of time), effective oversight also requires domain knowledge, AI literacy, suitable interfaces, access to evidence, and adequate training. The European Data Protection Supervisor warns that assigning someone to verify an automated output is insufficient unless the conditions for meaningful intervention have been established. Operators must be empowered to act substantively, rather than serving as a ceremonial stamp of approval.

The human must also have organisational permission to disagree. Suppose you are a reviewer who is tasked with overriding the AI if necessary, but you well know that doing so will delay a transaction, reduce a performance metric, or provoke an irritated manager. The faster the automated system operates, the more you become the bottleneck. The pressure is on you to reduce review time, increase caseloads, approve by exception, or allow the system’s recommendations to pass – especially given that AI errors then generate additional effort to justify rejection, review the justification, escalate concerns, correct errors, and audit approved corrections.

This is not well-governed, confident oversight. Accountability without information, time, or authority is just a convenient way to place blame.

The efficiency paradox

Putting this together, typical human-in-the-loop AI is fatally flawed. If humans genuinely check everything the AI has done, the anticipated efficiency disappears. So why bother using agentic AI at all?

The major systematic review and meta-analysis mentioned above examined 106 experiments and 370 reported effects comparing humans alone, AI alone and human–AI combinations. On average, the combined arrangements performed better than humans working alone, but worse than whichever of the human or AI performed best independently. The losses were particularly apparent in decision tasks, while creative tasks offered greater potential for productive collaboration.

This finding demands thought. Suppose the AI performed better on a certain task. Should we remove humans from that work? Allocation legal rights, ethical judgement, accountability, social legitimacy, and consequential decisions generally to machines has wider implications than a linear performance score.

What the finding tells us is that human involvement does not automatically bring human–AI synergy. In fact, it’s the reverse. Most of the systems in the meta-analysis used the typical human-in-the-loop pattern of giving an AI recommendation to a human, who then made the final decision. It was that familiar arrangement that failed to outperform the stronger party working alone. In fact, the researchers suggest that better results may depend on a more deliberate division of labour, assigning different subtasks to humans and AI according to their respective capabilities.

This gives us a new basis for putting humans in the loop – one that is more likely to bring true improvements in efficiency and effectiveness. Instead of trying to determine which point in the process a human should be brought into the loop (whatever “the loop” is), we should ask more sophisticated questions. For example,

• Where in this process is human judgement required?
• What skills, qualifications, and experience are required for those judgements?
• What evidence is required at each point for a human to participate effectively?
• Under what conditions should the process stop, change track, or escalate?
• How should escalation be implemented?
• And so on.

This looks a lot more like a traditional workflow than an agentic AI solution whose way of working is constructed on the fly.

Humans need oversight of the process (not just its outputs)

Going further, because an agentic AI solution constructs its way of working on the fly, this way of working may be largely opaque to the humans involved. By the time that outputs are presented to them for review, the AI system has already made important choices: which sources were included, how the problem was framed, what objective was optimised, which people were consulted, what counted as success, what uncertainty was hidden, and which possible harms were considered acceptable. These upstream design decisions are beyond the control, and even visibility, of a human reviewer, but they materially affect the quality of outputs.

Humans in the loop need meaningful oversight throughout the lifecycle: when the system’s purpose is defined, when knowledge is selected, when workflows are designed, when outputs are reviewed, and when evidence from actual use leads to revision. The European Data Protection Supervisor goes even further, arguing that oversight must include feedback from affected people and should become a continuing learning process rather than a symbolic intervention after the event.

Making the reasoning process itself transparent is also vital to allow recourse. A person affected by an AI-supported decision needs some practical way to question the information, challenge the reasoning, and obtain a reconsideration. Traceability, auditability and accessible redress are established components of the European Commission’s conception of trustworthy AI.

Making the loop fit for humans

Dedoctive was designed to address this problem. Its architecture gives the user confidence in outputs and enables agentic AI to be used within structured workflows that ensure practical human-in-the-loop guardrails.

Forensic analysis instead of plausible prose

Dedoctive grounds analysis in curated, validated knowledge rather than permitting a model to range indiscriminately across everything it encountered during training or can retrieve online. Specialist workflows created by human experts determine how difficult problems are analysed. Where the required evidence is unavailable, the system is designed to say so rather than fill the gap with plausible language.

Due to this careful architecture, the University of York (on behalf of the UK Ministry of Defence) were able to confirm that Dedoctive complies with all 6 GOV.UK Data Quality Dimensions: Dedoctive outputs are correct, complete, consistent, current, concise, and checkable.

Taken together, these features mitigate several different risks to the validity of an AI-generated answer. An answer may contain no obvious falsehood and still be misleading because it is incomplete. It may quote valid evidence that is no longer current. It may contain individually correct statements that contradict one another. It may bury the decisive point under so much irrelevant material that the reviewer cannot find it. Or it may be impossible to check because the connection between the claim and its source has disappeared.

Dedoctive’s compliance with all these principles ensures that humans are presented with better answers that have been generated in a transparent manner. Human reviewers can use hyperlinks (and a mind map view) to see the detailed provenance of every statement in output. They can also understand (and influence) how the output as a whole was put together. These aspects remove much of their cognitive burden. I’ll discuss each aspect separately.

Provenance for every statement

Dedoctive links the individual parts of an output to the evidence from which they were derived. Its granular provenance can lead directly to the relevant paragraph, table cell or even background text within an image.

This makes a huge difference to human reviewers. Without full provenance, a human reviewer may have to search documents, guess which passages were used, reconstruct the AI’s interpretation and determine whether evidence has been quoted accurately and in context. With full provenance, the person can move directly from claim to source.

That does not make the claim true merely because a hyperlink exists. The source may itself be incomplete, outdated or contestable. But provenance makes verification possible. It turns an unsupported assertion into an inspectable proposition.

This is particularly important in safety engineering, crisis management, planning, healthcare, skills and other domains in which the answer is rarely a single fact. The important output is usually an argument: a connected set of claims, evidence, assumptions, exceptions and recommended actions.

The human should be reviewing that argument, not conducting an archaeological excavation to discover where it came from.

Understanding how the output was put together

Dedoctive allows the human in the loop to understand, and have agency in designing, the loop itself.

It does this by combining agentic AI with workflows represented in the standard format Business Process Model and Notation (BPMN). AI is harnessed to search, extract, compare, classify, analyse, and draft. The workflow defines when those activities occur, what information they may use, what happens next, and where human judgement, approval or escalation is required.

This matters because an agentic AI system does more than generate text. It may select tools, retrieve information, carry out a sequence of activities and adapt its next step according to intermediate results. Given operational freedom, the way of working may vary unpredictably between usages, there is a high risk that the process may not always (or ever) use best practices, and it becomes deeply challenging for a human to assess how results were arrived at.

A Dedoctive workflow makes explicit:

• Which tasks the AI may perform agentically;
• Which sources and rules it may use;
• Which outputs require human review;
• What evidence the reviewer will receive;
• Who has authority to approve, reject or modify a result;
• Which conditions trigger escalation;
• How exceptions are handled;
• What is recorded for later audit; and
• How experience feeds back into improvement.

A Dedoctive workflow does not insert humans into every step (which would destroy much of the value of AI automation). Rather, it places human attention at the points where human abilities genuinely matter: determining goals, applying values, recognising unusual context, resolving ambiguity, considering consequences, exercising professional judgement and accepting responsibility.

Effective, efficient human guardrails

In the end, just saying a human is “in the loop” is too vague to deliver the efficiency and effectiveness improvements required from agentic AI. A human can be in a loop while poorly informed, overworked, distracted, deskilled, unable to challenge the system and unaware that responsibility has quietly migrated onto their shoulders.

An auditable AI system that produces defensible decisions by putting a human in the loop must provide trustworthy evidence, expose the reasoning that connects evidence to conclusions, allocate work intelligently, define authority, allow intervention and preserve an audit trail of what happened.

A system of this kind is not a technology deliverable – it is a socio-technical process. To evaluate it, it is not enough to ask how accurate it is on average and how often humans need to patch poor AI responses. Rather, we must ask whether the process as a whole helps real people make better decisions, with less avoidable effort, without concealing risk or weakening their ability to think for themselves.

We must also ask whether the system allows the humans in the loop to help provide effective guardrails against undesirable outcomes. Does the system empower the humans involved to identify, understand, and prevent potential failures? This is essential if we want to use AI without fear of causing catastrophic damage.

Dedoctive enables you to create, quickly and easily, AI-enabled systems that do more than drop a human into an opaque “loop”. It lets you build systems that humans are happy to be part of, that organisations can trust, and that deliver the performance improvements promised by AI.