I was cruising this month with my family in the Caribbean when I had some interesting conversations around the use of artificial intelligence tools at work. I heard about guardrails, responsible AI guidelines, and involving human in the loop (HIL) with people approving work, code, and build. This was wonderful and I also heard the terms human in the middle and human in the loop being used almost interchangeably. In my mind, the use of humans is for governance accountability and not just monotonous approvals. As I believe in the Accountable AI, I thought how to make this more practicable. The result is the LEAD reference framework. Let me elaborate.
It is important to realize that the neither the AI as a system or the AI agents as parts of the larger system hold accountability. They may provide evidence, build statistics, predict/forecast, monitor itself, etc. But, none of these things provide accountability to the end users who are impacted by the AI model. That accountability sits squarely with the humans. Here is where I think we need to call out the distinction between human in the loop from human in the middle.
Human in the Loop vs Human in the Middle
Human-in-the-Loop typically means inserting a person somewhere in the AI decision process—reviewing, approving, or overriding an AI-generated outcome. Although this is certainly useful, it can also become a checkbox. In that case, it is not human judgement but a formality involving human approval. Furthermore, as soon as we conceive it as a checkbox, it opens up the dangerous exposure to automating it removing that human connection. So, how is the human in the middle different?
If Human-in-the-Loop (HIL) asks: “Where do we insert a human?” then Human-in-the-Middle (HIM) asks: “Where must humans exercise judgment?” That judgment begins before the AI produces an answer and continues after it does. It includes what we ask AI to do, what data we provide, how we test it, what authority we give it, and what we do when its behavior produces unacceptable consequences.
This is where my LEAD framework becomes practical.
LEAD: Making Human Judgment Explicit
The LEAD framework stands for:
- L — Look beyond the obvious output.
- E — Examine the evidence and its consequences.
- A — Act when the evidence reveals unacceptable risk.
- D — Decide and Demonstrate who owns the outcome and why the decision was made.
The technical professional and the leader will approach LEAD differently. Both must collaborate together to ensure they demonstrate collective leadership in these questions answered.
- The technical professional asks: “Does the system work as intended?”
- The leader must also ask: “Should the system be doing this in the first place, and are we willing to stand behind its consequence"
Prompt engineering is a good example of why the AI world challenges some of our traditional Agile assumptions. The second Agile Manifesto principle reads, “Working software over comprehensive documentation.” The intent of this principle was never to eliminate thinking but to reduce unnecessary bureaucracy. With vibe coding and model driven development and powerful AI tools like Kiro, Claude Code, Codex (to cite a few), AI can now generate working software, requirements, test cases, analyses, documentation, and risks at extraordinary speed. So, the pattern is shifting to comprehensive documentation as software development itself is no longer a niche skill (I am sure it may evolve later).
Consequently, those that develop software need to master the art of Prompt Engineering discipline that requires people to think through the specific use cases, deliberate design around the constraints, assumptions, risks, and dependencies (CARD that I keep referring) further evaluating against bias, ambiguity, guardrails, and so on (Ethical considerations). It is important to note that the prompt itself may be small and succinct but the thinking behind the prompt certainly is not. The new pattern therefore is "Ethically Compliance over Faster Delivery." This thinking makes humans accountable for the outcome rather than an Approval Master in the tool chain.
Let us demonstrate this with an example of using AI to to generate test cases. Given below is the technical responsibility and leadership responsibility across the LEAD dimensions. I am avoiding the specific roles like designer, developer, tester, manager, scrum master, product owner, or project manager, or senior leadership teams in both these areas as the job titles differ across the industry globally.
| LEAD Area | Technical Responsibility | Leadership Responsibility |
|---|---|---|
| Look | What scenarios did AI generate? What scenarios did AI miss? | What risks are we not testing? Why are we not testing for them? |
| Examine | Are the tests reliable and repeatable? Are the tests automation-ready? Are the test sufficiently broad? | Are we optimizing the number of tests rather than meaningful risk coverage? |
| Act | How we are adding missing tests? How are we improving tests/prompts? How are we analyzing results (e.g.: FP, FN) | How are we verifying additional human judgement (governance)? |
| Decide | What evidence increases our confidence? What risks are we absorbing? What risks are we passing it on to others? | Who accepts the residual and secondary risks? How are we communicating them to users with autonomy options? |
Throughout the AI assisted product delivery value chain, whether it is software or not, the "risk management" is the language through which technical and leadership teams hold each others accountable. So, technical teams govern how the capability works. Leaders govern why it should be used, where it can be used, and what happens when it fails.
Although a model can be explainable, a prompt can be documented, a test suite can be automated, and a dashboard can show every metric, none of these holds the accountability cards. As we shift from HIL to HIM involving human judgement throughout the value delivery, we demonstrate that "The goal is not slow AI down. The goal is to avoid accelerating irresponsibility." To me, this is the foundation of "Accountable AI" mindset.
I think I landed on something here with the LEAD framework. What do you think?
