A practical model for deciding what AI can do, what humans must still own, and where product teams should draw the boundary.
An AI tool clusters hundreds of customer comments and names the largest theme “pricing friction”. The label sounds plausible, so the team asks it for pricing experiments, compares the options it returns and begins to discuss a roadmap change.
By the time someone approves the recommendation, the important decision may already have been made. The system has helped define the problem, selected a frame and made some alternatives less visible.
This is the practical governance question for product teams. Not simply what AI is capable of doing, but what authority it should have while doing it. Capability describes what a system can produce. Authority concerns what the organisation is prepared to let it decide, shape or set in motion.
I call the distinction bounded delegation. Give AI broad scope where exploration is cheap and reversible. Delegate analysis more substantially when its evidence, assumptions and uncertainty can be inspected. Tighten the boundary when a team is making a consequential commitment.
Authority should narrow as the stakes rise
It helps to distinguish between generation, analysis and commitment. They often blur together in a product workflow, but they carry different kinds of risk.
In generation, AI can usually range widely. It can draft interview questions, suggest alternative flows, summarise observations and produce prototype directions. These are starting points, not promises to customers or investments of scarce engineering capacity.
Analysis deserves more care. AI can compare evidence, surface patterns, identify trade-offs and test an argument, but the work needs to remain inspectable. A team should be able to see the evidence used, the assumptions made, what may have been excluded and where uncertainty remains.
Commitment is different. It is the point at which a team chooses what to build, allocates budget, changes a customer experience, makes a promise or accepts risk. AI can inform the decision; it should not quietly own it.
The practical rule is straightforward: AI authority should narrow as impact and accountability increase. A team can ask for ten onboarding alternatives with little concern. Asking a system to decide which customers lose access to a feature is another matter entirely.
A recommendation can decide the question too
The risk is not confined to a final automated action. It can arise much earlier, in the way an AI system frames the work.
Return to “pricing friction”. The same comments might indicate weak perceived value, confusing packaging, poor onboarding or a mismatch between the product and its intended segment. Once the label becomes the working frame, those explanations are easier to miss. Subsequent prompts can make the direction feel increasingly inevitable.
That is why final human sign-off is not enough. If AI framed the problem, categorised the evidence and narrowed the alternatives, the person approving its recommendation may have little meaningful judgement left to exercise. A human click does not, by itself, make a decision human-led.
Before acting on an AI-led synthesis, ask what other interpretations were plausible, what evidence would challenge the frame, what the synthesis flattened or excluded, and whether the team would have arrived at the same question without the AI-generated label. The consequential hand-off may have happened well before the recommendation appears.
This concern sits behind the anxiety explored in “When AI Stops Suggesting and Starts Acting”: delegation changes character when actions become harder to reverse and responsibility remains with people.
Four questions for setting the boundary
There is no universal line between AI and human authority. In practice, four questions are usually enough to establish a defensible one.
What happens if we are wrong?
Low-cost, reversible work can tolerate more delegation than decisions that are expensive, difficult to undo or potentially harmful. Generating alternative survey wording is easy to reverse. Choosing the customer segment that receives a year of investment is not. The consequence may be financial, but it may also concern trust, accessibility, privacy, legal exposure or reputation.
Who remains accountable?
Anyone expected to explain or defend an outcome later needs enough access to the evidence and reasoning to make a real judgement. Sign-off is a weak safeguard when the signer cannot say why the decision is sound.
A Microsoft Research study of where developers draw the line on AI autonomy found that willingness to delegate declined as accountability increased. The finding supports a simple organisational principle: responsibility should not remain human while meaningful authority quietly moves elsewhere.
Can the work be inspected?
Delegation becomes fragile when provenance, assumptions and uncertainty disappear behind a confident answer. If AI says users prefer option B, can the team trace that claim to the underlying research, see contrary evidence and understand the sample’s limits? If not, the output may still be a useful prompt for investigation, but it is a poor basis for commitment.
Is this generation, analysis or commitment?
This is often the quickest useful test. Generation usually permits the widest latitude. Analysis can support substantial delegation when it remains inspectable. Commitment warrants the narrowest boundary because it determines what the organisation will do, spend, change or promise.
Image created by the author with OpenAI image generation.
Delegation is not autonomy
The familiar copilot-versus-agent debate is too blunt for most product decisions. The important question is not whether AI waits for a prompt or acts independently. It is which decisions it is authorised to make as it works.
In Microsoft Research’s study of 22 AI systems developers want built, developers described systems capable of doing substantial work while respecting boundaries around authority, provenance, uncertainty and access. That is closer to the operating model product teams need.
A system can examine hundreds of interview notes without deciding which opportunity merits investment. It can generate interface directions without choosing which reaches customers. It can draft an experiment without deciding that the evidence justifies committing engineering time. The useful unit is not autonomy; it is bounded authority.
“Human in the loop” is not an operating model
The phrase sounds reassuring but leaves the essential questions unanswered. Which human is involved, at what point, for which decision, with what evidence and with what authority?
A designer reviewing AI-generated interface options has a different responsibility from an executive approving portfolio investment. A researcher checking an AI synthesis is doing different work from a product manager deciding whether an opportunity warrants another quarter of investment. Calling each situation “human in the loop” obscures the accountability structure rather than defining it.
Teams need deliberate hand-offs. State what AI may do, make material analysis inspectable, identify where authority narrows, and name the person or group that owns the commitment.
Judgment Points make the boundary operational
This is where the Discovery Judgment Framework, or DJF, connects to bounded delegation. DJF does not assume people should perform every discovery task manually. That would overlook much of what AI can do well.
Instead, it identifies consequential Judgment Points where a team makes an explicit decision before progressing. These points create natural boundaries for delegation.
Before a Judgment Point, AI can gather evidence, organise information, identify patterns, generate alternatives, expose assumptions, compare trade-offs and model scenarios. At the point itself, the team should be able to account for the evidence supporting the decision, the uncertainty that remains, the assumptions still in play, the alternatives considered and who owns the commitment.
Concept and image by the author, created with OpenAI image generation.
That is enough structure to preserve accountability without pulling people back into work AI can perform effectively. It also makes the boundary visible at the moment it matters, rather than relying on a vague promise of oversight.
Make the next delegation decision explicit
Bounded delegation is not an argument for using less AI. It is an argument for being more precise about authority.
Use AI to explore more options than a team could reasonably create by hand. Let it analyse evidence, test assumptions and surface patterns that deserve attention. But do not let increased capability quietly become increased authority.
For the next AI-assisted workflow, identify the first point at which the team is no longer exploring but committing. Decide, before the work begins, what evidence must be visible there and who has the authority to make the call. That is how AI can do substantial work without taking ownership of decisions the organisation must still be able to explain, defend and own.
Bounded delegation: How much authority should AI have? was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.
