AI can gather evidence, draft options and argue the other side. Seven product decisions still need a named human owner. Here is the list, the approval rule and a checklist.
What should PMs not delegate to AI? Seven decisions: which problem gets attention, what customers actually mean, which trade-off the roadmap accepts, what the team will not build, how stakeholder conflict ends, whether a risk is acceptable, and whether a bet should continue. AI can gather evidence, draft options and argue the other side. A named person still makes the call and owns the consequence.
Key takeaways
- Developers distrust AI output more than they trust it: 46% versus 33%, with only 3% highly trusting it (Stack Overflow, 2025, 33,000+ respondents).
- Anthropic's own engineers use AI in 59% of their work, yet most say they can fully delegate only 0–20% of it (Anthropic, December 2025).
- The EU AI Act requires people overseeing high-risk AI to be able to override or reverse its output and to guard against automation bias (Article 14).
- OpenAI's agent guide names two triggers for handing control back to a human: repeated failure, and actions that are sensitive, irreversible or high-stakes.
- The working rule: delegate the preparation, keep the decision. The higher the risk and the harder the undo, the closer human approval sits to the action.
Why is an AI recommendation not the same as a decision?
AI produces a confident answer from whatever context it receives. Fluent writing hides gaps: a recommendation can read as decisive while the evidence is incomplete, the customer request is unrepresentative, the trade-off is unstated and nobody has been named as owner.
The people who use AI most are cautious about trusting it. In Stack Overflow's 2025 Developer Survey, 84% of respondents use or plan to use AI tools, but 46% distrust the accuracy of the output and 33% trust it (Stack Overflow, 2025). The most common frustration, named by 66%, is output that is close to right but not quite there. Google's 2025 DORA research found the same gap: 90% of nearly 5,000 technology professionals use AI at work, yet 30% trust its output only a little or not at all (Google, 2025).
Product decisions are harder to check than code. A failing test tells you code is wrong within minutes. A wrong prioritisation call can look fine for two quarters.
What is automation bias, and why should PMs care?
Automation bias is the habit of accepting an output because a system produced it. The EU AI Act names it directly: people overseeing high-risk AI must stay aware of "the possible tendency of automatically relying or over-relying on the output" (EUR-Lex, Regulation (EU) 2024/1689, Article 14). The same article says the overseer must be able to disregard, override or reverse the output, and to stop the system safely.
Most PM tooling is not a high-risk system under the Act, but the principle transfers: a PM who cannot explain why they accepted a recommendation has not made a decision. They have forwarded one.
What should product managers never delegate to AI?
Seven product decisions should stay human-owned. AI can prepare the ground for each, but the decision rests on context, relationships or accountability that a model does not hold.
1. Which problem deserves the team's attention
AI can rank signals by frequency or similarity. Frequency is not importance. Ten requests from free users may matter less than one churn risk at your largest account, and only someone who knows the strategy, the timing and the opportunity cost can weigh that. Marty Cagan assigns value risk, whether customers will buy it or choose to use it, to the product manager (SVPG, 2017).
2. What the customer actually means
AI can tag themes across 200 transcripts in minutes. It cannot ask the follow-up question that separates a requested solution ("add an export button") from the underlying job ("my finance lead needs this number every Monday"). Keep each important claim linked to the original quote, and read the source before acting on a theme. Our guide to keeping evidence separate from interpretation in AI research synthesis shows the method.
3. Which trade-off the roadmap accepts
A prioritisation score is an input, not a verdict. Prioritisation decides who gets served, what waits, which risk is accepted and what the company chooses to learn. AI can run ten scenarios. A person has to pick one and explain the cost to the people who lose. The RICE prioritisation guide covers how to use a score without letting it decide for you.
4. What the team will deliberately not build
Saying no takes conviction and a relationship with the people who asked, and it matters more now that a prototype can take an afternoon. Pendo's analysis of 615 software subscriptions found that about 80% of features are rarely or never used (Pendo, 2019). Cheaper building makes that share easier to grow, not shrink. We cover the conversation itself in how to say no to "we can build this in one hour with AI".
5. How stakeholder conflict is resolved
AI can map competing goals and draft three compromise options. It cannot hold the relationship between product, a key customer, a sales leader and engineering when the answer disappoints one of them. Jeff Bezos's 2016 shareholder letter recommends that a leader with conviction but no consensus should disagree and commit (Amazon, 2016). That only works when a person is doing the committing.
6. Whether a risk is acceptable
Privacy, security, safety, compliance, reliability and irreversible actions need a responsible human to accept the risk. AI can describe the risk. It should never silently accept it. This applies to what AI builds as well: Veracode tested more than 100 language models and found that 45% of generated code samples introduced an OWASP Top 10 vulnerability, with security performance flat regardless of model size (Veracode, 2025).
7. Whether a product bet is still alive
AI can compare launch metrics against the hypothesis. Whether to continue, narrow, pause or retire the bet involves sunk cost, team morale, customer promises and the next opportunity. CB Insights found that 43% of 431 venture-backed startups that shut down since 2023 cited poor product-market fit (CB Insights, 2026). Ending the wrong bet early is part of the PM's job, and it needs an owner who will say so out loud.
The seven decisions at a glance
| Decision | What AI can do | What the human owns | Usually reversible? |
|---|---|---|---|
| Problem selection | Cluster and count signals | Weighing strategy, timing and opportunity cost | Partly |
| Customer meaning | Tag themes, pull quotes | Follow-up questions and the job behind the request | Yes |
| Roadmap trade-off | Score and compare scenarios | Choosing who waits and why | Partly |
| Not building | Generate alternatives | Saying no and keeping trust | Yes |
| Stakeholder conflict | Map positions, draft options | The relationship and the final call | Partly |
| Risk acceptance | Describe risks and failure modes | Accepting or rejecting the risk | Often no |
| Bet continuation | Compare outcomes with the hypothesis | Continue, narrow, pause or stop | Partly |
What can product managers safely delegate to AI?
Delegate work where a human can check the result against its source before it changes a roadmap, a customer promise or a production system. Good candidates:
- Meeting and call summaries with links to the recording or transcript
- Feedback grouping and deduplication
- First drafts of research readouts, PRDs and release notes
- Candidate interview questions and alternative hypotheses
- Draft acceptance criteria for review
- Decision-log updates and status-report preparation
- Checks for missing fields, stale evidence or inconsistent terminology
The common property is reviewability. Anthropic's study of its own staff is a useful reference point: engineers and researchers used Claude in 59% of their work, up from 28% a year earlier, and reported a 50% productivity gain, yet more than half said they could fully delegate only 0–20% of their work (Anthropic, 2025). Heavy use and light delegation sit together comfortably. The full method for sorting tasks is in what product managers should automate with AI.
How do you decide when a human must approve an AI output?
Use one rule: the higher the risk, irreversibility, ambiguity and relationship impact, the closer the human approval must sit to the action.
OpenAI's guide to building agents gives two triggers for handing control to a person: exceeding a failure threshold, such as misunderstanding a request several times, and high-risk actions that are sensitive, irreversible or high-stakes, such as cancelling orders or authorising large refunds (OpenAI, 2025). It also recommends rating every tool an agent can use as low, medium or high risk, based on read versus write access, reversibility, permissions and financial impact.
Bezos's two-way-door idea completes the picture. Many decisions are reversible and deserve a light process, and most can be made with about 70% of the information you wish you had. Save heavy review for the one-way doors. Our reversibility framework for product decisions shows how to tell them apart.
A four-factor approval score
Score each AI-prepared output from 1 (low) to 3 (high) on four factors, then add them up.
| Factor | 1 | 2 | 3 |
|---|---|---|---|
| Risk if wrong | Cosmetic | Annoys users or wastes a sprint | Money, data, safety or legal exposure |
| Reversibility | Undo in minutes | Undo in days | Cannot be undone |
| Ambiguity | Clear source and rule | Some interpretation | Contested evidence or strategy |
| Relationship impact | Internal only | One customer or team | Many customers or a public commitment |
- 4–6: automate it and spot-check weekly.
- 7–9: AI drafts, and the PM reviews before anything ships.
- 10–12: AI prepares evidence and options only. A named owner decides and records the reasoning.
The bands are a starting point, not a standard.
Worked example (illustrative)
Say a six-person team uses AI to triage 140 pieces of feedback a month. Tagging and deduplication scores 5: low risk, reversible in minutes, clear rules, internal only. It runs automatically, and the PM samples 10 items every Friday.
The same AI then recommends retiring a reporting feature used by 12% of accounts. That scores 11: users lose data, the change is hard to reverse, the usage evidence is contested, and many customers are affected. The AI assembles the evidence pack. The PM interviews three affected customers, makes the call and logs why.
Checklist: before you accept an AI recommendation
Copy this into your decision template.
- I can name the sources this recommendation rests on, and I opened at least one.
- I know which assumptions it depends on and which evidence contradicts it.
- I asked the AI for the strongest case against its own answer.
- Customer quotes are real and linked, not paraphrased into something new.
- I scored it for risk, reversibility, ambiguity and relationship impact.
- A named person owns the consequence, and that person is not "the AI".
- The decision and its reason are written down.
- There is a date to check whether the decision was right.
If the ownership line is blank, the decision is not ready. A product decision log template makes the last two boxes a two-minute job.
What are the most common mistakes?
Treating a summary as the source. A summary of 40 interviews is a claim about those interviews. Spot-check quotes before they reach a roadmap.
Delegating by vagueness. "Prioritise our roadmap" is not a task an AI can do well. It is a decision with missing inputs. Give AI narrow jobs whose output you can check.
Letting volume set the agenda. AI makes ideas, specs and prototypes cheap. Product Institute's 2026 survey of 309 product leaders found that only 17% named strategic planning and roadmapping as an area where AI had high impact, against 50% for engineering and development (Product Institute, 2026). The output grew faster than the thinking behind it.
No owner after launch. The continue-or-stop decision is the one most often skipped, because no meeting forces it. Put the review date in the calendar when you ship.
Over-correcting. Refusing to delegate anything wastes time AI can give back. Martin Fowler's note on agentic engineering strikes the balance: "Humans are still responsible for what the software does and how it works" (Fowler, 2026).
Where Specky fits
Specky handles the preparation side of these decisions. It reads customer calls, feedback, Slack, Jira and analytics into a Product Graph, so an AI-drafted PRD or recommendation arrives with inline citations to the conversations behind it. Drafts land in the PM Inbox for human review instead of going straight out, and the Decision Log records who decided what and why. The AI does the retrieval and the first draft. You keep the call.
Frequently asked questions
What should product managers never delegate to AI?
Problem selection, customer interpretation, roadmap trade-offs, the decision not to build, stakeholder conflict, risk acceptance and the call to continue or stop a bet. AI can prepare evidence and options for each, but these decisions depend on strategy, relationships and accountability that a model does not hold. Name a human owner for every one of them.
Can AI prioritise a product roadmap?
AI can score items, run scenarios and show which evidence supports each option. It should not set the final order. Prioritisation decides who waits and which risk the company accepts, and someone has to explain that to the people affected. Treat AI scores as one input next to customer evidence, strategy and engineering cost.
Should PMs trust AI summaries of customer feedback?
Trust them as a starting point, not as evidence. Summaries can merge different complaints or drop context. Keep every important claim linked to the original transcript, ticket or metric, and spot-check a sample before a summary shapes a roadmap decision.
What does human-in-the-loop mean in product management?
It means a named person reviews and accepts an AI-prepared decision or action before it creates a customer, strategic, financial or production commitment. The reviewer must be able to understand the output, reject or reverse it, and stop the process. That mirrors what Article 14 of the EU AI Act requires for high-risk AI systems.
How much of a PM's work can AI do?
Much of the preparation and little of the deciding. Anthropic's engineers used AI in 59% of their work in 2025, yet most said they could fully delegate only 0–20% of it. For PMs, research synthesis, drafting and status work move to AI, while problem choice, trade-offs and accountability stay human. Our data review asks whether AI can replace product managers.
Sources
- Stack Overflow — 2025 Developer Survey: AI
- Google — How are developers using AI? Inside our 2025 DORA report
- EUR-Lex — Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14
- SVPG — The Four Big Risks (Marty Cagan, 2017)
- Pendo — The 2019 Feature Adoption Report
- Amazon — 2016 Letter to Shareholders
- Veracode — 2025 GenAI Code Security Report
- CB Insights — The top reasons startups fail
- Anthropic — How AI is transforming work at Anthropic
- OpenAI — A practical guide to building agents
- Product Institute — State of AI in Product 2026
- Martin Fowler — Agentic Engineering
This story was originally published on 15 October 2026 at https://www.specky.space/blog/what-product-managers-should-never-delegate-to-ai.
