AI did not remove product work; it moved the bottleneck to judgment, review and ownership. Here is the data, a three-lane PM–engineering contract and a review-capacity check.
AI did not remove product work; it moved it. Producing drafts, prototypes and code got cheap, while deciding what deserves attention, reviewing what was produced and owning what reaches users got more expensive. That is the new AI product workload. Teams with high AI adoption merge 98% more pull requests, yet review time rises 91%. The fix is a clear contract for who prepares, who decides and who owns production.
Key takeaways
- Across 10,000+ developers, high-AI-adoption teams completed 21% more tasks and merged 98% more PRs, but PR review time rose 91% and PR size 154%, with no measurable company-level improvement (Faros AI, 2025).
- DORA's 2025 research found AI now raises delivery throughput but still lowers stability, because more change volume overwhelms weak controls.
- 27% of AI-assisted work at Anthropic consisted of tasks that would not otherwise have been done. Cheaper creation means more to evaluate.
- PM's share of tech hiring grew from 16% in 2021 to 21% in 2026, while US software-engineer hiring fell to under half its 2022 peak.
- Split the work into three lanes: automate PM preparation, keep product judgment human, and share production responsibility with engineering.
What new workload did AI create for product teams?
The new workload is evaluation. When creation is cheap, a team can produce more candidate work than it can judge. The clearest measurement comes from Faros AI's 2025 telemetry study of more than 10,000 developers across 1,255 teams (Faros AI, 2025). On teams with high AI adoption:
- Developers completed 21% more tasks and merged 98% more pull requests.
- Average PR size grew 154%.
- PR review time grew 91%, which Faros identifies as the main bottleneck.
- Bugs per developer rose 9%.
- Company-level delivery showed no significant improvement.
Google's DORA programme found the same pattern at system level. Its 2025 report, based on nearly 5,000 technology professionals, found AI adoption is now positively related to delivery throughput, reversing 2024's result, but still negatively related to stability. The authors attribute this to higher change volume overwhelming weak controls such as automated testing and feedback loops (Google Cloud, 2025).
It also creates entirely new work. Anthropic's study of its own staff found that 27% of AI-assisted work consisted of tasks that would not have been done without AI (Anthropic, 2025). Some of that is valuable. All of it needs someone to decide whether it should exist.
Where the work moved
| Activity | Before AI coding tools | With AI coding tools |
|---|---|---|
| Writing a first draft (spec, code, prototype) | Slow, scarce | Fast, abundant |
| Choosing which draft to pursue | Implicit, because few existed | Explicit, frequent decision |
| Reviewing changes | Proportional to team size | Grows faster than the team |
| Keeping production stable | Familiar change sizes | Larger, more frequent changes |
| Connecting work to customer evidence | Often skipped | Skipped at higher cost |
| Learning from what shipped | Monthly at best | Needed weekly to keep up |
Who absorbs the new product workload?
Each role picks up a different share.
| Role | New work AI created | What to stop doing |
|---|---|---|
| Product manager | Choosing between many drafts and prototypes; writing briefs agents can follow; checking that what shipped matches the decision | Hand-writing status reports and first-draft specs |
| Engineer | Reviewing larger and more frequent changes; guarding architecture, security and reliability; maintaining agent context | Typing boilerplate |
| Engineering manager | Sizing review capacity; keeping batches small; setting AI-use policy | Counting lines, commits or accepted suggestions |
| Founder or head of product | Saying no more often, because yes is now cheap | Treating demo speed as evidence |
The hiring data reflects this shift. TrueUp and Live Data Technologies data analysed by Lenny Rachitsky shows product management was the only core role whose share of hiring grew, from 16% in 2021 to 21% in late 2026, while US companies hired about 113,000 software engineers in the year to April 2026, against 239,000 at the 2022 peak (Lenny's Newsletter, 2026). Fewer engineers per PM means each PM is closer to more output. Our piece on the PM-to-engineer ratio in the AI era looks at what that does to the job.
What is the three-lane PM–engineering contract?
A contract makes the split explicit before work starts, so nobody discovers the gap in production.
Lane 1: automate the mechanical PM work
Use AI where the input exists, the transformation repeats and a person can inspect the output. Good candidates:
- Grouping and deduplicating feedback
- Extracting themes from calls and support conversations
- A first research summary with source links
- A draft brief from approved evidence
- Candidate tickets from a reviewed brief
- Release notes, stakeholder updates and decision-log entries
- Checking that an artifact contains its evidence and acceptance criteria
The standard is not "the AI did it". The standard is that the result is easier to review than doing the work from scratch. The full task list is in what product managers should automate with AI.
Lane 2: keep product judgment human
The PM stays accountable for which problem deserves attention, whose problem it is, what evidence is strong enough, what trade-off the team is making, what the hypothesis predicts, what will count as learning and what should not be built. AI can challenge a framing and find missing evidence. It should not quietly turn a plausible pattern into a roadmap commitment. We list these in what product managers should never delegate to AI.
Lane 3: share the production contract with engineering
PM and engineering agree together on:
- Architecture and system boundaries
- Security, privacy and data retention
- Reliability, performance and operating cost
- Evaluation and fallback behaviour for AI features
- Permissions and irreversible actions
- Rollout, observability and rollback
The PM brings the user, the problem, the evidence and the intended outcome. Engineering brings system constraints, failure modes and the cost of keeping the promise true in production. Martin Fowler's 2026 description of agentic engineering puts it plainly: humans remain responsible for what the software does, and understanding the problem domain matters more, not less (Fowler, 2026).
| Lane | PM owns | Engineering owns | Shared artefact |
|---|---|---|---|
| 1. Automate preparation | Rules, review of exceptions | Tooling and data access | Source-linked drafts |
| 2. Product judgment | Problem, evidence, trade-off, outcome | Feasibility input | Decision record |
| 3. Production contract | User impact, acceptance criteria | Architecture, security, reliability | Brief with guardrails and rollback plan |
Should product managers become builders?
Yes, if "PM builder" means a PM who builds enough to improve a product decision. No, if it means a PM owning production systems without the engineering skills, review and accountability to run them safely.
A PM builder can make an interaction tangible, expose ambiguity in a requirement, test a workflow with users, explore a question with data, create a disposable prototype and automate their own coordination work. A PM builder should not use AI to bypass architecture review, ship sensitive changes without security review, or turn an experiment into a production promise by accident.
One working definition: a PM builder builds enough to improve the decision, not to avoid making it. Our longer piece on whether product managers should become builders covers the skills involved.
How should PMs manage engineers and AI builders now?
The manager's job is not to maximise generated output. It is to make the system trustworthy enough that generated output can move safely. Eight operating rules:
- Prefer small batches. Smaller changes are easier to review, test and roll back. DORA lists working in small batches as one of seven capabilities that amplify AI's benefits (Google Cloud, 2025).
- Keep acceptance criteria close to the evidence. Engineers should see why the work exists, not only what the ticket says.
- Require agents to show their work. Files changed, tests run, assumptions made and open questions. Then review against the brief with a PM acceptance checklist for AI-built changes.
- Treat generated code as a change proposal, not an authority.
- Review behaviour and operations, not only the diff. Ask what the user experiences and what the team must support.
- Measure outcomes. Tokens, lines of code and suggestion acceptance are activity signals. See how to manage engineers using AI coding tools.
- Improve the context available to tools. DORA's model names AI-accessible internal data and healthy data ecosystems as capabilities in their own right. Our guide to giving AI coding agents product context shows what to include.
- Publish an AI-use policy. DORA's first capability is a clear and communicated AI stance: which tools are allowed, what data they may see and which actions need approval.
Operating model matters more than tool choice. Product Institute's 2026 survey of 309 product leaders found 47% of organisations with a mature operating model say AI is strengthening how they work, against 29% of those with an immature one. Among organisations with a weak model, 42% say AI is exposing weaknesses or making things worse (Product Institute, 2026).
How do you check whether review capacity can keep up?
Measure the queue, not the output. A simple weekly check:
- Count changes opened this week and changes reviewed.
- Note the median hours from opened to first review.
- Note the median change size in lines or files.
- If opened exceeds reviewed for two weeks running, or median review time doubles, cap work in progress before adding more agent tasks.
Worked example (illustrative)
Say a team of five engineers and one PM merged 20 changes a week before AI coding tools, at a median of 150 lines each, with first review inside four hours. Six weeks after adopting agents, the team opens 40 changes a week at a median of 380 lines. Reviews now take 11 hours to start, and 12 changes are waiting at the end of each week.
The team sets a limit of three open agent-generated changes per engineer, asks agents to split any change over 200 lines, and has the PM pre-check every change against its brief's acceptance criteria before engineering review. Within three weeks the queue falls to three changes and first review returns to five hours. Output settles at about 30 merged changes a week: fewer than the peak, more than before, and reviewable. Your numbers will differ; the method is the point.
Template: weekly AI workload review
Copy this into your team's weekly meeting notes.
Changes opened / reviewed / merged this week: Median time to first review: Median change size: Agent-generated changes still open: Prototypes created / promoted to roadmap / retired: Work generated but not linked to a decision: Where did AI remove friction? Where did it add review or operational load? What do we stop, shrink or roll back? The line about unlinked work is the most useful. Anything generated without a decision behind it is the new workload in its purest form.
What are the common mistakes?
Celebrating output. A 98% rise in merged PRs means little if review time and bugs rise too. Watch the system, not the individual.
Leaving review to whoever is free. Review is now the scarcest capacity on the team. Plan it like any other capacity.
Making the PM a part-time engineer. Building to learn is good. Shipping to production without engineering review moves risk, it does not remove it.
Ignoring the developer view. Atlassian's 2025 survey of 3,500 developers and managers found developers spend 16% of their time coding and that 63% say leaders do not understand their pain points (Atlassian, 2025). Ask engineers where the new load sits before deciding.
Planning headcount without agents. Microsoft's 2025 Work Trend Index proposes a human-agent ratio as a planning metric and found 36% of leaders expect managing agents to be part of their job within five years (Microsoft, 2025). Somebody has to direct and review those agents; budget their time.
Where Specky fits
Specky is built around this contract. It reads customer calls, feedback, Slack, Jira and analytics into a Product Graph, so a brief or ticket carries inline citations to the evidence behind it. Its MCP server gives coding agents that same product context, so engineering starts with the reason, not just the request. After release, Specky checks whether the bet worked. The PM keeps the judgment; engineering keeps production; nobody loses the thread between them.
Frequently asked questions
Should PMs learn to build with AI?
Yes, when building helps them test a product question, clarify a requirement or remove repeated mechanical work. Prototypes are a fast way to show a user or engineer what you mean. PMs should still involve engineering for production architecture, security, reliability and operations, because generated code that reaches users becomes engineering's responsibility.
How should PMs manage engineers who use AI coding tools?
Set a clear decision contract, keep work in small batches, ask agents to document assumptions and tests, define data and permission boundaries, and measure product and delivery outcomes instead of generated volume. Watch review time and change size every week; they are the first signs that output is outrunning the team's capacity to check it.
Why did AI increase our workload instead of reducing it?
Because creation got cheaper faster than evaluation did. Faros AI found high-AI teams merged 98% more PRs while review time rose 91%. Anthropic found 27% of AI-assisted work would not otherwise have been done. More drafts, prototypes and changes all need someone to decide whether they should exist and whether they work.
What is the right role for a PM builder?
A decision accelerator. A PM builder creates enough to learn and align, through prototypes, data exploration and automated coordination, then works with engineering to turn a validated direction into safe, maintainable software. The test is whether building improved a decision, not whether the PM shipped code.
Will AI replace product managers?
The data says no. PM's share of tech hiring has grown since 2021, while AI absorbs tasks such as synthesis, documentation and status reporting. Choosing problems, making trade-offs and owning outcomes still need a person. See our review of the 2026 data on whether AI can replace product managers.
Sources
- Faros AI — The AI Productivity Paradox Report 2025
- Google Cloud — Announcing the 2025 DORA report
- Google Cloud — Introducing DORA's inaugural AI Capabilities Model
- Anthropic — How AI is transforming work at Anthropic
- Atlassian — State of Developer Experience Report 2025
- Lenny's Newsletter — State of the tech job market: late 2026
- Microsoft — 2025 Work Trend Index: The year the Frontier Firm is born
- 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/ai-created-new-product-workload-pm-engineers-builders.

