Product Management Role Splits as Builder PMs Ship AI Code

The generalist PM is losing ground

Product management coverage in 2026 keeps coming back to one point. The single role of "product manager" is splitting into different jobs, and the generalist in the middle is getting squeezed. A widely shared August trend piece from Userpilot puts it most bluntly. It says the split has been visible since mid-2025. One side is the AI-native "builder-PM" who ships prototypes personally. The other is the high-EQ "integrator-PM" who handles cross-functional alignment in messy B2B organizations.1 The author argues that the PMs under the most pressure are those who do neither job especially well.1

That finding has a security consequence the trend pieces mostly skip. If PMs now write and ship code, much of it generated by AI, then securing that code becomes part of the product job. This article covers both: the evidence that the role is splitting, and the open source security tooling that teams are adopting as more people write code.

Six trends, one direction

Userpilot names six forces:

• The builder/integrator split.
• Specialization. Product coach Pawel Huryn puts 2026 US pay at about $245,000 for AI-focused PMs, against $123,000 for traditional PMs.1
• A move from "operator" to "governor". AI takes over the assembly work of digging through dashboards, so the PM's main job becomes judgment.1
• Go-to-market ownership.
• Hiring that has recovered but carries weaker signals. Openings are up, yet the job applications they attract, many now generated with AI, reveal less about who can actually do the work.1
• Capital costs. Expensive capital shifts strategic authority away from PMs and toward leadership.1

The split shows up across many outlets, though they draw the dividing line in different places. The Institute of Product Management frames it as strategic PMs versus execution PMs. In its account, strategic roles gain scope and pay while execution roles, built around specs, sprint ceremonies and coordination, get squeezed by automation.3 Product Management Society founder Gabriela Naumnik describes four tracks: AI product manager, AI-powered PM, strategic PM, and AI-native product ops.4 A September analysis says one pole is deeply technical (evals, retrieval design, prototyping, cost engineering) and the other deeply strategic (pricing, positioning, executive alignment). It also says the ticket-writing middle is being compressed hardest.20 Writers disagree on how many tracks there are, but they all say the generalist middle is shrinking.

LinkedIn's restructuring is the concrete example most writers cite. Chief product officer Tomer Cohen said on Lenny's Podcast that the company replaced its Associate PM program with a "Product Builder" program. That program trains people across product, design and engineering.59 Product School says roles overlapping like this is now how teams normally work, not a malfunction.2

Where the numbers disagree

The coverage differs most on headcount and ratios, and the differences matter.

PM-to-engineer ratios. Userpilot argues the old ratio of one PM to eight engineers is breaking down because the PM job now includes shipping code.1 Product School cites Andrew Ng saying that as engineering output speeds up, some teams are moving from roughly one PM per four engineers toward two PMs per engineer.2 Naumnik reads it the other way. She notes Linear ran with just two PMs in 2026 and treats that as evidence that AI-first teams need fewer PMs per engineer.4 Ant Murphy quotes a leader who expects a 1:1 ratio because "we actually need more PMs."5

These claims point in opposite directions. My reading is that both describe real effects in different places. Where AI speeds up engineering, product judgment becomes the bottleneck and the ratio moves toward more PMs. Where PMs themselves become builders, the separate PM headcount shrinks. Neither is a universal law.

Job counts. Userpilot cites Murphy's figure of about 42,000 open PM roles on LinkedIn in early 2026, double the year before. It adds LinkedIn data from May showing the market up 14% year over year and Associate PM roles up 33%.1 Lenny Rachitsky's early-2026 report, built on TrueUp data covering more than 9,000 tech companies, found over 7,300 open PM roles. That was the highest level in over three years and about 75% above the 2023 low.8 The counts are far apart because they come from different datasets, but both say hiring is up.

The Institute of Product Management adds a less comfortable figure. It reports that senior PM hiring grew 34% in 2025 while junior and mid-level hiring fell 12%.3 That conflicts with the reported 33% growth in Associate PM roles.1 It is also striking given that LinkedIn just retired its APM program.5 The safest conclusion is that entry-level hiring is being redesigned rather than simply growing or shrinking.

AI PM pay. The AI PM premium is real but covers a small group. AI PM roles make up roughly 8–10% of open PM positions, with nearly half of them in the US.5 Estimates of the premium vary. Huryn's $245,000 figure1 is well above another practitioner's range of $133,000 to $200,000-plus.11 One US jobs guide adds a useful distinction: 61% of PM postings mention AI, but only 23% of senior postings require experience actually shipping AI products.8

The builder-PM writes code, and someone has to secure it

The builder track rests on prototyping. Aakash Gupta, writing for Atlassian, calls AI prototyping the biggest change to the PM job in years. He advises PMs to deliver working prototypes alongside their strategy documents.14 FourWeekMBA reports that AI-native companies including Anthropic, Cursor, Linear and Vercel hire builder-PMs. In their interviews, candidates build a prototype instead of working through a case study.15 Product School describes PMs using Claude Code and vibe coding (building software by describing it to an AI) to close the gap between product intent and the actual code.18

Some writers raise concerns about this. One January essay warns that vibe coding without architectural oversight is already producing codebases that are hard to maintain.7 Naumnik notes that AI analysis can still be wrong, with roughly a 20% false-positive rate. She also says high-risk areas such as pricing, legal disclosures and account deletion still need strict human approval.4 She adds that PMs now carry more risk around data governance, privacy, bias and trust, and that assuming legal or engineering will handle those issues is a risky bet.4

The security data backs up her concern. Cycode's 2026 State of Product Security report, as cited in a startup-focused guide, found that every surveyed organization had AI-generated code in its codebase. It also found that 81% lacked visibility into AI usage across the software development lifecycle.24 A developer-community roundup argues that AI coding assistants ship code faster than review can keep up. It says these models readily reproduce SQL concatenation, insecure deserialization and hardcoded credentials, mistakes common in the code they learned from.28

The open source security stack product teams are adopting

The free tooling here is mature enough to cover a lot. The same startup guide argues a two-person team can now build a serious AI security stack from free tools alone. It names Garak for red teaming LLMs (probing them with attacks), Promptfoo for testing inside CI/CD pipelines, LLM Guard for screening inputs and outputs at runtime, and NeMo Guardrails for enforcing conversation policies.24

A separate August list adds tools for agent-era risks:

• Obot: an MIT-licensed gateway for MCP, the protocol that connects AI agents to external tools and data, that controls and audits what agents can access.
• PyRIT: Microsoft's MIT-licensed framework for automated red teaming.
• Guardrails AI: a framework for validating model inputs and outputs at runtime.

The authors present these as layers of one workflow rather than competing products.22

For conventional code and the software supply chain, the same few names keep appearing:

• Trivy: scans container images, infrastructure-as-code and repositories in one tool.28 With AI/ML plugins it can also scan Hugging Face model files and ML dependencies.24
• Semgrep: its open rulesets flag insecure patterns in AI-generated code before they reach a pull request.28 It is available in open source and commercial editions.27
• Gitleaks and Trufflehog: catch secrets before or after they enter git history.28

Even vendor-written comparisons, such as Checkmarx's DevSecOps roundup, credit Trivy as a fast, lightweight open source option. They also point out it only scans statically and lacks runtime enforcement.25

Many of these rankings come from vendors that put their own products at the top. Xygeni ranks itself first on its list of open source security tools.23 Arnica names itself the best supply-chain tool on its own list.21 Read them for what each tool covers, not for the order.

The commercial direction is clear. Endor Labs is described as keeping code generation separate from security verification, with an independent enforcement layer that sits across AI coding agents.27 Semgrep's Guardian product scans AI-generated code as it is written.27

The takeaway

Read together, the coverage leads me to a firm view: the builder-PM and AI governance are one trend, not two. The PM who ships a prototype on Monday is the same PM who, under Naumnik's model of an agentic team, ends up as a "Responsible AI Owner." That person manages how models fail, sets acceptable error rates and maintains guardrails.4 One analyst writes that PMs now own eval design for AI features and carry regulatory classification work that used to sit with legal.20

Security tooling has caught up with that shift. Promptfoo in the CI pipeline, Semgrep on every pull request and Gitleaks in pre-commit hooks are mostly free and easy to set up.2428 So a builder-PM who skips them is not simply moving fast. They are shipping code they cannot vouch for.

The role is not disappearing. Every source reviewed here says the low-leverage parts of PM work are going away while judgment matters more.34 What changes in 2026 is that judgment now has to cover whether AI-generated code is safe to ship, not just whether a feature is worth building.