Product manager

EU market. The regulation was published in the Official Journal on 12 July 2024 and this article applies from 2 August 2026; the July 2026 amendment (Regulation (EU) 2026/1744) left that date in place but gave generative systems already on the market before it until 2 December 2026 to meet Article 50(2). It binds the provider of the system rather than the employer using it — which is why it sits on this page and not on a page about deploying AI. Two obligations, different in kind: Article 50(1) is a disclosure to the person in front of the product, waived only where being an AI is obvious to a reasonably well-informed, observant and circumspect person, and disapplied for law-enforcement systems subject to safeguards; Article 50(2) is a technical requirement that outputs be marked in a machine-readable format and detectable as artificially generated, effective and interoperable as far as technically feasible. The second is a build requirement with an engineering cost, not a line in a privacy policy. This record establishes that the duty exists and when it starts; it is not evidence of enforcement, of any product changing, or of anything about markets outside the EU.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

A risk factor in a software company's final IPO prospectus, written by the company while selling shares. It says its product designers, product managers and engineers gather inputs through user research, surveys and analysis of customer support interactions, including AI-enabled tools that summarize and categorize user feedback, and that these inputs may be incomplete, misleading, biased or unrepresentative and may be misinterpreted. It describes current practice at one company; it gives no figures on how much of the work the tools do and says nothing about the number of product managers.

An employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.

The regulator's own announcement, which matters because the site already holds the rule and this is the rule being applied. The listed failures divide into two groups and both read as product specifications that were not written. For generation services: no visible label on generated content; no visible label carried into exported files; no implicit label in file metadata containing attribute information, the provider's name or code and a content number; and implicit labels placed in non-conforming positions. For distribution services: not verifying implicit labels, not adding a visible notice around published content, not writing propagation information into file metadata, and not giving users a function with which to declare that content is generated. The stated measures were regulatory interviews, orders to rectify within a deadline, and removal or taking offline. Two limits are firm. The announcement names no application and gives no count — a batch is all it says — so nothing here supports a claim about scale. And it concerns services provided in China; the same feature list is not law elsewhere, though the EU's labelling duty runs in the same direction.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

California regulations binding businesses whose processing of personal information presents significant risk, including automated decision-making technology used for significant decisions. A business must conduct a risk assessment before initiating the processing and document whether it will initiate it; an individual who has the authority to participate in deciding whether the business will initiate the processing must review and approve the assessment; and the stated goal is restricting or prohibiting the processing when the risks to consumers outweigh the benefits. It binds businesses, not product managers by name, covers personal-data processing rather than every product decision, and first submissions to the regulator fall due later.

Failure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.

LinkedIn's then chief product officer describing changes to LinkedIn's own product organisation. He lists what stays with people — vision, empathy, communication (aligning and rallying others around an idea), creativity and judgement — and says of everything else that the company wants to automate it; the company was piloting new team structures, running small pods capable of building end to end with its tools, and launching an associate programme in which builders will not just write specs. It is one executive's account of a pilot, with no figures on how many pods exist or on product-manager numbers, and it records August 2025.

Small-scale trial in a real setting. Tells us the deployment conditions are being tested, not that they hold — so one pilot is never enough on its own; two independent ones are.

A mixed-methods study (arXiv preprint) of product managers at Microsoft, conducted from April to June 2025: 885 PMs answered a survey, 731 consented to telemetry analysis and 15 were interviewed. Among individual-contributor PMs' uses of generative AI, 22% said they use it to conduct user research or data analysis; participants more readily delegated low-accountability tasks, but hesitated with high-accountability work such as writing specifications or journey maps. Microsoft sells the tools studied, and the respondents are its own staff at an early-adopter company; self-reported use correlated only moderately with usage logs.

Measured, large-scale use of a tool for real work, where the decision to use it was the worker's rather than an employer's. It is more than a capability record — the work is real, not a demo — and less than a deployment record, because no employer put it into production, required it, or built a process around it. Weighted `cautious`: `automating` means the machine can do the task AND there are adoption signs, and this is an adoption sign — but usage can be experimental, and much of the measurement comes from a party with a stake, so one record is never enough and two independent ones are. Note who is counting. Vendor telemetry sees this directly and sells the tool, so such a record names that stake in its scope; a statistics agency asking firms whether their workers use AI in tasks sees the same channel with no stake at all, and that is the better source where it exists.

The work you will be given first — writing tickets and specs — is the exposed half, and doing it well will not by itself make you a product manager. What converts the role into a career is being trusted to decide what does not get built, and that trust is earned by being right in public about small things first.

If your week is mostly documents and status, that is the configuration under pressure, and it is also the configuration many companies have quietly created. If your week is mostly refusals you can defend and decisions you own, the exposure is low. The difference is not seniority, it is whether the role carries a decision.

Stay and strengthen

Make sure the role carries a decision

The exposed version of this job is the one that writes down decisions other people made. The durable version makes them. Which one you hold is usually a function of the company, not of your skill.

Reshape the role

Own what AI features are allowed to do

Every product with a generative feature needs someone who decides what it may get wrong and what it must refuse. That is a product decision, it is currently unowned in most teams, and in several markets it is becoming documented obligation.

Adjacent move

Move to where the constraint is real

Regulated products, hardware, payments — domains where the cost of shipping the wrong thing is concrete rather than reputational, and where the decision therefore has to be held by someone.

The wrong question for this role, because the exposure depends on which version of the job you hold rather than on time. Writing specs and tickets is already substantially machine work. Deciding what does not get built, and answering for a bet that failed, has not moved at all. A signal you can check yourself: of the last five things your team shipped, how many did you decide rather than document.

Yes, and well — that is precisely why the document was never the job. A specification is a record of a decision, and the value was always in the deciding: which customer to disappoint, which quarter to spend, which problem is actually the problem. If the document is the deliverable your organisation measures you on, that is worth knowing about your organisation.

Enough to tell an estimate from a guess, and that bar has moved rather than risen. When generation makes a prototype cheap, the useful technical skill shifts from knowing what is hard to build toward knowing what is hard to operate and what is hard to undo — those are the costs that no longer show up in how long the first version takes.

Cheap building makes the deciding more consequential, not less: when a team can ship five things in the time it used to ship one, the cost of building the wrong five is five times larger. What it does change is that the role can no longer be justified by coordinating the build. It has to be justified by the quality of the choices, which is a higher bar and a more honest one.