New AI job titles

Why are new AI job titles appearing now?

Because model deployment stopped fitting inside the roles that already existed. An enterprise buying a frontier model is not buying software it installs; it is buying a capability that has to be designed into a workflow, evaluated against its own data, and then kept working. That job was landing half on sales engineering and half on product engineering, and neither side owned the outcome. The labs did the obvious thing and gave the work its own headcount line.

The mechanism is worth naming because it tells you which titles to take seriously. A title that exists because a job-to-be-done has no owner tends to persist. A title that exists because a capability is briefly novel tends to get absorbed back into an existing team within a couple of release cycles. Both kinds are on this page.

What is the difference between the two roles?

The split is pre-sale versus post-sale. The Applied AI Architect runs technical discovery, architects the proof of concept, and designs a deployment the customer can scale. The Forward Deployed Engineer arrives after the contract, sits inside the customer team, and writes the code in the customer environment. Anthropic runs both inside one Applied AI organisation, which is the clearest signal that the division is intentional rather than accidental naming drift.

For a CTO reading job postings, that distinction is the fastest way to tell what a company is actually hiring for. A posting heavy on discovery, architecture reviews and stakeholder workshops is the Architect. A posting that talks about shipping in the customer environment, on the customer’s stack, is the engineer.

What are the two risk titles, and why are they a pair?

Because one writes the rule and the other breaks it. Model Policy is OpenAI’s name for the seat that decides what a model will and will not do, and turns that decision into behavioural specifications and evaluation criteria that reach training and monitoring. The AI red team researcher then attacks the resulting product, at Microsoft across the Copilot surfaces and the Phi models, to find where the specification does not survive contact with a real user.

Neither title travels well. OpenAI is unusual in calling this function policy at all, where adjacent work elsewhere sits under Trust & Safety or AI Safety, and Microsoft’s AI Red Team titles describe work that other labs run under different names. That is the recurring lesson of this whole page: the job crosses companies and the label does not.

Titles to watch

These have appeared in real postings but do not yet show the multi-company pattern that would justify a full guide. No dedicated pages exist for them here yet. Listed with what each one appears to own:

• Model Welfare researcher — Studies whether and how a model’s own states should factor into deployment decisions. A research seat that did not exist as a job description two years ago.
• Societal Impacts researcher — Measures what a deployed model does to labour markets, information ecosystems and political discourse, and feeds that back into release decisions.
• Frontier Counsel — In-house legal specialised in frontier model release, licensing and liability, rather than general technology law.
• AI Systems Engineer (Agents) — Infrastructure for agent execution: sandboxes, tool routing, long-running state, and the failure handling an agent loop needs to survive production.
• Offensive Security Engineer, Agent Products — Attacks the company’s own agent surfaces. The security discipline reshaped for products where the attack surface is a model that takes actions.
• Agent Infrastructure Engineer (DeepSeek) — Agent execution infrastructure named as its own title at a lab outside the US, which is how you know the pattern is not one company’s naming quirk.

Four of these six are safety, security or legal-risk seats, and the two titles that graduated to full guides on this page came from the same group. The labs are not only inventing delivery roles. They are converting judgment work that used to sit with a committee into a named individual job with a reporting line.

Should you copy these titles into your own org?

Copy the job, not the label. If your AI deployments keep stalling between the team that sold the idea internally and the team that has to run it, you have the gap these titles were created to fill, and naming it helps. If they are not stalling, a new title adds a level and a salary band without adding throughput. The AI roles map covers the roles that are already load-tested; the Chief AI Officer guide covers the executive seat above them.

Titles and role scope verified against company job postings on Anthropic’s job board, OpenAI Careers and the Microsoft AI Red Team postings (August 2026).