AI-native team structure - the roles that change

Every founder who calls us about "staffing for AI" asks the same first question, and it is the wrong one: how many people can I cut? We have built these teams. The companies that lead with headcount end up with a cheaper version of the old org and a worse product. The right question is what work survives.

Which roles to keep, which to redesign, and which net-new titles to hire once agents run the routine work.

Short answer: An AI-native team is different by design before it is smaller. Agents absorb routine execution: first-draft production, data entry, outbound sequencing, scheduling, basic support. That frees people for judgment, orchestration, and oversight. The roles that grow are the ones that direct agents, evaluate their output, and own the decisions agents cannot make.

Does an AI-native company actually need fewer people?

Leaner, yes. But the leanness is a byproduct of redesign, and treating it as the goal is where companies go wrong.

The data on size is real. The Harvard Business School AI Institute finds AI-native firms run 12 to 25% leaner than comparable non-AI startups, with fewer ops, admin, and entry-level roles, and flatter hierarchies. So a smaller org chart is a genuine pattern, not a vendor promise.

The mistake is reading that number as a cut list. Klarna shrank its workforce by roughly 40%, from about 5,000 to 3,000, and credited AI for part of it. Then it reversed course and began rehiring. CEO Sebastian Siemiatkowski admitted the over-rotation hurt service quality. The firm that cut to the number found out which work it had actually been doing.

The contrast that holds up is one of design intent. The leanest AI-native firms were built around the new division of labor from day one. Klarna cut to a number without redesigning the work; the redesign was the part it skipped.

Which roles go away when agents handle the routine?

The roles most exposed are the ones whose day is mostly routine execution: a single repeatable output, produced against a rubric, at the entry rung of a function.

The clearest signal is in the youngest cohort. The Stanford Digital Economy Lab and ADP Research "Canaries in the Coal Mine" study found that workers aged 22 to 25 in AI-exposed occupations saw a 16% relative employment decline after generative AI spread. More experienced workers in the same roles were unaffected, or improved. Experience is the variable. The work that an agent replaces is the work a junior person was doing to build toward the work an agent cannot.

The trend is not flattening. The 2026 update to the Stanford/ADP dashboard puts employment for 22-to-25-year-olds in highly AI-exposed occupations at a 3.8% annual decline. The rate has accelerated since April 2024. If you staff a function entirely from the bottom, that is the rung the agent takes first.

This does not mean delete the junior layer and move on. It means the work that used to train juniors now needs a new on-ramp, which is a design problem we return to below.

What does the manager role become?

The manager stops administering people and starts orchestrating a blended human-agent system. The job moves from oversight of tasks to direction of output.

McKinsey's work on managing for agentic AI estimates that 75% of current jobs will need redesign by 2030, with the manager role shifting from administrative oversight to running blended human-agent teams. The direction is what matters: the person who used to assign and check work now also decides what an agent runs, where a human takes over, and when the output is good enough to ship.

The same research names three management roles that did not exist before. M-shaped supervisors, who span several domains a single agent now touches. AI quality assurance leads, who own whether agent output meets the bar. Agent coaches, who improve how agents perform over time. These are not future titles. They are the redesign of the manager job already underway.

What net-new roles do you actually hire for?

The largest new category is not engineers. It is people in sales, service, HR, and operations who work alongside agents, and most of these roles require no code.

McKinsey's "Agentic Organization" describes six structural shifts for AI-native operations and finds the biggest emerging job category is "AI-augmented frontline workers," not technical staff. Ten of the 20 emerging roles it identifies need no code at all. The hiring implication is direct: you are staffing for people who can direct and judge agent work in a function, not only for the engineers who build the agents.

One genuinely technical role is worth naming because the market has already priced it. The GTM engineer sits between go-to-market and the systems that run it, wiring agents into the revenue motion. Per WorkLife's analysis, median compensation runs around $127,500, and top employers such as OpenAI, Vercel, Ramp, and Stripe pay between $184,000 and $252,000. The title barely existed two years ago. We cover where it sits in a revenue org in our AI-native go-to-market structure piece.

Watch out for the title that already burned out. Prompt engineering as a standalone job is largely obsolete, absorbed into workflow engineering and orchestration roles as models got better at interpreting plain instructions. If a vendor pitches you a prompt engineer hire in 2026, you are buying yesterday's job.

Staffing a function right now? Sign up for the newsletter - we write about org design for AI-native teams.

Is this a net loss of jobs or a net gain?

At the economy level the projection is a net gain, with heavy churn underneath it. Most roles turn over rather than disappear.

The World Economic Forum's Future of Jobs Report 2025, drawn from more than 1,000 employers covering 14 million workers across 55 economies, projects 170 million new roles created and 92 million displaced by 2030. That is a net gain of 78 million. The same report finds 39% of existing skill sets will be transformed or outdated by 2030. So the headline is growth, and the texture underneath it is a workforce being reskilled at scale.

The single-company version of that churn has a public example. Salesforce cut its support team from 9,000 to 5,000 as AI agents took over 50% of customer interactions, and support costs fell 17%. That is a real reduction in one function. It is also a reallocation: the dollars and the attention move to where humans still decide.

How do you redeploy the people whose tasks the agents now run?

Redeploy them up the value chain inside their own function, into the judgment and oversight work the agents created demand for. The freed capacity is the asset, and banking it as a cut throws the asset away.

Start by separating the person from the task. The task was routine; the person carries context, relationships, and domain judgment an agent cannot acquire. A support rep who knew which escalations were really churn risks becomes an AI quality assurance reviewer for the support agent. A sales development rep who understood why a sequence landed becomes the human in a blended pipeline.

Rebuild the junior on-ramp deliberately, because the Stanford/ADP data shows the entry rung is exactly where agents bite first. If routine work no longer trains your juniors, you need a path that puts them on agent oversight and judgment work early, or you starve your own future senior bench. This is the design problem most cut-first companies discover 18 months too late, after the pipeline is empty.

How an AI-native team is staffed: the decision

Staffing an AI-native team is four decisions, and they belong to the CEO and head of people, not the AI vendor.

1. Where is the routine-judgment line in each function? Everything an agent owns sits on the routine side; everything a human keeps needs context the agent cannot reach. Draw it wrong and you either automate nothing that matters or hand an agent a call it should not make.
2. Which roles get redefined, not removed? The manager becomes an orchestrator. Frontline workers become agent directors. The McKinsey estimate of 75% of jobs needing redesign is your planning base, not a layoff plan.
3. Which net-new titles do you hire? AI-augmented frontline roles first, since they are the largest category and need no code. A GTM engineer where the revenue motion justifies the comp. No prompt engineers.
4. How do you keep a junior on-ramp? The entry rung is the first thing agents take, so the path that used to train juniors has to be rebuilt around oversight and judgment work, or your senior bench goes empty.

Answer those four and you have a team designed for the work that is left. Skip them, cut to a number, and you get Klarna's first act: a smaller org, a worse product, and a rehiring announcement.

Frequently asked questions

What roles does an AI-native company actually need? Roles that direct agents, evaluate their output, and own the decisions agents cannot make. McKinsey finds the largest emerging category is AI-augmented frontline workers in sales, service, HR, and operations, and 10 of its 20 emerging roles need no code. The function names stay; the work inside them shifts to judgment and oversight.

Which jobs go away when AI agents handle routine work? The most exposed are entry-level roles built on a single repeatable output. Stanford and ADP found a 16% relative employment decline for workers aged 22 to 25 in AI-exposed occupations, while experienced workers in the same roles held steady. Experience, not job title, predicts who an agent replaces.

What is a GTM engineer and does my company need one? A GTM engineer wires AI agents into the go-to-market motion, sitting between revenue and the systems that run it. Median comp runs around $127,500, with top firms paying $184,000 to $252,000. You need one when agent-driven outbound and pipeline are central to revenue, not as a default first hire.

Is prompt engineer still a real job title? Largely not. Prompt engineering as a standalone role is now mostly obsolete, absorbed into workflow engineering and AI orchestration as models got better at reading plain instructions. The skill still matters inside broader roles. The dedicated six-figure title has faded fast.

How many people does an AI-native startup employ versus a traditional one? Fewer, by a measurable margin. The Harvard Business School AI Institute puts AI-native firms 12 to 25% leaner than comparable non-AI startups, with fewer ops, admin, and entry-level roles and flatter hierarchies. The leanness comes from designing the work around agents from the start, not from cutting an existing team to a target.

A team built for the work that is left

We put together a free AI-native staffing map: the routine-judgment line drawn for each function, the roles to keep, redefine, and hire, and the junior on-ramp most teams forget. Give us your email and we will send the link. Get the staffing map. If you would rather just follow the thinking, sign up for the newsletter to get the next piece on AI-native org design.

Sources

1. Harvard Business School AI Institute, "Less Headcount, More Valuation: How AI-Native Firms Change the Game." aiinstitute.hbs.edu
2. Stanford Digital Economy Lab / ADP Research, "Canaries in the Coal Mine: Six Facts About the Recent Employment Effects of Artificial Intelligence." digitaleconomy.stanford.edu
3. McKinsey, "Rethink Management and Talent for Agentic AI" (cite as McKinsey internal estimate). mckinsey.com
4. World Economic Forum, "Future of Jobs Report 2025" (1,000+ employers, 14 million workers, 55 economies). weforum.org
5. CNBC, "Klarna CEO says AI helped company shrink workforce by 40%" (2025). cnbc.com
6. CNBC, "Salesforce CEO confirms 4,000 layoffs 'because I need less heads with AI'" (2025). cnbc.com
7. Fortune, "Prompt engineering, the $200K six-figure role, is now obsolete thanks to AI" (2025). fortune.com
8. WorkLife, "The Rise of GTM Engineering: How AI Is Creating the Go-to-Market Job of the Future." worklife.vc
9. McKinsey, "The Agentic Organization: Contours of the Next Paradigm for the AI Era." mckinsey.com
10. Fortune, "What the Stanford/ADP Canaries dashboard shows about AI and entry-level jobs" (2026). fortune.com

---

Christopher Kliebenstein builds AI-native operating models and agent workflows for founders and operators at Kliebenstein AI Studio.