Adopt Davenport’s builder-type taxonomy — citizen automator, citizen developer, citizen data scientist — as the canonical role model, and add the role every taxonomy misses: the verifier, because agent-era work shifted from writing things to checking them.

Status is contested, and not only over which taxonomy wins. Whether citizen development generalizes to most employees at all is genuinely disputed; that dispute is at the bottom of this page rather than smoothed out of it.

Start with the definition everyone borrows. Gartner’s, as quoted in the peer-reviewed literature: an employee who creates application capabilities for consumption by themselves or others (a persona, not a title or targeted role), reporting to a business unit other than IT. Interview research adds two consistent attributes: no formal software-development training, and they build for the unit they work in. That is all pre-agent low-code research, so treat it as the baseline agents inherit rather than as agent-era evidence.

The governance consequence sits in “persona, not a title.” The object of governance is an activity. Interviewed experts report that many firms never officially recognize the term at all: people do the work as an on-top job folded into a regular one. Any inventory, offboarding process or role-based control that waits for a formal job title will find nobody.

The checklist

1. Map each builder to one type, and name the competing taxonomies rather than blending them. At least five are in circulation: a five-role low-code model (citizen developers, champions, platform support consultants, implementation consultants, professional developers); an HBR four-role model (scouts, designers, developers/automators, analysts); and two consulting sets (prompt engineer / agent orchestrator / human-in-the-loop designer, and AI architect / orchestration engineer / AI performance manager / AI training lead). The consulting sets are proposals with no evidence any firm staffs them, and two of the three roles in the first have not survived contact with practice as job titles. Pick the spine, cite the rest as variants, and never publish five taxonomies side by side.

2. Split owner from verifier, per build. Decompose each role three ways: ownership (defining success and constraints), verification (auditing outputs, handling exceptions), and execution, which is now agent territory. Ownership can’t be delegated to an agent because it turns on decisions about values rather than optimization. That framing is one author’s argument, not a measurement; the operational rule that follows is the load-bearing part. Owner and verifier must be different people, or verification inherits the build’s failure modes; see the reviewer pool.

3. Expect the job to become direction rather than production. In a preregistered randomized experiment with 2,234 participants across 11,024 advertisements, people working with an AI agent made 62% fewer direct edits to the artifact, produced 50% more of them per worker, and sent 17% more delegating messages. From the other side, 82% of engineers in a matched longitudinal cohort of 95 reported spending less time writing code, and the authors name the replacement: supervisory engineering work (directing, evaluating and correcting AI output). Self-reported perceptions on a small cohort, so the direction is corroborated and the magnitude uncorroborated. OpenAI’s own telemetry adds that more than 10% of Codex users manage three or more concurrent agents at some point in a week.

4. Put oversight responsibility in the job description, and reward its quality. Not as an implicit norm, and not measured by velocity. If a builder’s review counts shipped artifacts, they will ship artifacts and route around the gate.

5. Give every agent a named human who can answer for its actions. This has academic backing as a core function of agent infrastructure, and it is the field that offboarding and agent inventory both key on.

6. Never assign an agent a role meant for a person. The measured cost of putting agents on the org chart: personal accountability for errors fell nine percentage points, accountability attributed to the AI rose eight, and additional review rose 44%. The same study reports higher identity uncertainty and lower trust under that framing, behind a paywall, so carry those qualitatively. The effects appeared only among managers already exposed to org-charted agents, reviewing documents rather than code.

7. Screen, don’t conscript. Documented selection postures differ sharply: one firm uses a trait screen (logical mindset, technical competency, aptitude to learn) plus a job-fit screen favouring rules-based work; another trains anyone who asks and considers even program dropouts useful; a third runs a formal application, and one used a logic test as an entrance exam. Pick a posture deliberately. The best recruiting doctrine in the literature is also the cheapest: a firm’s existing shadow-IT offenders have already demonstrated initiative, business familiarity and problem-solving; see shadow agents.

8. Make the escalation path part of the role, not an admission of failure. A practitioner’s statement of the boundary: the builder takes the basic application, professionals take the complex one. And ownership carries maintenance: “you build it, you own it” means documentation and user support, not only creation. Where that lapses, the result is quality debt.

Who is already doing this

Not the population most firms would guess. In US survey data, managers use generative AI at close to the rate of computer and mathematical occupations (51.9% against 53.6%, with business and finance at 48.2%), more than double the rate of administrative workers despite similar predicted exposure. Adoption rises across the seniority distribution rather than concentrating in junior staff.

European workplace data, fielded in early 2024 and worth dating as such, shows adoption climbing from about 1.5% in the least AI-exposed occupational quintile to nearly a quarter in the most exposed. Three moderators survive the joint model: tertiary education, abstract task content, and having a say in how work is organized. Individual training does not independently predict adoption, though country-level training provision does; write both halves or neither. A gender gap of 4.1 percentage points overall widens to 7.5 within the most-exposed occupations and survives controls, so a builder pool left to self-select will skew exactly where the payoff is highest.

Why functional role definitions stop mapping is measurable. In a randomized experiment with 776 professionals, those working without AI stayed in their silos (R&D proposed technical solutions, Commercial proposed commercial ones) while individuals using AI produced balanced solutions regardless of background, matching the performance of teams working without it. That was a one-day task with a chat assistant, not agent building: it establishes the boundary-spanning mechanism, not the agent-era magnitude.

How you’d know it’s working

Every registered agent has a named owner and a different named verifier. Role assignments survive an org-chart pull without hand-waving: a person can be named for each of the three functions on any build, and the same name never appears twice on a build that matters.

What this doesn’t solve

Role labels don’t create skill. Behaviour predicts who amplifies and who rubber-stamps; title and seniority don’t. That is training and reviewer selection, not taxonomy.

Nor does a role model fix incentives. Where citizen development is uncompensated on-top work, the reward structure quietly undermines the governance structure: people ship fast and quietly rather than through gates.

And the central premise is contested by the only financial-services case in this corpus. One firm’s CIO reported that less technically savvy business staff lacked the mindset to build applications logically; the firm abandoned citizen development and reverted to trained in-house developers on low-code platforms. In the same study another firm gave up on transitioning existing employees, hired dedicated no-code developers instead, and now screens candidates with a cognitive assessment. The authors’ conclusion is blunt: the word “citizen” implies a large transferable pool, and that implication is unrealistic. Those interviews ran to August 2023, before generative tools moved the floor, and that is the reason to keep the case rather than dismiss it. It is the strongest counter to the everyone-can-build framing, and it comes from this audience’s industry.

The honest state of the question belongs to the practitioner who proposed business sense and systems thinking as the selection criteria, then added that he did not know whether either is innate or learnable. Neither do we. See reader-facing open questions.

See also

• The reviewer pool — how the verifier role gets staffed and rotated.
• Training curriculum — what each role type needs to learn, in hours.
• The support model — the structures these roles sit inside.
• The jagged frontier — why behaviour, not credentials, predicts who verifies well.
• Oversight decay — what happens when agents get treated as employees.
• Offboarding agents and owners — what happens to the role when the person leaves.