What role should AI play inside the organization — and what should remain distinctly human?
AI is playing the role of several people at once — and the metaphors we have don’t quite fit any of them.
Every CEO I’ve spoken with this year has asked some version of the same question underneath the questions: “What is the AI inside my organization actually going to be?” They mean it functionally — what will it do, what value does it create — but underneath that is a relational question. What is our relationship to this thing? Is it an employee? A tool? Something else?
The question gets sharper when you look at what AI is actually doing inside organizations already deploying it seriously. It drafts first versions of nearly everything. It extracts structure from what the organization produces — decisions from pull requests, action items from meetings, deviations from contracts. It connects silos, monitors continuously, executes within scope, and increasingly decides within boundaries. Some of that is junior work done faster. Some is senior work made tractable. Some is work no human was ever doing because no human had the bandwidth. The reason no single metaphor fits is that no single human role does all those things at once.
A better way to ask the question
Metaphors are load-bearing. The one a CEO reaches for quietly determines how AI gets governed, how employees are introduced to it, what authority it gets, and who feels threatened versus empowered. So instead of asking “which human role is the AI like?” — a question that forces a wrong answer — ask a more useful one: for each candidate role, what is its actual profile across five dimensions?
- Mental model: how they think about the organization
- Information: where they receive it
- Action: where they act on it
- Control: how much authority or gating they have
- Opportunism: reactive (waiting to be asked) or proactive (surfacing what no one thought to ask)
Run AI through those five dimensions and the profile doesn’t match any single human role — but it overlaps with several. That overlap is the actual answer.
The roles we keep reaching for
The all-knowing manager
- Mental model: hierarchical, information flows up, authority flows down
- Information: dashboards, status reports, performance metrics
- Action: directives, evaluations, gating decisions
- Control: high
- Opportunism: reactive — intervenes when metrics breach threshold
This is the worst metaphor and the most common one. The AI isn’t watching to evaluate — it’s reading to assist. Frame it as a manager and your people will treat it as one: defensively, performatively, looking for ways to game it. You’ll have built a panopticon when what you wanted was a partner.
Picture the engineer who fixes a production issue at midnight and walks into a calendar invite the next morning because the AI flagged her rollbacks as a “low collaboration signal.” The problem was solved. The relationship is broken.
The protective parent
- Mental model: developmental — people are growing toward independence
- Information: observation, intuition, signals of distress or readiness
- Action: guidance, intervention, obstacle removal
- Control: high but waning by design
- Opportunism: highly proactive — intervenes before harm occurs
Surveillance with affection added. It encodes infantilization, which is exactly the dynamic you don’t want professionals to have with AI. Senior people won’t perform well in a workflow that treats them like children, no matter how warmly the parent is framed.
Picture the senior PM who hits six “are you sure?” prompts in an hour. By the next week, he’s clicking past every flag — including the ones that would have mattered.
The assistant or junior employee
- Mental model: apprentice, learning from above
- Information: assignments and feedback
- Action: drafting, preparing, executing what’s delegated
- Control: very limited
- Opportunism: reactive — operates within tasks given
The metaphor most AI vendors use, and a misleading half-truth. A junior employee is on a development arc, has emotional needs, and operates as a separate agent with their own judgment. Treating AI as a junior leads to either over-trusting it or over-managing it.
Picture the partner who makes seventeen edits to the AI’s client deck and ships it — not noticing that across forty similar interactions, his judgment has quietly migrated into the AI’s defaults. He thinks he is supervising. He is being trained.
The chief of staff
- Mental model: the principal’s organization from the principal’s seat
- Information: everything the principal sees, plus what they miss
- Action: briefing, connecting, surfacing, capturing
- Control: none independently
- Opportunism: highly proactive but never directive
The closest human fit — embedded, non-evaluative, pattern-surfacing, no independent authority. But a chief of staff doesn’t do the work. The AI is doing real work: drafting, analyzing, executing, deciding within scope. This metaphor describes the posture and understates the contribution.
Picture the COO walking into a difficult client call, fully briefed on the open tickets, commercial issues, and the contract clause governing the slip — all surfaced before she sat down. That’s chief of staff. But in the next room, the AI is also drafting the renewal terms and routing the escalations.
The trusted senior colleague
- Mental model: the firm as a long arc of cases and lessons
- Information: decades of context, most of it not written down
- Action: consultation on request — “this came up in 2008, here’s what we did”
- Control: none formal; significant informal influence
- Opportunism: mostly reactive, occasionally interjects on familiar patterns
The metaphor most people don’t reach for first but probably should. AI overlaps with this role significantly — recalling what happened on a similar project, surfacing decisions that produced outcomes, making accumulated experience legible to people who weren’t there. The limitation: the senior colleague’s authority comes from earned trust. AI’s recall is more complete but less weighted by judgment about what actually mattered.
Picture the new associate prepping for a regulatory call who pulls up every prior interaction with this reviewer, including a note that they care disproportionately about traceability documentation. No human briefed her. The thirty-year colleague she’d have asked is on vacation.
The auditor or quality reviewer
- Mental model: work as a structure that should match its specification
- Information: the work product, the standards, prior cases
- Action: flagging gaps, deviations, risks
- Control: gating — work doesn’t ship without sign-off
- Opportunism: reactive, with strong proactive instincts inside their domain
AI genuinely checks work against standards — code against style guides, contracts against clause libraries, decisions against compliance frameworks. The danger is in the gating dimension. A human auditor’s authority comes from accountability. AI flags carry no such accountability, and treating them as gates produces a brittle outcome: humans defer to flags they shouldn’t and ignore flags they should heed.
Picture the engineering lead who overrides an AI flag on an architectural deviation — correctly. The deviation is exactly what lets the team ship on time. The AI was right and wrong simultaneously, and only a human could tell which.
The librarian
- Mental model: organizational knowledge as a structured collection
- Information: documents, decisions, artifacts
- Action: retrieval, synthesis, citation
- Control: none
- Opportunism: purely reactive — waits for the question
Fits retrieval and synthesis cleanly. Falls short on proactivity. The AI in an orchestrated network surfaces what you didn’t think to ask — the precedent you weren’t aware of, the risk invisible until three unrelated projects connect. The librarian is one slice of that; the proactive surfacing is the other, and the metaphor doesn’t cover it.
Picture the analyst who gets a full structured brief on the firm’s prior infrastructure deals in under a minute — but the AI doesn’t volunteer that two of them share a counterparty risk pattern with the deal she’s actually working on. She didn’t ask. The AI didn’t tell.
The on-call specialist
- Mental model: deep expertise in a narrow domain
- Information: specifics of the situation they’re called into
- Action: focused intervention, then exit
- Control: high within domain, none outside it
- Opportunism: pure reactivity — responds to escalations
A strong fit for narrow, deep, escalation-driven AI agents — fraud detection, incident triage, regulatory interpretation. Unlike most of the other metaphors, this one carries no misleading connotations. It just isn’t the whole picture.
Picture the platform team paged at 2am because the AI detected an anomaly that historically precedes a specific outage by 90 minutes. They trust it more than they trust themselves at 2am — on this, and only this.
Two metaphors that come closer
None of the human roles is wrong. Each captures a real slice. But the slices don’t add up to a person — they add up to something for which the closest analogues aren’t human roles at all.
Mycelial network: invisible infrastructure beneath the visible work, connecting decisions across projects and surfacing patterns no individual could see. It’s not above you. It’s beneath everything. No one owns it; everyone benefits from it.
Jazz ensemble: no conductor, every player doing real work, coordination through shared structure and real-time listening. Leadership shifts to whoever has the strongest line right now. This captures the peer relationship between humans and AI that the most advanced organizations are reaching toward — not hierarchy, not assistance, but genuine co-contribution.
What this means for how you build
Stop asking “what is the AI?” Start asking “which role is this particular AI playing, and is its profile across the five dimensions actually the one we want?”
A serious organization in 2026 isn’t deploying “AI” — it’s deploying a portfolio of AI roles. The fraud agent is an on-call specialist. The briefing agent is a chief of staff. The knowledge agent is a senior colleague. The compliance agent is an auditor. Each has its own profile, its own appropriate authority, its own human counterpart. Underneath all of them runs the mycelial layer — the connected knowledge graph that lets each draw on what the others have produced. On top, the jazz layer — where humans and agents actually work together in real time.
The leaders who build the next decade’s organizations won’t be the ones who pick the right metaphor. They’ll be the ones who notice the question itself was wrong — and replace it with a better one.
