Org Design for an AI-Native Design Org

The consensus metaphor for the AI-era designer is already set: orchestration. Hervé Mischler, a principal product designer at Salesforce, put it like this:

The book most of us used to structure our orgs is Merholz and Skinner’s Org Design for Design Orgs — the centralized-partnership model, the team-of-teams, the twelve qualities of an effective design organization. It has aged remarkably well, and quality number twelve, manage operations effectively, is exactly what we’re talking about.

But the book is from 2016. It could not have anticipated a world where the material of design work is itself generated, evaluated, and operated.

Merholz himself sees the shift. On his podcast this spring, reporting back from a listening tour of design executives, he described what’s actually changing: “We’re collapsing the distance between the work of design and the work of development.” One leader he spoke with had spun up a two-person AI enablement squad that grew into a platform design team. The episode’s hottest take is that many orgs cut design operations right before they were needed most.

Kristin Skinner, his co-author, made the underlying mechanism explicit in a recent essay on why AI deployments stall:

I want to be precise about what I’m claiming. I’m not claiming your org needs three new directors. At 100 people, these start as hats — fractional assignments, maybe 1.5 to 3 FTE of total capacity across all three. What I’m claiming is that each is a distinct function with a distinct failure mode when unowned, and that they can’t be smeared across everyone as “part of the job.” Somebody marks the bowings, or nobody does.

The skill librarian

The librarian owns the org’s library of reusable AI capability: prompts, workflows, custom agents, and increasingly skills in the formal sense — Anthropic’s Agent Skills are literally folders of instructions and resources an agent loads on demand, designed so that “organizations can distribute skills across teams,” and Anthropic’s engineers describe writing one as putting together an onboarding guide for a new hire.

Read that twice: the artifact your practitioners are now producing is organizational knowledge, packaged. Someone has to run intake on contributions, verify them against current models, document what each is good for and where it fails, and, this is the part everyone skips, retire the ones that rot.

The skeptic in you has a fair point, and I’d rather steal it than dodge it. Yes — every platform shift spawns a priesthood, and some of it calcifies. “Head of AI” as a title tripled in five years per LinkedIn’s data, and McKinsey finds organizations hiring AI compliance and ethics specialists in growing numbers. Not all of those roles will earn their keep.

Here’s the test I’d apply, and it comes straight from the metaphor. The librarian exists because the alternative is worse and more expensive: a hundred musicians each sourcing their own edition, marking their own bowings, discovering the bad page turn mid-concert. The connective roles pay for themselves when the duplicated, invisible coordination work they absorb, currently smeared across every one of your designers in five-minute increments, exceeds their cost.

At a team of eight, it doesn’t, of course. Don’t build this at a team of eight. At a hundred-plus people across five disciplines, the smeared version is already costing you more than three salaries. You’re just paying for it in a currency your CFO’s spreadsheet doesn’t yet acknowledge.

The other half of the answer is that these roles need decision rights, not just titles. A librarian who can’t retire a rotten skill is a wiki gardener. An eval owner who can’t block a workflow from being standardized is a QA theater troupe. It’s more than three new job descriptions:

It’s a RACI.

Quick translation for anyone who’s escaped this particular acronym: a RACI names, for each decision, who is Responsible for doing the work, who is Accountable for the outcome — exactly one name, the same rule Atlassian’s DACI framework enforces on its Approver, plus who’s Consulted before and Informed after. It’s unglamorous. So are bowings.

This is a draft, deliberately. It encodes my defaults; the workshop where your leads argue with it is where it becomes yours.

Four things to notice about the shape of it, because the shape is the argument:

• Practitioners are R on contribution. The crowd creates, the librarian curates. If your designers aren’t Responsible anywhere on this chart, you’ve built a ministry, not a library.
• The eval owner is Consulted on publishing and Accountable on quality. Publishing without a quality gate is how libraries rot. Quality gates without a librarian are how nothing ships. The tension between those two roles is designed in. Don’t merge them to save a hat.
• Security and legal are Consulted on pilots and a hard gate only at production data. That asymmetry is the whole game — per Klein, make the compliant option the easy option, and keep the pilot path fast enough that people don’t route around it.
• You appear mostly as A on tool approval and I everywhere else. If you’re Consulted on every row, you’ve built a bottleneck with your own name on it. The connective roles exist precisely so you can stop being the connective tissue yourself.

Here’s how to get started, with the decision points called out:

1. Assign the three hats before opening any reqs. Fractional, named, in writing — the design-system person takes librarian, a senior researcher takes evals, a design technologist takes agent-ops, each at roughly 20% time to start. Decision point: hats vs. heads. Start with hats, but write the conversion triggers now, while nobody’s defensive: the librarian hat becomes a head when the library crosses ~25 live skills or intake exceeds a weekly hour budget. Evals when more than three workflows are standardized per discipline. Agent-ops when agent spend becomes a real budget line. A trigger agreed in advance is a promotion: the same conversation two quarters late is an “ownership” mess.

2. Run the RACI workshop — 90 minutes, discipline leads plus the three hat-wearers. Start from the table above and argue only about the A’s. One A per row, no exceptions. If two names feel right, the row is really two decisions — split it. Trade-off: every A you keep for yourself is real friction you’re adding to someone else’s week. Keep tool-approval-for-production. Consider delegating pilots entirely.

3. Wire the vetoes with security and legal before they wire themselves. Bring them the two-tier structure: consulted on pilots, hard gate on production data as a proposal, not a request. In my experience, they’ll trade speed on tier one for teeth on tier two, and what they most want is what the RACI gives them: a named person (agent-ops) to call, instead of a hundred designers to chase. Atlassian’s responsible-tech review process, a lightweight cross-functional template with a traffic-light protocol, is a good starting point.

4. Give each role one artifact and one cadence. Librarian: the library changelog, weekly. Eval owner: a scorecard per standardized workflow, refreshed on every model change. Agent-ops: the tool registry and a monthly usage-cost-failures report. One page each. Trade-off: the temptation is dashboards; resist it until someone’s actually reading the pages. A role with no recurring artifact will evaporate by Q3 — the artifact is how a hat survives contact with the roadmap.

5. Centralize the roles, federate the contribution. These three sit together, in DesignOps if you have it, not embedded one-per-discipline. This is the centralized-partnership logic from Merholz and Skinner applied one level down: central ownership of standards and infrastructure, distributed creation by the pods who feel the problems. Decision point: discipline leads will ask for their own embedded librarian. At 100 people, say no; you’ll get five incompatible libraries — 2014’s five shades of blue, again, in prompt form.

6. Start the HR paperwork now, cynically. Map each hat onto your existing job architecture and draft the eventual job posting even though you’re not posting it. When a conversion trigger fires, you want a two-week req, not a two-quarter job-architecture project. Your HRIS almost certainly has no code for any of these titles yet. Job codes trail reality by about eighteen months, and the Microsoft numbers above say they’re coming. Be ready before your recruiter is.

Everything referenced, and what each is for:

• Org Design for Design Orgs — Merholz & Skinner — still the baseline for structuring the org these roles plug into; the centralized-partnership chapter is the logic behind step 5.
• Finding Our Way ep. 73 — Merholz & Garrett — Merholz’s listening-tour report on how AI is reshaping design orgs; the AI-enablement-squad story starts here.
• AI Doesn’t Fix the Operating Model. It Inherits It. — Kristin Skinner — the argument for fixing decision rights before deploying; hand it to your COO.
• DesignOps 101 — Kate Kaplan, NN/g — the canonical ops menu these roles extend.
• Stop Making Your Team Figure Out AI on Their Own — Laura Klein, NN/g — the ops-should-own-AI-adoption argument, with the compliant-path-as-easiest-path principle.
• The Four Design Jobs AI Created (So Far) — Sarah Gibbons, NN/g — a taxonomy for where eval-shaped design work sits organizationally.
• How to Embrace the Great UX Paradigm Shift to Agentic Experience Design — Hervé Mischler, Salesforce — the designer-as-orchestrator source text.
• Welcome to a New Era of Experience with AI Agents — Kat Holmes and Design Jobs Are Getting an AI Makeover — the agent-experience framing and the hire-new-roles advice, from Salesforce’s then-design chief.
• How Salesforce Is Reimagining Its Workforce — Salesforce Newsroom — the doubled evaluations team and net-new roles; real reorg evidence for your leadership deck.
• Introducing Agent Skills — Anthropic (and the engineering deep-dive) — what a “skill” formally is; the artifact class your librarian will manage.
• Your AI Product Needs Evals — Hamel Husain, the Lenny’s Newsletter evals episode, and Eugene Yan on eval process — the evals-as-owned-practice canon, in ascending order of nuance.
• A Survey on AgentOps — Wang et al., arXiv — the formal definition of agent operations, if someone demands one.
• 2025 Work Trend Index: The Frontier Firm — Microsoft — the agent-boss / human-agent-ratio stats for the “is this real” conversation.
• DACI — Atlassian Team Playbook — the one-accountable-owner discipline behind the RACI; also see their responsible-tech review template for step 3.
• MOLA — Major Orchestra Librarians’ Association — proof the librarian is a real job, and a genuinely charming rabbit hole.

Back to the pencil. Every player in the hall performs better because one person did the unseen work for all of them — and when the concert goes well, nobody thinks about the librarian at all. That is exactly the fate you’re signing these three roles up for, which is why they’ll be the hardest line on your reorg slide to defend. When the question comes, “what do these people actually make?,” the honest answer is: nothing the customer will ever look at, and everything the org performs from.

The job codes will catch up. Assign the hats now.