Every AI-adjacent design leadership search on your board belongs to someone graded on one of two numbers: the rate at which AI spreads through the organization, or the cost of the first time it goes wrong.
You will hear both described. Only one is on their review, and it is never the one in the posting.
I read ten live senior design searches this week. Nine were checkable for a time-bound success target tied to the person they hire. None of the nine names one. Every one of them talks outcomes; not one names the instrument that will measure them.
Reset your default. All three tiers point the same direction. Adoption is what gets funded. Control is what gets said. Enterprise postings sound the most control-oriented and score the most adoption-driven. Growth-stage postings name both and rank neither.
Control language in a posting is a hypothesis. Never a signal. Assume adoption until the person in the room proves you wrong.
So you have to hear it live.
The five-condition count
"Human in the loop" is an empty container. Every buyer says it. Almost none of them has paid for it.
A count tells you something. A person can genuinely stop an automated system only when five things are true at the same time:
- They can see the state. Not the output. The path the system took to reach it.
- They have a window. Real time exists between output and consequence.
- Their objection is legitimate. Raising a flag reads as doing the job, not as drag on it.
- The system obeys. Halt means halt, not "flagged for later review."
- Stopping is cheap for them. No career cost attaches to the person who paused it.
The framing is mine. It sits alongside NIST's AI risk framework, which refuses to treat "the human" as a single role, and its measurement playbook, which asks organizations to log overrides and escalations instead of asserting oversight. Condition five comes from Madeleine Elish's moral crumple zone: a human placed close enough to absorb the blame, too far from authority to have prevented anything.
Run the count in the first ten minutes. Ask which of the five their product currently has. An adoption buyer affirms all five and has funded none, because to them the question sounds like a compliance box. A control buyer has funded two or three and names the missing ones before you ask, usually with some irritation.
Three and five carry the diagnosis. Nobody fakes those two.
AI-native startup — adoption, unconcealed
Say: craft and velocity held together. Suno wants a leader who "moves fast without sacrificing craft", holding speed, experimentation and polish at once.
Evaluate: output per head against the model release cadence. The test is whether you can produce without an approval layer, because none is being built. Craft here is a quality floor, not a brake.
Tells. Interruptions land on duration, never on sign-off. Tooling questions, not reviewer questions. One signal flips the read, and it is specific: when this buyer asks what did your agent get wrong, and how did you find out, they have had a production incident. You are now talking to a control buyer inside an adoption company, and none of that appeared in the posting.
Wins them. Things they can open in a browser. Lead Agentic Labs, Carrier IQ first, because insurance quote automation runs the full handoff sequence in a regulated context on live APIs. Then place the current role verbally: Head of Product at TinyFish, enterprise web agent platform, Series A, three products in three months, agent traces in production daily. Keep that spoken. Currency, not case study.
Loses them. Org design vocabulary in the opening exchange. Headcount plans. Any answer where "governance" arrives ahead of "shipped."
Growth-stage platform — contested, adoption ahead
Say: coherence at speed with no bottleneck in the middle. Amplitude's Head of Product Design posting wants pods to keep moving and rejects the approval-committee model outright. Gusto wants a leader who models AI fluency from the front and puts prototypes in front of customers directly.
Evaluate: whether you will cost them throughput. Velocity gets measured by someone above this buyer. Incidents get measured by nobody, right up until there is one.
Tells. The word bottleneck, aimed at the role rather than the product. Push once and out come review-cycle time, ship count, prototype-to-feedback lag. Gusto's engineering leadership published the cleanest public account of the sequence: the organization stopped asking "are you using it?", started counting features shipped and production bugs, then scaled review rigor by risk while keeping payroll and tax code human-heavy. Note the order. The control layer arrived because throughput had doubled.
Wins them. Every control mechanism named as a speed mechanism, with a number attached to it.
Posture. Refuse to promise you will add no review steps. Ask instead what a production incident costs them today and who owns it. If the answer is nobody, you have located the mandate the posting could not name. Price it out loud.
Loses them. Proposing a ritual. Any safeguard without a throughput number reads as tax.
Enterprise software — adoption in control vocabulary
Say: control. Risk, safeguards, responsible deployment.
Evaluate: adoption, and usually of internal tooling. Capital One's Director of AI in Experience Design asks the hire to be a catalyst for accelerating AI adoption and to scale AI-driven design processes and tooling. Read that operationally and it is workforce upskilling plus tool penetration inside the design function. The risk language wrapped around it is sincere. It is not the scorecard.
This is the most expensive misread on your board. You see enterprise plus regulated, you lead the Thermo Fisher decision gate, and you have just pitched a customer-facing safety mandate to someone whose quarterly number is probably the percentage of their designers shipping with an AI coding assistant. None of the postings in the set disclose the real target.
Tells. One question resolves it. Who is the user of the AI in this role, our customers or our designers? AI pointed at employees means an adoption buyer speaking risk vocabulary. AI pointed at customers making consequential decisions means control is live. Ask early.
Wins them. On an internal transformation mandate, invert your usual order: the Trust essay and Agentic Labs lead, with Alibaba and Thermo Fisher underneath as ballast. The hook is that you have already built the agentic workflows they are trying to install.
Loses them. Assuming the role owns customer surface.
Where control buyers actually concentrate
The default read is adoption. Three conditions reliably invert it.
- Post-incident organizations. The tell is retrospective rather than aspirational: they ask what went wrong in your work before they ask what you shipped.
- AI pointed at a customer making a consequential decision. Not an employee. Not a draft. A commitment.
- Regulated verticals with a named signature. Clinical, financial, compliance. Somebody's name goes on the output and everyone knows whose.
Two of the three present, lead control. Fewer than two, lead adoption and hold control back for the follow-up.
Words that land, words that cost you
| Lands with adoption buyers | Lands with control buyers |
|---|---|
| shipped, throughput, weeks to value | gate, reversibility, exception |
| unblocked, cadence, per-head output | what it got wrong, and how I found out |
| adopted by, without a review layer | who signs, override, escalation |
Costs you with adoption buyers: governance, framework, alignment, rollout plan. Costs you with control buyers: velocity, ship fast, remove friction, autonomy.
Contested buyers get one bridge sentence: every control mechanism described as a speed mechanism, with a number attached.
One case, two readings
Thermo Fisher is the most flexible asset you own. The published facts read both ways and nothing has to be stretched.
Adoption read. Six of six pharma partners adopted. Nine sites, three regions, zero to live in twelve months, 42% out of overhead, $20M-plus in annual margin recovered. The claim underneath: my work gets used, fast, by people who were never obligated to say yes.
Control read. Exceptions were surfacing at the delivery gate, three to five times costlier there than a week upstream. You rebuilt the product exception-first. In the agentic redesign, five agents run continuously and exactly one human gate survives, batch QA release, because the regulatory signature cannot be delegated. The claim underneath: I know which gate to keep, and I can defend every one I removed.
Both readings are true, so truth is not the variable. Pick by which fear you are answering.
Whether one buyer flips, and why it barely matters
Working hypothesis, labeled as one. Organizations do progress. Atlassian revised its AI adoption metric three times inside eighteen months, pushing superusers from roughly 14% to 34% before concluding that volume rewarded repetitive prompting and missed actual workflow change. That is movement toward changed work. It is not yet movement toward authority.
The pattern you can act on is different: mandates get split, not sequenced. Webflow named a CPO holding product and design authority over the customer platform and, the same day, a first SVP of AI Transformation owning internal adoption. Two executives inside one company, carrying two different metrics.
Act anyway. Assume both frames live inside whoever you are speaking with, one funded and one aspirational, and lead with the frame the tier predicts. A posting that names speed and control with no ranking between them is an unresolved argument between two executives, and the hire is being recruited to settle it. So ask who the role reports through. My earlier decoder argued that a buyer reveals themselves in what they probe after an ordinary question. Reporting line is the sharpest instance I know. Probe it and they hand you the scorecard without noticing they did.
Who else is in the room
Inference from market pattern. I have no visibility into any actual pipeline. Confidence is labeled per row.
| Rival archetype | Objection they create | Your counter | Confidence |
|---|---|---|---|
| AI-native leader from a model or agent company | Your labs look small beside production scale at a name-brand company | Put Agentic Labs next to Alibaba's $50B-plus GMV platform. Most of that field brings scale or judgment. Show both inside one sentence. | Moderate as pattern, speculative per search |
| Enterprise design director | A product title in the current role reads as a detour off the design track | Lead the Alibaba mandate build: you named the structural gap from data, gaze tracking and 32 interviews, then won executive sponsorship to close it. TinyFish establishes present-tense currency and nothing more. | Moderate |
| Staff or Senior Staff craft candidate | Fresher hands-on artifacts, more recent prototyping in code | Agentic Labs first, solo-built on live APIs. Do not compete on tooling recency. Compete on what the artifacts decide. | Moderate |
| Vertical-domain specialist | Direct tenure in fintech, security or clinical simply looks safer to a panel | Match on consequence rather than industry: Carrier IQ for regulated automation, Thermo Fisher for compliance gates, Allē for provider-plus-consumer. | Moderate |
Action tiers
Ordered by contact priority. HackerOne and Amplitude go first this week. Both are searches where the proof you hold has no obvious substitute in the field. Ramp, Brex and Rubrik appear above as tier examples only and are not in this queue. Run the connection check before any of this becomes a sent message: a direct colleague means you ask them for the introduction, not the target.
| Company | Buyer read | Narrative fit |
|---|---|---|
| HackerOne — Director, AI Product Design | Control, funded | Your agents act across multistep workflows, but a bounty payout still waits on a human by default. I designed that same decision inside a pharma supply chain: five agents running continuously, one signature surviving at batch QA release. |
| Amplitude — Head of Product Design | Contested, trust lives in the product | An insight is worth only what someone will act on without re-deriving it first. I wrote the five-handoff framework for exactly that moment, and I read agent traces in production every day. |
| Maven Clinic — VP of Design | Control, clinical | Clinical outcome and conversion pull against each other until a non-expert can calibrate how far to trust what they are being shown. I have built that calibration for 30M-plus members, and for disaster caseworkers disbursing funds against a clock. |
| Vanta — Head of Design | Control, process-side | Definition of done and launch review are the same problem as a compliance gate: what has to be true before this ships, and who is permitted to say no. I designed that gate where a late catch cost three to five times an early one. |
| Gusto — Head of Design | Contested, control retrofitted after throughput doubled | The hard question in a unified service platform is who decides an answer goes out, and how quickly. I rebuilt a supply chain so exceptions surfaced a week earlier, where they cost three to five times less, and took 42% out of overhead getting there. |
| Capital One — Director, AI in Experience Design | Adoption, in risk vocabulary | You are installing agentic workflows across an experience design organization inside a bank. I have built those workflows solo, end to end, in a regulated vertical on live APIs, and published the framework for how humans keep authority while they run. |
After you send it
- AI-native (Suno, Brex tier): 48 to 72 hours is healthy. Silence past five business days usually means the message did not land. One follow-up, anchored to a new specific signal. Never "just checking in."
- Growth-stage (Amplitude, Gusto, Ramp, Vanta tier): three to five business days is normal. Follow up on day eight with one fresh hook, then move on.
- Enterprise (Capital One, Rubrik tier): five to seven business days is normal. Follow up day ten, then day twenty. Enterprise silence is process, not rejection.
- All tiers: a reply that engages the design problem you named, rather than thanking you for reaching out, is the strongest signal available. Answer inside 24 hours. A portfolio walkthrough request, or calendar availability in the first reply, moves that company to the top of tomorrow's queue. Two unanswered follow-ups is the ceiling.
-
Vanta's evaluation machinery: Before the Epling conversation, read how Vanta says it builds AI products — task-specific evals, governance-expert reviewers, real-workflow shipping thresholds — because it converts their "trust" language into named gates you can ask who owns.
-
The Gusto workforce mandate: Amy Thibodeau set a June 1 deadline for every designer to transform their workflow after hackathons and a pull-request requirement failed to stick, documented in Gusto's own account of the shift; read it as the adoption-metric buyer's playbook written down.
-
HackerOne's before and after: Compare the 2023 copilot-and-efficiency framing against the 2026 position on scoped permissions and mandatory verification — different authors, three years apart, and the clearest public record of an organization moving from oversight as principle to oversight as control.
-
The unresolved Amplitude org chart: Kim Lenox still appears as VP of Design on Amplitude's site while the open Head of Product Design role reports to the CPO; no public source explains whether that is succession, a new layer, or a stale page, so make it your first question in the screen rather than your assumption going in.