1. The Objection
"Has she actually built AI products, or is this traditional design with an AI label on it?"
Every tier generates this. The trigger point differs. Frontier-lab companies (OpenAI, Anthropic) surface it during portfolio review. AI-augmented platforms (Ramp, Gusto, Headway) surface it on screening calls. Enterprise AI (Salesforce, Adobe) surface it in final rounds when the panel debates whether enterprise experience transfers. Same surface question, three fears underneath. Answer the wrong one and you've confirmed the concern instead of resolving it.
2. What They're Actually Asking
Three buyer types, three anxieties. Diagnose which one is active before you open your mouth.
Frontier-lab buyers fear you can only skin a model output. OpenAI's Model Designer role describes shaping model behavior and interaction patterns. Anthropic's Claude.ai design team builds the interfaces and interactions that turn a capable model into a product people actually use. Neither company is hiring someone to decorate outputs. The fear: she'll design a wrapper, not a system.
AI-augmented platform buyers fear you'll break trust during the redesign. Ramp's Director posting describes redesigning spend management around AI agents. Gusto's published AI principles center on keeping humans informed and in control during consequential HR and payroll decisions. The fear: she understands trust as a concept but hasn't redesigned a live workflow where a wrong agent action costs real money.
Enterprise AI buyers fear you don't speak governance. Salesforce describes its Agent Fabric as a control plane with human checkpoints, trusted agent identity, and mobile approval for high-stakes actions. Adobe's responsible AI framework centers pre-market evaluation, access-control roles, and continuous human oversight. The fear: she'll optimize for delight when they need someone designing for auditability.
3. The Strongest Case Against You
- Zero frontier-lab tenure. No time at OpenAI, Anthropic, DeepMind, or any organization where design directly shapes model behavior through RLHF or interaction-pattern decisions.
- Agentic Labs projects (Brand Pulse, Retail Velocity, Carrier IQ) are solo-built explorations. No real users bearing real consequences.
- The Alibaba agentic sourcing-loop and Thermo Fisher five-agent coda are speculative redesigns appended to enterprise cases. A skeptical reviewer reads "If I were building this today" and "Same problems. Different architecture" as thought exercises, not shipped work.
- TinyFish is a product role, not a design role. The work is not in your published portfolio.
4. The Grain of Truth
You have not worked inside a frontier lab. For companies that specifically want someone who has shipped model-behavior changes from inside a research organization, no reframe closes that gap. The deeper technical-depth question gets its own dossier.
Your AI work is not cosmetic. But that distinction only carries weight if you make it concrete, one control mechanism at a time.
5. The Reframe
Every company on your list is hiring for the same core design problem, whether they frame it that way or not: when an agent acts, what does the human need to see, control, reverse, and verify? Trust is the theme. Control is the mechanism. You have a published, inspectable body of work that names and designs those control mechanisms with more specificity than nearly any candidate at this level.
Start with the five handoffs framework. Each moment in an agentic workflow gets a designed control surface: Intent-Setting, In-Progress, Output Review, Decision Gate, Loop Feedback. The trust ladder (Watch → Verify → Delegate) describes how users actually build confidence in agent systems. Your line "Accuracy is what the model achieves. Consistency is what the design delivers" lands because it is operationally precise, not aspirational. And the framework is not theoretical. The essay derives it explicitly from designing trust systems across Alibaba, Thermo Fisher, and Red Cross over a decade. Three different domains, three different stakes profiles, one accumulated instinct for where humans need control.
The agentic codas prove you can apply that instinct architecturally. The Thermo Fisher coda maps five supply-chain modules to five agents and identifies the one irreplaceable human gate: batch QA release, where a regulatory signature cannot be automated. The Alibaba coda maps the full autonomous loop: an AI sourcing agent that parses a procurement brief, traverses search and PDP in real time, closes the order autonomously. Thermo Fisher shows where the human stays. Alibaba shows the loop the human oversees. That range matters because Salesforce's own human-at-the-helm patterns (mindful friction, confirmation steps, citations) map almost directly to your five handoffs. You arrived at the same design intelligence from a different direction.
TinyFish adds production currency across all three tiers. You are building and deploying enterprise agents daily, working through agent traces, auditability, and governance in production. Use it verbally as context. Do not cite it as portfolio proof.
Tier Notes:
- Frontier-lab: Lead with the essay framework and Agentic Labs. The five handoffs give you shared language for how design shapes agent interaction. Do not overclaim model-behavior influence. Moderate confidence. Logical reframe, untested against frontier-lab panels.
- AI-augmented platforms: Lead with Thermo Fisher (consequential workflow, real money, redesigned for agent architecture) and the trust ladder. High confidence. Multiple evidence points directly address this buyer's fear.
- Enterprise AI: Lead with the Thermo Fisher human gate, the Alibaba autonomous-loop vision, and the Decision Gate framework. Your governance vocabulary already matches Salesforce's control-plane language and Adobe's responsible-AI framework. High confidence.
6. What to Say
For platform and enterprise conversations, lead with the first response. For frontier-lab, lead with the second.
"I design control surfaces: visibility, escalation, reversibility, decision gates, audit trails. At Thermo Fisher, I mapped five supply-chain modules to five agents and identified the one human gate that can't be automated — regulatory batch release. That's the design problem I solve. Where the human stays, what they see, what they can reverse."
"My published framework names five handoffs in every agentic workflow — Intent-Setting through Loop Feedback — each one a designed surface. I've been building and testing these at TinyFish in production and through three live AI systems in my portfolio."
7. What Not to Say
- "I'm passionate about trustworthy AI." Every candidate says this. It signals you think about trust as a value, not a set of control mechanisms you can name and build. The interviewer hears enthusiasm where they needed specificity.
- "My traditional design experience translates directly." The word "translates" asks the panel to take a leap of faith on your behalf. You built three live systems and a published framework. Lead with what you shipped, not with a request for credit transfer.
- "I've been studying AI closely." Studying is not building. You built. Say what.
8. Early Signals
- The follow-up cluster: They ask about your AI work, then immediately ask how close you were to the model, then ask who made interaction-pattern decisions. Three questions in rapid sequence. The frontier-lab variant is active. Redirect to the five handoffs and name specific control mechanisms before the fourth question arrives.
- The workflow question: "How would you approach redesigning [specific workflow] with AI?" is the platform variant surfacing. They want to hear you name what stays human and why before you describe what gets automated. Lead with the gate, not the agent.
- The governance probe: The interviewer asks about audit trails, compliance review, or approval workflows within the first fifteen minutes. Or the panel includes someone from security, compliance, or legal. Enterprise variant. Lead with the Thermo Fisher human gate and the Decision Gate framework. Name auditability, reversibility, and human authority before they have to ask whether you've thought about them.
Objection: "Is she really AI-native or is this traditional design with an AI label?" · Real fear: Frontier-lab: she'll wrap, not shape. Platform: she'll break trust in the redesign. Enterprise: she doesn't speak governance. · Lead with: Five handoffs framework + Thermo Fisher human gate (adjust by tier) · Say: "I design control surfaces — visibility, escalation, reversibility, decision gates, audit trails. My published framework names five handoffs in every agentic workflow, and I've tested them in production." · Avoid: "Trustworthy AI" — name the mechanisms instead.
- Gusto's design-org transformation: Gusto publicly documented how its design organization went from traditional to AI-native within a single quarter, including designers shipping PRs and design systems rebranded as "Builder Enablement" — useful vocabulary if you're interviewing there.
- Salesforce's trust pattern taxonomy: Salesforce's Responsible AI team published five human-at-the-helm pattern categories including Mindful Friction and confirmation steps at critical junctures, which map closely enough to your five handoffs that you should reference them by name in any Salesforce conversation.
- Gusto's three open design questions: Gusto's June 2026 AI principles post names three active problems — showing what an agent is doing, designing intelligent escalation, and preserving agency when automation makes consequential decisions — each one a direct match to your framework.
- Backdoor references are growing: A June 2026 Wall Street Journal report covered the rise of informal reference checks beyond a candidate's provided list, especially at senior levels, which means every claim in this dossier needs to survive repetition by people you didn't choose as references.