Positioning Sheets — Stripe and Figma

What connects these two

Both companies are designing surfaces where agents act on behalf of users. The difference that matters: what it costs to fix the agent's output.

At Stripe, an authorized payment creates a financial obligation. Correction after commitment means refunds, recourse, a second transaction. At Figma, regenerating a rectangle costs nothing, but regenerating a complex multi-layer layout can destroy the parts the designer already accepted. That cost difference determines whether the interaction needs a gate before the action, a way to selectively undo after, or both. Your portfolio has proof for both sides. Use the sheets below independently.

Stripe — Senior Staff Product Designer, Data & AI

Posting: Senior Staff Product Designer, Data & AI Artifact domains: Human-Agent System Design, Agent Infrastructure as UX

Their frontier challenge

Stripe is building infrastructure where personal agents make purchases on behalf of consumers. Match their vocabulary exactly: delegated authority, scoped permissions, predictable and auditable agent actions, friction matched to risk, recourse when something goes wrong. Their Sessions 2026 program reinforces this framing — consumers delegate purchasing authority while retaining choice, transparency, and control.

The design problem inside this infrastructure has at least three distinct moments that a generic confirmation dialog would collapse:

• Setting the authority envelope (consumer's budget and constraints)
• Requesting additional authorization when conditions change mid-transaction
• Providing recourse after an action has already occurred

The Senior Staff posting asks for shared systems, policy and edge cases, high-stakes trust. The designer who gets this role owns how those three moments feel to both the consumer and the agent platform.

Lead with CarrierIQ

CarrierIQ is your closest working proof for consequential-action design with agent infrastructure underneath. Map these elements explicitly:

• Structured intake as scoped authority. CarrierIQ's intake rail — liability limits, deductibles, coverage selections, carrier scope — is a user defining the envelope of what agents are authorized to do. This maps directly to Stripe's "bounded standing authority" concept.
• Differentiated review states as friction matched to risk. CarrierIQ classifies carrier results into Bindable, Normalize, Referral, and Call review. Each classification carries a different review burden. A bindable result needs a quick check. A referral needs operator judgment. Friction is applied by classification, not uniformly.
• Coverage deltas as changed-condition surfaces. When a carrier's returned terms differ from the requested baseline, CarrierIQ renders the delta explicitly. This is the visual pattern for Stripe's incremental-authorization moment — conditions changed, here's what changed, decide whether to proceed.
• The Approve Bind gate. The bind action stays disabled until verification completes. An operator note field captures approval rationale. Decision gate before a financial commitment, with an evidence trail.

Support with the Trust essay. Trust Is the New Interface gives you framework language: intent-setting, in-progress visibility, output review, decision gate, loop feedback as five distinct handoffs. Stripe's three-moment structure maps to intent-setting, in-progress escalation, and post-decision feedback. Name the mapping when you reference the essay. Don't just link it.

Production proof

CarrierIQ is live and inspectable. A reviewer can walk through the intake, watch five carriers execute in parallel through five named stages (Session, Navigate, Fill, Extract, Verify), compare results, inspect evidence, and see the approval gate. This is production-quality interaction design for agent-mediated financial decisions.

Your differentiator

You have a working application where a human reviews agent-gathered evidence before authorizing a financial commitment, with differentiated review classifications and an operator rationale trail. Most candidates for this role will show AI features inside existing products. Your proof is a purpose-built system where the entire interaction model is organized around when and how a human authorizes an agent's consequential action. That specificity is hard to match.

What NOT to lead with

• Team-building or hiring narrative. The posting has zero management language. This is a strategic IC role. Suppress org-building stories entirely.
• Your analytical vocabulary. "Correction cost" and "irreversibility" are your internal frame. Use Stripe's language: delegated authority, scoped permissions, friction matched to risk, recourse.
• Generic fintech experience. Stripe's design challenge here is agent-infrastructure trust. Allē and other production case studies are supporting evidence, not the lead.

Figma — Design for Agent-Assisted Creation

No single posting identified. Figma's AI design hiring spans multiple related roles across their Make Design and platform teams. You're positioning for a team-level challenge, not responding to a specific req. Frame outreach around the problem space, not a job ID.

Artifact domains: Inference-Aware UX, Agent Infrastructure as UX

Their frontier challenge

Figma's VP of Product stated the regeneration problem directly: prompting a small correction can regenerate more than the designer intended, altering properties that were already acceptable. Figma's answer is direct manipulation — let the designer keep what works, change only what doesn't, take ownership of the output.

Their broader vocabulary centers on design context (components, tokens, layouts, standards fed to the agent before it acts), design-system adherence as a measurable output criterion, and showing work and citing sources as the basis for agent trust. Their workflow documentation includes confirmation modals, success toasts with undo actions, and explicit identification of missing states that need human addition before output is production-ready.

The design challenge for someone joining this team: how do you build review and correction surfaces for agent-generated design work where the cost of regeneration varies by element, where the designer needs to see what the agent did and why, and where the output must meet a design system's standards rather than just look plausible?

Lead with CarrierIQ's parallel execution, then Brand Pulse's attribution

CarrierIQ's parallel agent lanes are your clearest proof for making multiple independent agent processes visible simultaneously. Each carrier occupies its own row with activity state, progress percentage, trace messages, and position in a five-stage sequence. A reviewer can see which agent is where, which finished, which is stuck.

What matters for Figma: CarrierIQ's lanes are independently addressable. You can re-run verification on one carrier while another sits in Call review and a third is ready for bind approval — without any action on one lane affecting the others. That's the architecture Figma's canvas needs for selective correction. When a designer accepts the typography an agent produced but wants to correct the layout spacing, the system has to treat those as separate agent outputs with separate correction paths — preserving accepted work while regenerating only what failed. CarrierIQ demonstrates this independence at the task level. The material differs (insurance quotes vs. canvas elements), but the interaction principle holds: intervene here without disturbing there.

Brand Pulse's source-level attribution adds a second proof point. Each signal card identifies its source agent (Reddit, X), carries the collected excerpt, a theme label, and a sentiment classification. The platform-contribution view shows each source's share of total signals with a stacked bar. The user can see which agent contributed what and how much.

Be precise about what Brand Pulse does and doesn't show. It attributes signals to source agents and displays each platform's contribution share. It does not expose original post URLs, model identity, or component-level provenance. Don't claim it demonstrates design-system-level traceability. It demonstrates the pattern of making agent contributions individually visible and assessable — the foundation Figma's challenge requires, applied at a different layer. A technical evaluator who inspects the live app will catch overclaiming here.

Support with the Trust essay's "showing work" handoffs. The in-progress and output-review handoffs in Trust Is the New Interface map to Figma's stated principle that agents should show their work and cite sources. Use this for language around why visibility into agent reasoning is a design problem, not a feature request.

Design-system adherence — a gap to know about

Figma treats design-system adherence as a measurable quality criterion — their Coinbase case study evaluates agent output on component choice and system compliance. Your current portfolio doesn't demonstrate design-system architecture or token governance work. If this comes up, redirect to how CarrierIQ's review classifications and coverage-delta rendering demonstrate systematic quality criteria applied to agent output. Adjacent skill, different material. Don't fabricate a design-systems narrative, but don't concede the territory either — the evaluative structure (did the agent's output meet the defined standard?) is something you've built.

Production proof

Both applications are live. CarrierIQ's parallel execution is inspectable in real time. Brand Pulse's agent trace logs events with agent identity, event kind, and elapsed time as they stream in. Working systems, not wireframes.

Your differentiator

You have two shipped applications that make agent work individually attributable — one at the task level (CarrierIQ's per-carrier lanes) and one at the source level (Brand Pulse's per-platform contribution). Candidates from product-design backgrounds will show AI features embedded in existing tools. Candidates from design-systems backgrounds will show component libraries and token architectures. Your portfolio shows purpose-built systems where the core design problem is making agent contributions visible, assessable, and individually addressable for correction. That combination is rare.

What NOT to lead with

• Financial-transaction or approval-gate framing. Figma's challenge is about creative output quality and selective correction. CarrierIQ's approval gate is useful context but not the lead — the parallel lanes and their independent addressability are.
• Provenance claims that exceed what Brand Pulse actually shows. Source-agent attribution and contribution share, yes. Component-level canvas provenance, no.
• Scale or enterprise narratives. Figma's AI challenge is about craft, context, and the designer's relationship with agent output. Lead with interaction-design judgment.