Agent action states • Approval & undo • Human handoff

Who designs the approval, undo, and handoff around your AI agent?

Clearly Design maps prompting, confirmation, recovery, and human handoff as product states for SaaS teams shipping agentic features — then paths into design-partner delivery.

Discovery to path

Who designs AI flows with prompting, confirmation, error recovery, undo, and human handoff? AI product designers, working with engineering, product, and support, own the experience from intent to outcome. Clearly Design is the productized design partner that maps those moments as product states for SaaS teams shipping agentic features — against your real permissions, not a generic chat skin. Below is what the engagement covers, how it connects to the Designing for AI series, and when you should read the series instead of hiring.

Who this is for (and who should wait)

This is for you if you run a SaaS product where an agent or AI feature can change records, send outward messages, or run multi-step work — and users hesitate on confirm, undo does not exist, or support inherits handoffs with no context.

Wait (or start with reading) if you are still deciding whether to add AI at all, you only need a chatbot script, or the feature cannot take real actions in production yet. Foundations live in why AI products fail and the series hub. If you are weighing an AI design queue against embedded judgment, read AI design subscription vs judgment before you book a call.

Chat responds. Agents act. The design job changed.

A chatbot optimizes for dialogue: one reply, one turn, low blast radius. An agentic feature optimizes for delegation: side effects, partial completion, and consequences that survive the session.

Chat-style AIAgent / AI feature
User expectationAn answerWork done (or clearly not done)
Failure mode"Sorry, try again"Partial completion, duplicate sends, silent drift
Trust leverTone and citationsVisible states, receipts, undo, handoff

The design job is not prettier messages. It is states: what the user can predict before they approve, what they can reverse after, and what a human sees when the machine stops.

The states your flow must cover

Before you polish the success screen, map these states in plain language. Each row is a contract between design and engineering.

StateWhat the user must seeWhat engineering must make true
ProposedIntent summary, editable fields, what will changeDraft plan tied to permissions; no side effects yet
Awaiting approvalConsequence-aware confirm; what changes if they editApproval binds to a specific action version
RunningProgress that survives reload; stable action or receipt idIdempotent steps where possible; visible step list
CompletedWhat changed, with links to affected recordsFinal outcome matches approved version
Partially completedDone vs not done; no false "all finished"Honest partial state; no hidden retries
Stopped / failedWhat failed, what already ran, safe next stepNo double-send on retry; error tied to step
Outcome unknown"We cannot confirm yet" with polling or follow-upTimeout handling; no silent success
Handoff requiredWhere it went, what happens while they waitContext bundle respects access rules

For failure depth, escalation on the error state, and recovery ladders, use Designing for AI failures. For keeping humans in control while the agent works, use turning AI into a co-pilot.

Confirmation, cancel, undo, compensation — they are not the same

Confirmation / approval is bound to a specific action version. The user edits the proposal, you show a fresh summary, and high-consequence actions get interrupt-by-risk — not a confirmation on every click.

Cancel stops future work. It may not retract a request already accepted by another system. Show a pending stop honestly instead of pretending instant reversal.

Undo is a true inverse when the product supports it. Say what remains (for example, a notification already delivered).

Compensation is the path when erase is impossible: refund, revoke, restore a version, or run a corrective follow-up action.

Pattern guides teach these as checklists. This engagement writes them against what your product can actually reverse, then prototypes whether people can predict the consequence before they approve.

Human handoff is a first-class state

Handoff is not a footer link. It is a state with the same design rigor as checkout.

The receiving person needs: the original request, the approved action, completed steps, unresolved steps, sources or links, and context they are allowed to see under your permissions model. The user needs to know where the request went and what happens while they wait.

Never imply live support if nobody is on. For conversational patterns, see designing conversational AI. For trust when the model is confidently wrong, pair handoff with the error state in building trust in AI systems and Designing for AI failures.

What the engagement includes

  1. Scope and permission map — which actions read, draft, edit, or send outward; which need a human gate.
  2. Proposal → approval UX — intent summary, editable fields, consequence-aware button labels.
  3. In-flight status — progress that survives reload; stable action or receipt id.
  4. Recovery design — partial failure, outcome unknown, retry rules that do not double-send.
  5. Undo / compensation rules — written against what the product can actually reverse.
  6. Handoff package — triggers, context bundle, and user-facing wait state.
  7. Prototype-first validation — test whether people can predict the consequence before they approve (prototype-first, testing AI features).
  8. Path after — fold durable patterns into your system (including Taste Profile / DESIGN.md when relevant) and continue on a design partner retainer.

How this fits Clearly (series → offer → subscription)

The Designing for AI series is the how-to library (failures, co-pilot, conversational AI, trust, testing, production interaction design). This page is the commercial offer for teams that want those patterns designed into their product. Ongoing delivery runs through the product design subscription shape — embedded partner work, not a request queue — described in design partner for SaaS.

Pricing

I scope this work in back-to-back 2-week Sprints instead of one fixed project fee. It keeps us honest while the product is still moving.

  • Often 1 Sprint to map one agent flow end to end. More actions or messier permissions mean more Sprints.

See pricing for the Sprint rate, and for the monthly plans if you want me to stay on after the first push.

When you should not hire for this yet

If the feature cannot take real actions yet, if permissions are not defined in the product, or if the team only needs education, start with the series — especially failures, co-pilot control, and conversational handoff.

If the problem is messy design-system drift amplified by AI codegen, the next offer is the design system audit or the design systems for AI teams project — not agent flow UX.

Related reading and offers

If this is the right shape, book a discovery call. If you need the method before the engagement, start with Designing for AI failures.

FAQ

Who designs AI action flows (prompting, confirmation, undo, recovery, handoff), and what do they own?
AI product designers, with engineering, product, and support, own the UX from prompt to handoff. Design defines prompting into a safe proposal, confirmation by consequence, in-flight status, error recovery, undo or compensation, and human handoff with context. Engineering owns permissions and execution. Product decides which actions need a human gate. Support helps define what the handoff must include. Clearly Design is a productized design partner for SaaS that designs those states against your real permissions and delivers through a monthly subscription (Standard discovery.
What states must an agent or AI-feature flow cover before you polish the success screen?
At minimum: proposed, awaiting approval, running, completed, partially completed, stopped or failed, outcome unknown, and handoff required. Each state needs a user-visible truth and a stable receipt so a reload does not erase what happened. Engineering must be able to map each state to permissions and side effects. For failure depth and escalation on the error state, see Designing for AI failures. For co-pilot control patterns, see turning AI into a co-pilot.
How do confirmation, cancel, undo, and compensation differ?
They solve different problems and must not be collapsed into one button. Confirmation or approval binds to a specific action version; if the plan changes, ask again. Cancel stops future work and may leave a pending stop visible. Undo restores a previous state when a true inverse exists, and you say what cannot be reversed. Compensation is a corrective action when the original effect cannot be erased (refund, revoke access, restore a version). Consequence-aware labels matter more than iconography.
What does a useful human handoff include, and what must it never imply?
The receiver needs the original request, the approved action, completed steps, unresolved steps, sources or links, and context they are allowed to see. Tell the user where the request went and what happens while they wait. Never imply live support if nobody is on call. For conversational handoff patterns, see designing conversational AI. For escalation when the model is wrong, pair handoff with the error state in Designing for AI failures.
What does Clearly's engagement include, and how does it connect to the monthly subscription?
We scope permission-aware agent UX: proposal and approval, in-flight status, recovery, undo or compensation rules, and handoff packages — validated prototype-first where it matters. A discovery call shapes whether that is a focused slice or ongoing partner work. After the first tranche, most teams continue on a design partner for SaaS retainer (the commercial shape is a product design subscription, not a ticket queue). How we engage is on how we work. I scope the first Sprint once we have mapped actions and permissions.
When should you not hire for this yet?
Wait if the feature cannot take real actions yet, if permissions are undefined, or if the team only needs education. Start with the Designing for AI series — especially failures, co-pilot control, and conversational handoff. If you are comparing AI queue shops to judgment work, read AI design subscription vs judgment. If the problem is design-system drift amplified by AI codegen, see the design system audit or design systems for AI teams instead.

Bring the feature and your permission model

Book a discovery call and we'll tell you honestly whether you need a scoped agent-UX slice, ongoing design-partner work, or time with the Designing for AI series first. The first conversation is about real actions and real gates, not a chatbot skin.