Designed software
Stable screens for known paths. Best when the work repeats.
Part 01 / What it is
Software can adapt without giving up control.
Concepta thesis
Generative UI composes an approved interface for the task at hand. The application still owns the truth.
Fixed software is reliable but cannot anticipate every case. Chat is flexible but often hides state, evidence and available action. Generative UI keeps the flexibility while restoring a visible, checkable interface.
Conversational AI
Flexible input, but state, evidence and possible actions can disappear inside the answer.
Generative UI
Approved components adapt to the task. The application retains authority.
What it is not: arbitrary code generation, a chatbot with cards, or a replacement for the design system, permission model or product architecture.
Part 02 / Where it fits
Start where variation helps.
Choose a workflow where cases genuinely differ, actions are bounded, and a person or system can verify the outcome.
Review + approval
Contracts, escalations, policy exceptions, release risk and security findings.
Triage + investigation
Incidents, claims, requests, anomalies and operational exceptions.
Decision support
Scenario comparison, assumption changes, evidence inspection and rationale.
Configuration
Deployment, onboarding, entitlement, service options and workflow setup.
Role-specific reporting
One governed result, composed for executive, operator, auditor or client.
Complex data exploration
Charts, filters, drill-downs, annotations and follow-up actions.
Part 03 / How it works
The interface can change. The contract cannot.
Do not generate a new application for every request. Compose a focused surface over one governed capability.
The surface proposes. The app makes the change.
Part 04 / How it responds
Not every interaction needs the model.
The layout may regenerate when the job changes. Ordinary interaction should remain immediate, and the workflow must continue even if the surface disappears.
When the job changes
Recompose deliberately.
Structure and emphasis may regenerate when the intent, phase or material context changes enough to justify a new arrangement.
Expected behaviorValidated generation latency
Part 05 / Who decides
Put human judgment where the risk is.
Require review when the action is consequential, hard to reverse or uncertain. Approval everywhere only creates another queue.
AI may propose
Summarize, draft, compare, rank, surface missing context, arrange evidence and recommend a next step.
The system must govern
Identity, permission, app state, deterministic rules, policy checks, evaluation, tool execution, logging and rollback.
People retain judgment
Set goals, correct context, decide exceptions, approve high-impact action, override, reject, escalate and record rationale.
Pause when it matters.
- A named approver is required
- The action is irreversible or client-visible
- Evidence is incomplete or contradictory
- An evaluation enters a review band
- Novelty exceeds an operational threshold
- Sensitive data, money or privilege is involved
Part 06 / What review needs
Show the evidence, not just the request.
A reviewer should immediately see why the workflow paused, what is proposed, what supports it and which actions are available.
Approve a 10% canary or hold for remediation.
Policy: review bandTwo checks need judgment
Test coverage is complete. The change introduces a new payment-provider dependency, and rollback rehearsal finished above the preferred recovery-time threshold.
100%required tests passed
+1external dependency
14mrollback rehearsal
AUDIT / Awaiting a named decision and rationale.
Part 07 / What stays stable
Five layers keep the surface safe.
The generated surface is only the visible layer. Identity, contracts, action control, audit and recovery remain application-owned.
Stable application shell
Identity, navigation, history, global settings, help, recovery and the deterministic fallback preserve orientation.
Intent router + contracts
Select a supported job, then declare its approved components, context, state references, actions, rules, semantics and limits.
Composition + state projection
Propose and validate a surface against the contract. Resolve app-owned state while keeping ordinary interaction local.
Action + capability runtime
Map controls to declared actions. Recheck authorization, policy and evaluations at execution—not when the control is rendered.
Audit, observability + recovery
Record request, evidence, versions, decisions, action and outcome. Support retry, cancel, rollback and safe stop.
Declarative is safer only when it is restricted.
Every action must be pre-registered, authorized at execution and audited. A schema does not remove privacy, isolation or data-minimization risk.
Schema validity is not accessibility.
Use semantic primitives, deterministic fallbacks, keyboard and assistive-technology testing, responsive validation, recovery and representative users.
Part 08 / Where to start
Five workflows to test first.
Each candidate has a variable case, a named human decision and a measurable outcome. Each also keeps a deterministic fallback.
Compose the evidence for this case.
Human decision: approve, edit, reject or escalate client-visible communication.
Adapt the controls to the evidence.
Human decision: ship, hold, narrow rollout or require remediation.
Expose clauses, precedent and missing context.
Human decision: accept, request revision or escalate to legal.
Change priority with impact and SLA.
Human decision: assign, reprioritize, merge or trigger response.
Make assumptions visible and editable.
Human decision: select a scenario, accept a trade-off or request analysis.
Part 09 / How to adopt it
Prove value before you scale.
Baseline the real workflow, define authority, keep the deterministic path, then test bounded composition on production volume.
Understand the work
- Observe queues, handoffs, workarounds and decisions.
- Choose one bounded, repeatable decision.
Exit: the normal path supports the complete task and recovery.
Define authority
- Package the governed capability.
- Declare components, context, actions and limits.
Exit: every action rechecks authorization and a fallback exists.
Add composition
- Vary hierarchy and emphasis inside the contract.
- Pilot on real volume in shadow or recommendation mode.
Exit: outcome, burden, error, accessibility, latency and cost beat the comparator.
Reuse the capability
- Operate policy, evaluation, audit and rollback.
- Publish into micro-apps, APIs and agent actions.
Exit: ownership and version lifecycle work in production.
Measurement rule
Measure the whole workflow—not the novelty of the screen.
Define the unit, baseline, comparator, success threshold and stop condition before the pilot begins.
Outcome
Completion, decision quality, business result, downstream error and rework.
Flow
Time to useful outcome, total cycle time, queue time, review SLA and handoffs.
Human burden
Review rate, time per review, evidence requests, override and escalation.
Interaction
Correction, abandonment, predictability, explanation use and fallback use.
Governance
Policy/eval pass rate, unauthorized attempts, audit completeness and rollback.
Accessibility
Semantic review, keyboard and screen-reader completion, zoom, reflow and testing.
Runtime
Composition and local latency, cost, failure, schema rejection and fallback rate.
Part 10 / Concepta point of view
The interface changes. Accountability stays.
Capabilities first. Delivery forms second. Governance always.
Stable shell, bounded surface. Predictability and adaptation can coexist.
Verification beside generation. Model quality is only one part of production quality.
The next move
Choose one review-heavy workflow. Define the approved actions and the human decision. Keep the existing path. Then test whether an adaptive surface improves the production outcome.