Design the relationship between people, models and product state
Tcules helps teams create AI-native products, add AI to established software and redesign workflows where probabilistic output changes how people decide or act.
- AI-native products
- AI Product UX
- Product state
An AI feature spans several interaction patterns
AI Product UX is the wider product relationship: how people understand, direct, evaluate, control and recover from model behaviour.
Before choosing chat, copilot or agent, define the role the model plays.
The article on generative AI patterns for SaaS compares common product roles before a team settles the interaction.
- generate a candidate;
- transform supplied material;
- retrieve and synthesise evidence;
- recommend;
- predict or detect;
- orchestrate steps;
- act with delegated authority.
Each role introduces different uncertainty and control. A drafting assistant needs editing and source constraints. A recommendation needs basis and alternatives. An action needs permission, preview, state, stopping and recovery.
The AI product responsibility record
| Field | Product decision |
|---|---|
User work | Which decision or task deserves assistance? |
Model role | What may the model produce or influence? |
Evidence | What data and sources support the output? |
Uncertainty | What can be missing, wrong or unstable? |
Authority | What can happen without human approval? |
Evaluation | What constitutes useful and safe performance here? |
Recovery | How can a person correct or reverse the result? |
This record follows the product into design and engineering. It prevents an interface label from carrying more certainty than the system has earned.
A consequential AI scenario is enough to define what the model may do, what a person can inspect and how recovery must work.
How is an AI feature evaluated before release?
Evaluation begins with representative user decisions and failure conditions, not model accuracy alone. The acceptance set can test source coverage, useful output, missing evidence, appropriate human review, action scope, correction and recovery.
Technical and operational checks follow the agreed delivery scope. Tcules makes the evidence reviewable; client owners retain approval of domain, policy and residual risk where those authorities belong to them.
How product responsibility combines
- opportunity framing and product boundary;
- research with target users and domain experts;
- AI interaction and workflow design;
- source, confidence and explanation patterns;
- human review, control and recovery;
- evaluation scenarios and acceptance;
- coded prototypes and product engineering;
- integration into existing product roles, state and design systems.
Tcules has worked on 12 AI products across copilots, AI-native tools, AI-powered products and integrations. The cases below show several distinct product situations inside that experience.
Evidence across different AI conditions
Evidence across different AI conditions
- ConnectX
An intent-led workspace bounded through eight product scenarios and demonstrated in code.
- BuildTwin
Evidence, correction and escalation for an engineer reviewing AI quality-control results.
- Eden AI
Workflow creation and search for a developer-facing AI platform.
- Novus
A survey-builder product situation; public responsibility remains bounded by available source.
AI-assisted delivery is a separate claim
Tcules also uses AI in its own research synthesis, exploration, prototyping, coding and QA where appropriate. Generated work is reviewed against the product model, source evidence, implementation constraints and client requirements. Client AI restrictions override delivery defaults.
Authority changes with consequence
Informing, suggesting, preparing and executing are different product roles. For each one, define the consequence, reversibility, evidence a person can inspect, the human gate and the recovery path.
A system that can act across several steps needs those decisions per action, not one blanket label for the feature.
Define the first AI decision
Useful inputs are the user work, the current product state and the authority you expect AI to receive. If that scope is still uncertain, the AI Product UX Readiness Check is a free self-assessment, and an AI Product UX Readiness assessment can create the first evidence.
Use AI Copilots and Agents when the product is delegating a bounded task. Use Human-in-the-Loop Workflows when review, escalation and correction are central to the work. Trust, Control and Recovery addresses consequential failure and oversight. AI Search and Recommendation focuses on retrieval, ranking and a person’s ability to inspect the basis for a result.