An AI assistant might seem "simple" to launch—a chat box, a model, some docs. In enterprise, this is rarely the case. The best assistants win not because they "answer well," but because they **integrate into real work**, make their limits visible, and trigger reliable actions.
An AI assistant might seem “simple” to launch—a chat box, a model, some docs. In enterprise, this is rarely the case. The best assistants and chatbots don’t win because they “answer well,” but because they integrate into real work, make their limits visible, and trigger reliable actions (without creating new risks).
This AI design guide covers UX specific to assistants and chatbots in a professional context, with concrete patterns for designing trust, action, accessibility, and measurement.
Traditional UX designs deterministic paths: screens, forms, rules. With an LLM-based assistant, the interface becomes probabilistic:
The user formulates their need in natural language (often vague).
The system must interpret the intent, choose a strategy, and generate a response.
Quality varies according to context, available data, and phrasing.
Consequence: in AI design, the UX mission is not just to “make the interface pleasant.” It is to reduce uncertainty and make the product “steerable”: expectations, guardrails, error recovery, traceability, and KPIs.
If you want to lay the UX/UI foundations regarding vocabulary and methods, the Impulse Lab glossary sheet on UX/UI is a good refresher.
Many projects fail from the start due to a format error: deploying a “generalist” chatbot when the need is work assistance within a tool, or vice versa.
Here is a simple grid, useful for product scoping.
Format | Where it lives | What it's for | Dominant UX | Main risk |
|---|---|---|---|---|
Chatbot (web, support) | Site, helpdesk, WhatsApp, etc. | Answering, guiding, qualifying | Intent coverage, escalation | “Polite but useless answers” |
Internal Assistant (knowledge) | Intranet, Slack/Teams, portal | Finding info, summarizing, helping decide | Trust, sources, context | Hallucinations + obsolete info |
Tooled Copilot (actions) | In the business tool (CRM, ERP) | Executing tasks via tools (tool-calling) | Confirmation, control, audit | Incorrect actions, costs, security |
Semi-autonomous Agent | Orchestrator + tools | Chaining steps (workflow) | Governance, supervision | Silent errors + drift |
To clarify product terms, you can also reread the glossary definition of AI agent and chatbot.
Before opening Figma, your best UX lever is a very concrete intent framing.
A pro assistant must produce an actionable output: a sourced answer, a draft, an action in a tool, a decision with options, or a transfer to a human. If the output is vague, the UX will be too.
Describe, for 5 to 10 real scenarios:
The trigger (e.g., “client requests a refund”)
The available context (data, history, knowledge base)
The expected result (e.g., “refund policy + next step + internal link”)
The guardrail (e.g., “if amount > X, manager validation”)
In AI design, an explicit contract reduces frustration:
What the assistant knows how to do (short list)
What it doesn't do (and why)
What it does in case of doubt (question, source, escalation)
This contract must be visible: onboarding, field placeholder, examples, and contextual “help”.
In enterprise, an assistant's UX is primarily a UX of trust. Without trust, low adoption. With unjustified trust, high risk.
Indicate the level of certainty: “I'm not sure,” “I didn't find this in your internal sources,” “Here is what procedure X says.”
Display sources when possible (excerpts, links, update date). RAG (retrieval-augmented generation) type assistants are made for this. See the definition of RAG.
Trace the perimeter: “I answer based on your Support base (internal articles), not based on your CRM.”
Offer an alternative: “I can create a ticket” or “I can escalate to an agent.”
The user often assumes that “the bot remembers.” In enterprise, this is precisely where problems arise.
Good AI design:
explain what is memorized (session, preferences, history)
make memory editable (delete, correct)
avoid “fake personalization” (which gives the impression of surveillance)
On compliance, keep a simple principle: minimize data and make usage explicit. The CNIL regularly publishes useful recommendations on personal data and compliance.
An assistant that “talks” but does nothing quickly creates friction. Conversely, an assistant that acts without control is dangerous.
The good UX compromise for many SMEs and scale-ups: chat + guided actions.
Structured responses: instead of a block of text, use cards (summary, steps, fields).
Action buttons: “Create a ticket,” “Pre-fill an email,” “Open client file,” “Generate a summary.”
Preview before execution: the user validates what will be sent or modified.
Explicit confirmation for any irreversible action.
Assistants can repeat an action if latency is long, if the user retries, or if the orchestration “retries.” Even without getting into architecture, UX must provide:
a clear status (“action in progress,” “completed,” “failed”)
an action history
protection against duplicates (“already sent 2 min ago”)
This becomes even more important when the assistant is integrated with multiple tools. Impulse Lab works precisely on integration with existing tools and automation, so that AI doesn't remain an isolated screen.
In AI design, errors are not a marginal case; they are part of the product.
Type of failure | User-side symptom | Expected UX response | Example guardrail |
|---|---|---|---|
Lack of info | Vague answer | Ask a targeted question, offer options | “Which product are we talking about?” |
Uncertainty / contradiction | Incoherent answers | Display doubt, cite sources, ask for validation | “Procedure A and B diverge, which do you wish to apply?” |
Critical error | Risky action | Block, escalate, require strong confirmation | “High amount, validation required” |
Avoid generic errors (“An error occurred”). A good assistant must say:
what happened (without jargon)
what it can do now
what the user can do
if a human is required
Conversational interfaces seem “simple,” but they can become difficult to use: keyboard focus, screen readers, contrasts, dynamic states, timing.
Accessible AI design reflexes:
make states readable (loading, result, error)
guarantee keyboard navigation (buttons, suggestions)
avoid “push” messages breaking reading flow
offer alternatives to long text (summaries, steps)
For a broader checklist (WCAG, testing tools, best practices), the Impulse Lab sheet on web accessibility is a practical base.
Testing an assistant like a classic UI (static mockups) is not enough. You must test conversations, results, and actions.
Loop 1 (qualitative): user tests on 10 to 20 scenarios. Objective: understand language, blockers, need for control.
Loop 2 (replicable evaluations): build an “evaluation set” (golden set) of questions and expected answers, to measure evolution.
Loop 3 (production): instrumentation, error tracking, and weekly iteration.
Impulse Lab is used to delivering in short cycles (weekly cadence). This cadence is particularly adapted to AI design, because UX and quality are built through iterations and measurements.
Stage | UX Artifact | Recommended test | Success signal |
|---|---|---|---|
Scoping | Scenarios, assistant contract | Business review + risks | “Frequent” scenarios validated |
Prototype | Scripts + conversational mockups | Guided user tests | The user reaches the expected output |
MVP | Assistant connected to sources | Golden set + situational tests | Stability, sources, low rate of “unjustified” escalation |
Pilot | Limited deployment | Adoption + quality + cost KPIs | Measurable gains, errors under control |
To properly scope prompts and behaviors, the sheet on prompt engineering can help, but keep in mind that a good assistant is not “saved” by the prompt. It is designed through UX + data + integrations.
As you scale, the assistant must not depend on “the person who knows how to talk to the model.” You must standardize.
Typical elements of a conversational design system:
Tone and style (level of formality, short sentences, forbidden terms)
Response templates (summary, steps, next action)
UI states (loading, streaming, error, action in progress)
Components (sources, citations, buttons, forms)
Escalation rules (when to switch to human)
If your designers work in Figma, the Impulse Lab glossary on Figma can be useful for structuring components and versions with the dev team.
Without measurement, you will have endless debates on “it answers well.” With measurement, you steer.
Some often actionable UX metrics (to adapt to the case):
Resolution rate (without human) vs escalation rate
Time to output (e.g., actionable answer, created action)
Reformulation rate (the user repeats their question)
Correction rate (the user modifies the draft)
Post-interaction satisfaction (simple, 1 question)
For an ROI and instrumentation-oriented approach, you can also read the Impulse Lab article on AI chatbot KPIs.
If you are an SME or a scale-up, your advantage is not to aim for “a universal assistant.” It is to quickly deliver a useful assistant on 1 to 2 frequent paths, then industrialize: reliable sources, actionable integrations, trusted UX, and measurement.
At Impulse Lab, the typical approach for this type of topic combines:
AI opportunity audit (scoping, risks, prioritization)
Custom design and development (web + AI)
Integration with your tools
Training for adoption at the point of usage
If you want to secure an assistant (or revamp a chatbot already in place) with a UX oriented towards action, compliance, and ROI, you can start with a conversation via Impulse Lab.
Notre équipe d'experts vous répond rapidement pour comprendre vos besoins et vous proposer la meilleure solution.
Vous avez des questions ? On a les réponses.
Leonard
Co-fondateur
Continuez votre lecture avec ces articles
Vous avez probablement déjà vécu la scène : quelqu’un poste un prompt “miracle” sur Slack, un autre teste un outil gratuit “juste pour voir”, puis, deux semaines plus tard, vous ne savez plus qui utilise quoi, avec quelles données, ni si les résultats sont fiables.
Un prototype d’agent IA peut impressionner en 48 heures, puis se révéler inutilisable dès qu’il touche des données réelles, des utilisateurs pressés, ou des outils métiers imparfaits. En PME, le passage à la production n’est pas une question de “meilleur modèle”, c’est une question de **cadrage, d’i...