Service 01 / 04 · AI assistants, agents

AI assistant design that plans for the wrong answer.

We design AI assistants, voice apps and agents that do tasks, for your customers and staff. Each shows its sources and passes hard questions to a person.

Start an AI design project See what we make

Typical length: 6–12 weeks to a test version
Works in: Chat · voice · camera · agents
Standard: Tested on your questions

Illustration of an assistant chat. A user asks which enterprise contracts renew before March, and whether any pricing terms changed. The assistant shows a read-only plan of 3 steps. It answers with 3 numbered sources. It says 2 scanned contracts could not be read. Then it offers to draft reminders to 3 account owners and waits for approval.

You

Which enterprise contracts renew before March, and did any pricing terms change?

Plan · 3 stepsRead-only

contracts.searchrenews < 31 Mar · enterprise14 found
clauses.readpricing, indexation × 1412 read
clauses.diffagainst the previous term3 changed

Assistant

14 contracts renew before 31 March. Three changed pricing terms: two now rise with inflation ¹ ² and one removed a volume discount ³.

1MSA-2291 · §7.2
2MSA-2310 · §7.1
3MSA-2187 · Sch. B

Partial2 contracts are scanned images and were not read.

Proposed action · needs approval

Draft renewal reminders to 3 account owners

Edit planApprove

What we make

6 parts of an AI product we design and test.

They set what the AI does alone, what it asks first and what it shows. People then know when to trust it.

01 / 06

Conversation design

The first message, questions, refusals and recovery, written as a script and tested with users before any AI is connected.

Scripted · tested
02 / 06

Agent plans and approvals

Agents are AI assistants that do tasks. People see the plan, approve each step and can undo it later.

Plan · approve · undo
03 / 06

Voice and camera input

Voice, camera, document and screen input designed as one, with a switch to typing in noise or poor light.

Voice · camera · document
04 / 06

Sources and confidence

Sources, confidence, edit history and a record of what the AI used, so people check before they rely on it.

Sourced · checkable
05 / 06

Delays, errors and backups

Streaming text, part results and clear messages for the 4-second wait and the wrong answer.

Streaming · backups
06 / 06

User testing

Users try tasks on fixed test questions, odd cases included. We score success, effort to fix and trust.

Task success · trust

Plan and timing

From first idea to build plan in 4 stages.

We build a test version on a working AI early. A static picture cannot show how AI answers behave. Timings are typical, set per project.

01 Wk 01–02
Define

We agree the users, their tasks, the data the AI may see and what it may do. We list risks before features.
- Experience plan
- Task & risk map
- Test questions v1
02 Wk 02–06
Design

Flows, conversation scripts and screens, built with your design kit and tried on a working AI.
- Screens and flows
- Conversation scripts
- Test version
03 Wk 05–09
Test

Users try tasks from the test questions, including ones the AI gets wrong. We redesign around what they do.
- User test findings
- Revised flows
- Test questions v2
04 Wk 09–12
Hand over

Your engineers get the specification, screen parts and test questions, plus the checks that say when it is done.
- Design specification
- Screen parts
- Sign-off checks

How it works

A list of what can go wrong, written and tested before launch.

For each failure we list the cause, what people see, how they recover and the test for it.

Failure states · Support assistant5 of 23 states

Failure states · Support assistant: illustrative sample
State	Cause	What the person sees	Recovery	Tested
No source found	Search score too low	Says so, offers search	One tap to a person	41 / 41
Low confidence	Quality score < 0.7	Answer with warning + sources	Ask a follow-up	38 / 40
Action needs approval	Refund over limit	Plan shown, approve or edit	Named approver	25 / 25
Hidden instruction	Instruction in a document	Ignores it, flags the file	Logged for review	18 / 20
Out of scope	Legal or medical question	Declines, explains why	Handover with context	9 / 12

States designed: 23All before launch
Test pass rate: 94%Target 95% to launch
Handovers with context: 100%No repeat questions

Illustrative Screens marked Illustrative are examples, not client results.

What we measure

3 measures of trust, taken on the test version.

We measure before design starts, then on the test version. Figures shown are typical. Targets are set per project.

01

Errors people can correct

People can spot a bad answer, correct it and carry on with their task.

Recovery after a wrong answer

Before35%
Target80%+

Measured by Guided sessions on test questions

02

People stay in control

People see the plan, approve big actions and can undo anything the AI did.

Agent actions approved first time

Before50%
Target85%+

Measured by Approvals in the test version log

03

Proof before the build

We test each design with users on a working AI, so the build starts from what worked.

Task success on test questions

Before58%
Target85%+

Measured by Tasks completed ÷ tasks attempted

What you get

The plan your engineers build from, in your own files.

Scripts, test questions, the failure list and the tested version sit in your code and design files as we go.

A hand sketches a flow diagram on paper, the pen resting on a yes-or-no branch — **Every branch drawn**the wrong answer included

File list · AI Apps7 items

AI Application Design items and their formats
#	Item	Format
01	Experience plan, task & risk map	Doc · board
02	Screen and flow design	Figma
03	Scripts & AI instructions	Doc · test questions
04	Test version on a working AI	Demo · code
05	AI screen parts library	Figma · Storybook
06	User testing report	Report · recordings
07	Design specification & sign-off checks	Doc · tickets

Who does what

Design comes after strategy and guides the build.

AI Application Design covers what people see and use. Its sign-off checks become the tests your engineers build to.

AI Design AI Strategy & Consulting
Which job AI should do first, and why.
AI Design · this page AI Application Design
The conversation, controls and error states, tested on a working AI.
Technology & Intelligence AI engineering
Agents, document search and servers, built to the specification.
Both Live testing
The same test questions, re-run each release.

The test questions written in week 1 are used at all 4 stages.

Tools we use

The design tools and models we test ideas with.

Design, AI and test-build tools, most used first. We pick models by testing them on your questions.

Figma
Anthropic
Google Gemini
Mistral AI
LangGraph
TypeScript
React
Next.js
Storybook
Vercel
PostHog

Also in use

OpenAI
LlamaIndex
Playwright

Standards we build to

Public standards your reviewers already know.

These standards shape the checks, error states and records we hand over. We build to them. We hold no certificates for them.

WCAG 2.2 AAWeb Content Accessibility Guidelines

Rules that make websites usable by people with disabilities. Version 2.2 adds visible focus, tap size and easier log-ins.
EU AI ActArtificial Intelligence Act (EU) 2024/1689

Europe's AI law sets rules by risk: banned uses, duties for high-risk AI and general-purpose models, and AI labels.
NIST AI RMF 1.0AI Risk Management Framework

Managing AI risk in 4 parts: Govern, Map, Measure and Manage. A second guide covers generative AI (NIST AI 600-1).
OWASP Top 10 for LLM ApplicationsSecurity risks in generative AI applications

Risks in AI apps, such as hidden instructions that trick the AI, data leaks and AI with too much power.
GDPRGeneral Data Protection Regulation (EU) 2016/679

Europe's data law: a legal reason to use data, people's rights, privacy by design, breach reports and risk checks.
DPDP Act 2023Digital Personal Data Protection Act, 2023

India's data law and its Rules: notice and consent, people's rights, breach reports and extra duties for significant firms.

Offers & plans

AI assistant design offers, bought alone or together.

Buy the design of one assistant, agent or voice app, or of the whole product. We test every design with users on a working AI.

Categories: 04
Offers: 14
Plans: 04

Not sure what you need? Tell us the problem

How to buy

01Pick offers. Ask about one, or add several to one enquiry.
02Choose a plan. A small job, a fixed project or ongoing help.
03Send your enquiry. We reply within one working day.

Browse by category

Timelines are typical. We write the scope before we quote.

01Ideas and concepts

3 offers

Typical timeline: 2–3 weeks

AI feature check-up

We review a live AI feature: where people give up, where trust drops and what to fix first.

What’s included

Review of sessions and chat logs
Expert review against AI interface patterns
Interviews with people who gave up
Ranked list of fixes with effort

First step
Check-up
Ranked fixes

PostHog
Figma

Best forTeams with a live AI feature whose usage is falling.

Typical timeline: 2 weeks

AI concepts in 2 weeks

We turn an idea into 3 concepts, each demoed on a working AI, so you choose on how they behave.

What’s included

Kick-off session with your team
3 concepts designed and demoed
Cost, risk and build notes on each
Recommendation and next steps

Sprint
Demos
Decision

Figma
Anthropic
OpenAI

Best forLeadership teams choosing between several AI ideas.

Typical timeline: 3–5 weeks

AI product definition

What the product does and will not do, where a person approves and how the work runs behind the screens.

What’s included

Users, tasks and what can fail
What it can and cannot do
Service map including the human steps
Success measures and sign-off checks

Definition
Service map
Limits

Miro
Figma

Best forTeams whose AI product scope changes in every meeting.

02Interaction design

4 offers

Typical timeline: 3–6 weeks

Conversation design

The assistant’s opening, questions, refusals and recovery, written as a script and tested before any AI is connected.

What’s included

Conversation scripts for the main tasks
Main AI instructions and tone
Wording for refusals, fixes and handovers
Script testing with users

Scripts
Tone
Refusals

Anthropic
OpenAI
Figma

Best forAssistants and chat products where the writing is the interface.

Typical timeline: 4–8 weeks

AI agent design

How an agent’s plan is shown, paused, edited and approved. People see what it will do first, and can undo it.

What’s included

Plan preview, step by step
Approvals, permissions and spending limits
Pause, retry and undo states
Record of every agent action

Agents
Approval
Undo

Figma
LangGraph
Anthropic

Best forProducts where AI acts for a customer or colleague.

Typical timeline: 4–8 weeks

Voice and camera design

Speech, camera and document input designed alongside typing, with steps to fix anything misheard or misread.

What’s included

Voice and camera interaction design
Steps to confirm, pause and correct
Typing option for every spoken step
Noise, low light and blur states

Voice
Camera
Corrections

Figma
OpenAI
Google Gemini

Best forProducts used hands-free, in the field or on a phone camera.

Typical timeline: 3–5 weeks

Sources and control design

Screen parts that help people judge an answer: sources, what the AI used, its doubts and how to fix it.

What’s included

Ways to show and open sources
Showing doubt and confidence
Ways to fix, rate or overrule
Clear AI labels on every screen

Sources
Confidence
Overrule

Figma

Best forProducts for decisions where a wrong answer is costly.

03Systems & build

3 offers

Typical timeline: 6–10 weeks

Shared AI design kit

Reusable screen parts for AI: streaming text, sources, plan steps, approvals and error states. Designed once, documented in code.

What’s included

Parts with shared colours and fonts
Streaming, loading and part-answer states
Accessibility rules for each part
Documentation in Storybook

Shared kit
Screen parts
Documented

Figma
Storybook
React
TypeScript

Best forOrganisations adding AI to several products and repeating the same work.

Typical timeline: 4–8 weeks

Working test version

A test version on a working AI with your content. Try it with users or show it to a board or investors.

What’s included

Your data or a safe sample
Document search or lookups if needed
Test questions covering awkward cases
Test sessions and a findings report

Test version
Working AI
Testable

Next.js
TypeScript
LlamaIndex
Anthropic
Vercel

Best forTeams that need proof before they fund the build.

Typical timeline: Ongoing, by sprint

Design support through the build

Designers work with your engineers through the build. They answer what the plan cannot and check each release against it.

What’s included

Designer in your planning and reviews
Plan updates as the AI changes
Design checks against the agreed plan
Accessibility checks before release

In your team
Design checks
Your pace

Figma
Linear
Jira

Best forEngineering teams building AI without an AI designer.

04Test & improve

4 offers

Typical timeline: 3–5 weeks

Testing with users

Users try tasks on fixed test questions, wrong answers included. We score success, effort to fix and trust.

What’s included

Test questions for the awkward cases
Guided user sessions
Scores for success, effort and trust
Ranked findings with design changes

Ease of use
Task success
Trust

PostHog

Best forTeams about to launch or to invest more in an unproven feature.

Typical timeline: 2–4 weeks

Answer quality review

We review what your assistant says: tone, accuracy, refusals and answers that hurt trust. Then we rewrite its instructions.

What’s included

Chat samples checked against your tone
Main instructions rewritten and versioned
Refusal and safety wording review
Before-and-after test on fixed questions

Chat logs
Tone
Rewrite

Anthropic
OpenAI
Python

Best forLive assistants whose answers no longer sound like the brand.

Typical timeline: 2–4 weeks

Accessibility review for AI

A WCAG 2.2 AA review of what general audits miss: streaming text, agent actions, voice input and changing content.

What’s included

Screen-reader tests of streaming text
Keyboard control of every agent action
A typed option for voice input
Findings with fixes and retest

WCAG 2.2 AA
Screen readers
Keyboard

Playwright
Figma

Best forPublic-sector, regulated and enterprise products that must be accessible.

Typical timeline: Monthly, ongoing

Monitoring after launch

A monthly review of how the assistant is doing: dropped chats, corrections, handovers and rejected answers.

What’s included

Tracking of AI events
Monthly review of drop-offs and fixes
Test questions refreshed from live chats
Ranked list of design fixes

Monthly
Tracked
Fix list

PostHog
Mixpanel
OpenTelemetry

Best forProducts where the AI feature is now a main part of the service.

Your enquiry

Tick “Add to enquiry” on any offer, choose a plan, then continue. Or enquire about one offer.

Plan

How we work with you

Ways to start, from a question to a quote.

Ask a quick question, send your project details or request a quote. The person who will lead the work reads each one in full. Services in your enquiry go with it.

Or book a 30-minute call

What are you sending?

01

About 2 minutes3 short answers

For a first chat, a press request or anything not ready to plan yet.

You get A reply from a named lead
02Recommended

About 8 minutes5 short steps

Goals, who it is for, budget and timing, so our first reply can outline the work.

You get Options and prices after one call
03

About 15 minutesYour documents attached

Your documents, deadlines and the purchasing and security rules to meet.

You get Confirmation and a named lead

How it is priced

Each plan shows how it is priced. Work starts once we agree a written scope and price.

Typical length
1–3 weeks

Pricing
Fixed fee
Typical length
4–12 weeks

Pricing
Fixed price
Typical length
3–9 months

Pricing
Fixed price per stage
Typical length
Ongoing · 6-month minimum

Pricing
Monthly fee

Compare what each plan includes

What each plan includes and who it suits
Plan	What you get	Best for
SprintA short, fixed job that answers one clear question.	Scope and result agreed before we start A senior lead and the right specialists A weekly review with you A clear answer or a test version	An audit, a brand check or a test version of an app
ProjectOne clear job for one fixed price.	A written scope: what you get and how we agree it is done A named project lead and a fixed team A shared plan with checkpoint dates Source files, yours once paid for	A brand identity, a website, a simple app or a set of tools
MilestoneA bigger build in stages you approve and pay for one at a time.	Stages, each with its own scope and approval A go or stop review at each stage Re-plan between stages as you learn Pay for each stage you approve	Work too big for one contract, such as a large app
RetainerA monthly team that runs and improves what we built.	Team time set aside for you every month Agreed response times for requests and fixes A monthly review and a to-do list Planned improvements, not only fixes	Monthly social media and ads, or care for live sites and apps

Questions

Your questions on designing AI assistants.

What buyers ask before AI assistant design work starts.

Book a call

How is this different from Technology & Intelligence?

We design what people see and do: the conversation, controls and answers. Technology & Intelligence builds and runs the AI behind it. On larger projects, the 2 teams work together.

Do you design for one model or several?

Several. The design does not depend on one model. We test on those that suit the task, such as Gemini, OpenAI, Anthropic, Mistral and open models. You can switch later without a redesign.

How do you design for wrong answers?

We treat a wrong answer as normal. Sources and confidence are always shown. A fix is one tap away, big steps need approval and you can always reach a person.

Can our engineers build it from your designs?

Yes. They get a specification, screen parts and test questions. What we make for you is yours once it is paid for. Tools we already had stay ours, and you get a free, permanent licence to use them.

Can an AI interface be accessible?

Yes. We design to WCAG 2.2 AA, so people with disabilities can use it. Screen readers read text as it appears. Every action works by keyboard, and every voice step can be typed.

Also in AI Design