UX design for AI features in B2B SaaS.
The model works and the usage chart is flat. We design the part of an AI feature the user sees: where it sits in the product, how it shows what it is sure of, and what happens when it is wrong.
Copilots, assistants, and agents in B2B products · Prague and Dubai · 4.9 on Clutch
Problems we see
Five patterns from AI features inside B2B products. If two of them sound familiar, start with the audit.
- 01 The copilot is ignored It lives in a side panel behind a sparkle icon. The user already has a way to do the task, and the panel asks them to learn a second one. They close it once and never open it again.
- 02 Nobody trusts the answer The output arrives with no source, no confidence, and no way to check it. A finance or legal user cannot act on that, so they redo the work by hand and the feature turns into a suggestion box.
- 03 The empty state asks for a prompt A blank field that says “Ask anything” is where most sessions end. People do not know what the model can do, so they type something small and leave when the answer is generic.
- 04 Every model error looks the same A wrong answer, a timeout, a refusal, and a hallucinated field each need their own state and their own next step. When they share one red banner, the user learns that the feature is unreliable, not that one request failed.
- 05 Onboarding into AI is a tour A tooltip walkthrough on first login explains the feature to someone who has no task yet. The moment to teach it is when the user is doing the thing the feature is for, inside the flow, with their own data.
How we solve it
Four steps. The first one is a product of its own.
Audit the feature you have
The AI feature UX audit, 1–2 weeks. We read the usage data you have, watch sessions, and map where people find the feature and where they leave. We also look at what they ask it. Adoption and completion become numbers your team can argue about.
Redraw the screens that decide it
The entry point, the answer, the failure. We redesign those first, in your components, so engineering can ship them in a normal sprint.
Prototype and test with your users
A clickable prototype with output that looks real, put in front of 6 of your customers. Whether they trust the answer shows in whether they act on it.
Ship and measure
Your team ships the screens. We compare the numbers from step 1 with the numbers after release and pick the next feature. That continues as a sprint or a retainer, whichever fits the roadmap.
Examples
Three products with an AI feature we worked on. We walk through the screens on a call.
Cluely
A real-time AI meeting assistant from a startup backed by a16z, where the AI is the whole product. We designed the UX. The live suggestion panel, the meeting summary, and the search across past calls.
Neptune
Prenups for the modern couple, with an AI feature inside the flow. The guided questionnaire and the shared draft view.
Afternoon
An AI-first product for financial advisors in the UK, built that way from the start. We did the UX. The assistant that drafts client reviews, and the adviser’s review queue.
What we do differently on AI features
Where to start
Two doors: one if the feature exists, one if it does not.
AI feature UX audit
For a copilot, search, chat, or agent that is already live. Adoption and completion numbers, the screens where users leave, 2–3 redrawn screens, and a ranked list for the next sprint. $9,000, credited against the sprint that follows.
Product design sprint
For a feature that is still a slide. Research, flows, the answer and failure states, a prototype your users can try, and handover to engineering. From $24,000.
Join hundreds of teams. And counting.
Questions?
Can you redesign an AI feature we have already shipped?
Yes. Most teams come to us with the feature already live, so we begin with the AI feature UX audit: 1–2 weeks on the feature as it is, reading the usage you have and watching people use it. The result is a ranked list of what to change and 2–3 redrawn screens. If you go on to a sprint with us, the audit fee is credited against it.
We have no usage data for the AI feature. Can you still start?
Yes. Missing data is itself a finding: it usually means the feature was never instrumented for adoption and completion. We agree on the events to track in the first week and run sessions with your users in the meantime, so the first numbers are there by the time the redrawn screens are.
Do you need access to the model, the prompts, or the code?
No. We work on the interface between the model and the user: entry points, states, wording, and what a person can do with the output. We do need a working environment or recordings of the feature in use, and a conversation with whoever owns the prompts, because the wording on screen and the wording in the prompt have to agree.
Which format fits an AI feature, and what does each one cost?
The audit is 1–2 weeks at $9,000. A redesign sprint on the screens the audit picks is 2–4 weeks at $24,000. A retainer for continuing AI feature work is $14,000 a month. Every price is confirmed on the call, after we have seen the product.
What to read
Three pieces on AI in products from our blog, and the audit.
Last updated
