This guide is for product owners, designers and engineering leads adding a generative answer, draft or suggestion to a business product. It sits under our product engineering service, and the scoping work it describes usually starts in product discovery and roadmapping.
In this guide
- Why does an AI feature need its own interface patterns?
- Which patterns signal uncertainty?
- Which patterns show where an answer came from?
- Which patterns let users correct or reject output?
- Which patterns does each risk tier need?
- How should a team specify these patterns before build?
- Where AI fits
- What delivery record exists, and what does not
- Alternatives: other ways to handle AI uncertainty
- What are the limits of this guide?
- Frequently asked questions
- Next step
Why does an AI feature need its own interface patterns?
A conventional feature either works or shows an error. An AI feature can return a fluent, well-formatted answer that is wrong, and the interface gives no hint. The Nielsen Norman Group's guidance on AI hallucinations puts it plainly: trust depends on acknowledging the system's limits, not on hiding them.
NIST's generative AI profile lists confabulation, confidently stated false content, among the risks specific to generative AI. The model cannot remove that risk alone, so the interface has to carry part of it: by telling users how much to rely on an output, by letting them check it, and by making a wrong output easy to fix.
The failure to avoid is uniform presentation. When a verified fact and a guess look identical on screen, users either trust everything or trust nothing, and both outcomes defeat the feature.
Which patterns signal uncertainty?
- Categorical confidence. Show High, Medium or Low instead of a percentage, with a defined action for each band. Google PAIR's chapter on explainability and trust warns that numeric confidence assumes users understand probability. Skip it when the model's scores are not calibrated against real outcomes.
- Hedged wording. Let the answer itself say "I could not confirm this" for weak evidence instead of a generic footer disclaimer. Skip it for structured outputs such as extracted fields, where a flag works better.
- Field-level flags. In a form filled by document AI, mark the individual fields the model was unsure of, so reviewers check three fields, not thirty.
- Several options. Offer two or three candidate drafts or classifications so users apply their own judgement. Skip it when choosing between options costs more than editing one.
- Designed refusal. When the evidence is missing, say so and offer a next step (search, a form, a person) instead of guessing.
Which patterns show where an answer came from?
- Claim-level citations. Attach each source to the sentence it supports, not as a list at the end. A list of links looks like proof but tells the user nothing about which claim each link backs.
- Passage preview. Clicking a citation opens the exact passage, highlighted, so checking takes seconds. This matters most in a knowledge assistant built on retrieval, as covered in our guide to RAG architecture for knowledge assistants.
- Source metadata. Show the document's owner and last update date next to the citation, because a correct quote from an outdated policy is still a wrong answer.
- Grounded versus inferred. Mark content the model inferred or summarised differently from content quoted from a source. Skip it in casual drafting tools, where everything is a suggestion anyway.
Which patterns let users correct or reject output?
- Edit in place. The output is editable where it appears, and the edited version is what gets saved or sent.
- One-step dismiss and undo. Rejecting a suggestion is as fast as accepting it, and an accepted action can be reversed.
- Reasoned feedback. A thumbs-down opens a short choice of reasons (inaccurate, out of date, wrong format) that the team can act on. PAIR's chapter on errors and graceful failure recommends feedback both on errors and alongside correct output.
- Review before action. When the AI acts on the user's behalf, such as sending a message or updating a record, a confirmation step shows what will happen. Skip it for reversible, low-impact actions, where it only adds friction.
Which patterns does each risk tier need?
| Risk tier | Example feature | Uncertainty and source patterns | Correction patterns |
|---|---|---|---|
| Low: a wrong output costs a few seconds | Draft reply, tag suggestion | Hedged wording; citations optional | Edit in place, one-step dismiss |
| Medium: a wrong output reaches a customer or a record | Support answer, extracted invoice fields | Field-level flags, designed refusal, claim-level citations with passage preview | Reasoned feedback, undo |
| High: a wrong output carries money, safety or legal effect | Policy answer, automated update to an order | Categorical confidence with defined actions, source metadata, grounded versus inferred | Review before action, logged corrections |
The rule behind the table: add a pattern only when the cost of a wrong output justifies the friction it adds. A tag suggestion with a confirmation dialog will be switched off by its users; a policy answer without sources will be trusted when it should not be.
How should a team specify these patterns before build?
-
List the failure modes
For each AI output, write down how it can be wrong: invented fact, outdated source, wrong field, wrong tone.
-
Assign a risk tier
Use the cost of the worst plausible wrong output, not the average one.
-
Pick the minimum pattern set
Take the row of the table above and remove anything the feature does not need.
-
Define the confidence bands
Agree what High, Medium and Low mean in model terms and what the user should do in each.
-
Design the empty and refusal states
Most mockups show only the good answer; draw the "no evidence" screen too.
-
Prototype with real wrong answers
Test with outputs you know are wrong and watch whether users catch them; our guide to usability testing without a research team covers small-sample rounds.
-
Wire corrections to evaluation
Route edits, dismissals and reasoned feedback into the test set the engineers use before each model or prompt change.
-
Review after launch
Track how often users edit or reject output; heavy editing of accepted output means the feature is failing quietly.
Where AI fits
AI tools help with the specification work itself: generating batches of plausible wrong answers for prototype tests, drafting failure-mode lists from sample outputs, and grouping user feedback by reason. Designers still decide which patterns ship and engineers still check every test case, because a test set written by the same kind of model it tests can share its blind spots. Netbase works with the major commercial and open-source AI tools and models, chosen per project, under human control.
- In delivered work
RAG knowledge assistant
A retrieval-augmented assistant that answers from an organization's own documents; the client is not named.
- In delivered work
Recommendation engine
A recommendation engine delivered for 4over4's online printing store, suggesting products from browsing history and previous purchases.
What delivery record exists, and what does not
Netbase has delivered AI projects whose interfaces face these questions: retrieval-augmented knowledge assistants, document AI and MLOps pipelines for clients that are not named, and an AI chatbot for WhatsApp Business connected to a client's CRM, also for an unnamed client. The named AI feature in the portfolio is the 4over4 recommendation engine.
Netbase delivery teams combine business analysts, project managers, solution architects, developers, QA and UI/UX designers, and discovery and strategic alignment is the first step of the delivery lifecycle, which is where the risk tier and pattern set are agreed. Netbase works with commercial and open-source models chosen per project, so the patterns here do not depend on one vendor. Most Netbase projects are delivered on fixed-price contracts.
What does not exist: Netbase publishes no usability, trust or correction-rate metric for any of these interfaces, and this guide claims none. The pattern library is a design method, not a measured result.
Alternatives: other ways to handle AI uncertainty
- Keep a human between the model and the user. Staff review every output before anyone sees it; choose this when volumes are low and errors are costly. Our guide to agentic AI with human-in-the-loop controls covers where that checkpoint sits.
- Constrain the output. Replace free text with choices from a fixed list or a template; choose this when the answer space is small.
- Do not ship the feature yet. If the team cannot name the failure modes or build a test set, the interface cannot compensate; choose discovery first.
- Commission the design. Ask a delivery team to run the eight steps above; our UI/UX design service describes how that work is scoped. If the AI feature arrives as part of a wider interface overhaul, see redesign or iterate a legacy interface.
What are the limits of this guide?
- It covers text answers, drafts, extracted fields and suggestions; image and voice outputs need patterns of their own.
- The patterns reduce the cost of wrong output; they do not make a weak model or poor retrieval accurate.
- It does not replace an accessibility review of the new states and controls.
Plan the next step with a Netbase consultant
Frequently asked questions
Usually not. Most users cannot tell whether a given percentage is good or bad, and model scores are often not calibrated. Categories such as High, Medium and Low with a defined action for each are easier to act on, and they only work if the bands are tested against real outcomes.
No. Citations make wrong answers easier to catch, but only when each source is tied to the claim it supports and opens the exact passage. A list of links at the end of an answer looks like evidence and is rarely checked.
Corrections help only when the team uses them. Route edits, dismissals and the reasons users give into the test set that runs before each model or prompt change, and review edit rates after launch to find outputs that users accept and then rewrite.
When the action is hard to reverse or reaches a customer, a record or a payment. For reversible, low-impact actions, a one-step undo protects users without slowing every interaction down.
Next step
If you are adding an AI answer, draft or suggestion to your product, tell us about the feature and its users and we will reply with a proposal for the risk review and the interface work that follows.
Related services and solutions
AI-enabled software product engineering from discovery to operation Netbase provides software product engineering for companies that need a web, SaaS, mobile or custom platform built and kept running, with AI features and AI-assisted engineering where they pay off. We cover the full lifecycle, from discovery to operation, and the same team stays accountable for architecture, delivery and handover, so nothing gets lost between phases.
Learn More
Generative AI development with evaluation and human review Netbase builds generative AI features for product and operations teams: drafting, summarising, rewriting and content generation inside the software your people and customers already use. Every feature ships with an evaluation set, a human review step where output matters and limits on cost. Generative AI is a capability we are growing, and we label its maturity honestly.
Learn More
UI/UX design services: research, interfaces and design systems built in code Netbase designs the interfaces of web apps, SaaS products, mobile apps and online stores: user research, flows, wireframes, clickable prototypes and visual design, plus a design system shared by designers and developers. Because UI/UX designers sit inside the delivery team, the same people who design a screen see it built, tested and released.
Learn More
AI knowledge assistant: answers from company knowledge, with sources and access control An AI knowledge assistant is an internal tool that answers employees' questions from your company's own documents and systems, cites the source behind every answer and shows each person only what they may see. Netbase offers it as a growth capability, designed with retrieval, evaluation and human review from the first pilot, and has delivered one for an unnamed client.
Learn More
Work with Netbase remotely
Every Netbase engagement is delivered remotely, so clients in any country and any industry can work with us; Netbase has worked with clients in 80 countries. Contracts and invoices are issued by Netbase JSC, and payment is made by bank transfer to Netbase JSC's company account or PayPal Business. Talk to the CEO, Huy Nguyen (David), on WhatsApp or by email.
Discuss a project
Netbase JSC helps organizations design, build, modernize, and operate digital products and AI-enabled business systems.project@netbasejsc.com
+84 937 869 689
91 Nguyen Chi Thanh, Dong Da, Hanoi, Vietnam
Get in touch
Tell us what you want to build, modernize, or operate.