insights

An assistant that can’t cite its sources is a liability

Trust in an AI assistant is not a model property. It is an interface property.

category >

ai & agents

read >

4 min

updated >

Two of the systems we run in production answer questions all day: a support assistant that took over 42% of first-line tickets, and a knowledge platform that turned hours of searching into seconds. Neither would have survived a month without one design rule: every answer shows where it came from.

Wrong answers cost more than no answers

A support agent who re-answers a question loses minutes. An assistant that invents a refund policy loses a customer — and the team’s willingness to ever trust the tool again. The failure mode of question-answering systems is not silence, it is confident fabrication, and the damage lands on whoever repeated the answer.

Citations change how people use the answer

When an answer carries its sources, the person reading it stops treating the assistant as an oracle and starts treating it as a very fast colleague: check the link when it matters, skip it when it does not. Verification becomes a one-click habit instead of a leap of faith. In practice we watched users click through on the answers that touched money or contracts, and take the rest at face value — exactly the right allocation of attention.

Knowing when not to answer

The second half of trustworthiness is refusal. When retrieval comes back thin, the honest move is to say so and hand off — our support assistant escalates to a person instead of improvising. An escalation with the conversation attached is a good experience; a made-up answer is a complaint.

What this demands from the build

Citations are not a UI garnish; they constrain the whole architecture.

  • Retrieval has to be good enough that the right passage is actually there to cite — which means investing in how documents are chunked, indexed and kept current.
  • Answers must be generated from the retrieved passages, not merely decorated with them afterwards.
  • You need an evaluation set of real questions with known sources, run on every change — a demo that works once proves nothing.

If a vendor shows you an assistant without sources on every answer, ask where the answers come from. If the demo cannot say, production will not either.

faq

Questions

A company is AI-native when AI is part of how the product works and how the work gets done — not a feature added at the end. It changes what you build and how your team operates. We work that way ourselves, which is why we can tell you what it costs.

If you already know what to build, start with software. If you do not, start with consulting. Most companies start with a two-week assessment and move straight into a build.

One call. We look at your business, product and operations, then send a short written plan with scope, timeline and price. No questionnaire, no discovery deck.

An assessment takes two weeks. A first system usually runs in production within four to eight weeks. Larger programmes run as a monthly engagement.

Companies with a real operational or product problem — funded startups through to established mid-size companies. We are based in Portugal and work remotely across Europe.

rúben martins, founder

everything you need to know before the call

Got another

question?