An answer you can check

Everyone working near AI has now met the failure mode: an answer that is fluent, specific, well-structured, and wrong. In casual use it costs a laugh and a retry. Inside a company it compounds — a salesperson repeats the claim on a call, a marketer lifts it into the launch page — and by the time anyone checks, it has your logo on it and has spent the one thing you took years to build: the assumption that when your company says something about your product, it’s true.

The awkward part is that fluency was the easy piece. Large models produce confidence natively; what they know sits underneath, unevenly. Anyone putting an AI in front of product questions has to close the gap between how sure the system sounds and how sure anyone can actually be — and the industry has been closing it, in stages, for a few years now.

Where grounding got to

The first stage was retrieval: instead of asking the model what it remembers, fetch the relevant documents and ask it to answer from those. This helped enormously, and it is now table stakes — most serious product assistants are retrieval systems. The second stage was attribution: append sources, so the reader can see which documents fed the answer. Footnotes with links, a “sources” row under the reply. Also genuinely good, and increasingly standard.

Sit with attribution for a moment, though, and two gaps show up. The first is granularity. A citation that points at a forty-page document has told you where to start reading, and not much else. Checking the answer means doing the retrieval again yourself, manually, which is exactly the work the assistant was supposed to remove. In practice nobody clicks, which means the citation functions as reassurance rather than as evidence — a trust ornament.

The second gap is the corpus. Most assistants are pointed at everything the company has ever written: the wiki, the tickets, the drive folder nobody has opened since 2023. When the corpus is unbounded, nobody can say in advance what the system might assert, because nobody has read the corpus — and parts of it now contradict each other. The citations can be perfectly accurate while the cited material is perfectly stale.

The line-level bar

We hold Deputy to a bar we started calling line-level: every answer cites the specific line it came from, and clicking the citation opens the briefing scrolled to that exact line. Checking an answer takes one click and about two seconds. The difference between “source: the documentation” and “source: this sentence, updated May 12” turns out to be the difference between trust as a vibe and trust as a receipt.

The other half of the bar constrains the corpus rather than the citation: if a claim isn’t in the briefing, it doesn’t get said. Deputy speaks only from a briefing you edit and approve — a document deliberately small enough to read end to end. That gives the person accountable for the answers a guarantee that no unbounded corpus can offer: you can read, in advance, every word it will ever say. Bounding the corpus is what makes line-level citation meaningful — a receipt only reassures if the ledger it points into is one you actually vouch for.

A side effect we didn’t fully anticipate: the bar disciplines the writers, meaning us. When every answer must trace to a line, a capability claim can’t float in an old slide deck or in somebody’s head — it has to exist as a reviewable sentence before the system will repeat it. A wrong answer stops being a mystery to debug and becomes a line to fix, and fixing the line fixes every future answer that would have leaned on it.

What the bar costs

Bounded corpora decline more. Ask a Deputy something outside its briefing and it will tell you it doesn’t know and go ask the owner — the question waits, marked pending, instead of receiving a confident guess. Early on, that happens noticeably more often than it would with a general assistant, and a team optimizing for answer-rate would call it a regression. We think the trade runs the other way: a system that visibly declines what it can’t support is the only kind whose answers mean something when it doesn’t decline. The declines also aren’t dead ends — each one comes back as a new briefing line, which is its own story.

None of this is the finish line for grounding — the industry keeps moving, and the bar should keep rising. But if you are evaluating any AI that will speak about your product, ours included, the two questions worth asking first: when it cites, can I check the claim in one click — and can anyone on my team read everything it’s allowed to say?