Leadership brief · one page

Grounded Answers From Documents

Every organization owns thousands of documents whose contents nobody can query. This training builds a question-answering capability over exactly those documents - and is honest about what its answers prove, which is the part most such projects skip.

Self-paced · 6 modules, 22 lessons · about 7 hours

Chapter 6 in the Meridian sequence · 9 live so far — the sequence follows one fictional utility through the same modernization, so the examples build on each other, and this one picks up the story from Building Trustworthy Data Products and Zero Trust Implementation. Each training stands on its own; the order is the recommended path, not a prerequisite — the examples name the same fictional utility, but nothing from the earlier trainings is needed to follow them. The story continues in AI Assurance: System Risk and Release Decisions.

What this training covers

Six modules in the order the work actually happens: deciding whether a question needs this at all, against four cheaper ways to answer it; governing the corpus - what is authoritative, what is superseded, who may see what - before the system reads a page; building retrieval whose every answer cites a passage the reader can open; proving answers against their cited sources, including the questions the system must refuse; authorizing who can ask what, since a synthesized answer can exceed any passage it cites; and re-testing everything whenever any dependency changes. Every module teaches the work and the AI-assisted way to do it as one workflow, and every module ends by finding what the agent got wrong.

Why it matters

Fluent, confident, wrong answers are the specific product failure of this technology, and they are worse than no answer because somebody acts on them. The defensible version of this capability makes a modest claim: a grounded answer converts an unbounded question - is this true? - into a bounded one - does this cited passage say this, and is the passage current and authoritative? That bounded check is one a domain expert can perform in the time it takes to read the cited page — a minute or two, not a research task. Everything in this training exists to keep the capability inside that claim: evaluation before build, citations that open, abstention when the corpus cannot answer, and a human check that automation screens for but never replaces.

What changes in practice

Tags name what each shift affects most: calendar time, cost, contract risk, or an audit finding avoided.

1. 1

   The evaluation questions are written before anything is built

   Calendar time

   Why it matters: a fixed question set with known answers and a stated pass bar is what separates a capability from a demo, and it is cheapest before anyone is attached to the build · Module 1

2. 2

   Four ways to answer a question without building anything get tried first

   Cost

   Why it matters: reading the document, a search box, a query against a governed table, or admitting the question needs human judgment answers most candidates cheaper - the build is reserved for what survives all four exits · Module 1

3. 3

   The corpus is governed before the system reads a page

   Audit finding avoided

   Why it matters: which documents are authoritative, which are superseded, and who may see what get decided by people - retrieval only enforces decisions that were actually made · Module 2

4. 4

   Every answer cites a passage a reader can open

   Calendar time

   Why it matters: the check on an answer becomes reading the cited page instead of an unbounded research task - and answers that cannot cite, abstain · Modules 3-4

5. 5

   Support is checked by hand on a sample, every evaluation, forever

   Contract risk

   Why it matters: a citation proves where text came from, not that it says what the answer claims - and automated grounding checks screen answers, they do not certify them · Module 4

6. 6

   Answers are authorized at the answer, not just the source

   Audit finding avoided

   Why it matters: a synthesized answer can aggregate innocuous passages into something no single document was cleared to say, so who-can-ask-what is a design decision with a named owner · Module 5

7. 7

   Any changed dependency re-runs the evaluations

   Contract risk

   Why it matters: a corpus update, a model change, or an edited prompt silently invalidates yesterday's results - the regression habit catches it before a consumer acts on it · Module 6

Where the effort goes

Most of it lands in the corpus and the evaluation, not the retrieval. Deciding what the system may read, extracting documents that were never meant to be machine-read, and writing the question set with its pass bar is the real work; the retrieval build itself is short once an agent is doing the typing. What this does not require is a machine-learning team or a new vendor contract - the practice build runs on your existing AI CLI plus utilities already on a stock machine, with no new account and no admin request. What it does require is funded time: a named owner for the corpus, the sampled hand check as a standing cost, and the discipline to let "not in the corpus" stand as a correct answer.

How you'll know it worked

• A frozen question set with a pass bar exists before the build — including questions the system must refuse to answer
• Every served answer carries a citation that opens, and a sampled hand check runs on every evaluation
• Zero answers served from a corpus, model, or prompt version that has not been re-evaluated since it changed