Agent Library · By function · Internal Q&A · Reporting · Anomalies
Ask your own company a question.
Everything your company has decided, priced, promised and learned is written down somewhere nobody can find. A retrieval agent over your own documents is the single highest-leverage first build in most mid-market companies — and the one where sourcing every answer matters most.
The answer exists and is unfindable
Policies, precedents, past proposals and decisions live across drives, wikis and inboxes with no shared index.
The same question is asked weekly
New joiners and busy seniors both ask things the company already answered, in writing, two years ago.
Anomalies are found in hindsight
The number that broke was visible in the data for six weeks before the monthly review found it.
The agents
5 agents. Each one earns its autonomy.
Agent
What it does
Autonomy
Human gate
Internal Q&A
Answers questions from your own documents — policies, precedents, past work — with a citation on every claim.
Auto
—
Scheduled Reporter
Recurring reports assembled from source systems on a schedule, with the commentary drafted.
Notify
—
Anomaly Watcher
Monitors your key operating numbers continuously and flags genuine deviations with the likely cause attached.
Flag only
—
Decision Brief
Assembles the evidence pack behind a pending decision: what we know, what we assumed, what we decided before.
Review
—
Document Drafter
First drafts of SOPs, letters and proposals in your templates, grounded in your existing documents.
Review
—
Autonomy key — Auto: executes above 95% confidence. Notify: executes at 85–95% with notification and undo. Review: always queued for a human. Flag only: never acts, alerts a person.
What changes
Three shifts, not a feature list.
Institutional memory becomes queryable
What the company knows stops living in individual mailboxes and becomes something anyone can ask.
New joiners get productive faster
The questions they are afraid to ask a senior person get answered instantly, with sources.
Deviation gets caught in the week it happens
Not in the month-end review, by which point the cause is three weeks cold.
The eval bench
What these agents have to pass before they touch real work.
- Answer accuracy against 100 questions with verified answers drawn from your own documents
- Citation integrity: every factual claim links to a retrievable source, or the agent declines
- Adversarial set: questions whose answers do not exist in the corpus must produce "I don't know", never a plausible invention
Written against your data, not ours. The bench is built during Blueprint and runs on every deploy from then on.
Related
Where this connects.
See it on your Google Workspace.
A Blueprint maps which of these agents pays off first in your exact setup — and which to leave alone.
Apply for a Blueprint