AI Agent Guide¶

Audience: Users · Version 0.2 (draft) · Updated 2026-06-11

The AI Agent is an in-app assistant that answers questions about companies using Capital Workbench's own curated data — the same numbers you see on the pages. It does not browse the web and it never invents figures.

Open it with the AI Agent button in the header. It appears when you're signed in and chat is enabled for your account.

Always verify before you act

AI-generated answers can be incomplete or wrong. Treat the assistant as a research aid, not a decision-maker — check any figure against the page or the source filing before you rely on it.

This guide replaces the older cw_ai_agent_doc.md, ai-tools-and-capabilities.md, and ai-chat-prompt-examples.md for everyday use. (The architecture/governance notes in cw_ai_agent_doc.md remain for developers.)

What it can do¶

The assistant works by calling a fixed set of read-only tools over CW data. In plain terms, it can:

Find a company and switch your selected company.
Summarize a company — a latest-period snapshot of the key KPIs, plus recent growth and profitability trends.
Show KPI history — quarter-by-quarter values for a metric (margins, DSO, cash ratios, etc.), with a link to the page that shows it.
Explain data quality — why a metric is missing or flagged, in plain language.
Look up filings — filing metadata and links (not full filing text).
Compare against peers — a sector-based peer comparison on curated KPIs.
Compare against an industry — where a company sits within its industry cohort (percentile and distribution).
Pull raw statement lines — specific balance-sheet / income / cash-flow line items for one or several companies.
Compute statement ratios — ratios not shown as KPI tiles (e.g. CapEx/Revenue, interest coverage, effective tax rate).
Propose a report — when you ask it to build/generate a report, it prepares a confirmation card. You click Generate — it never creates the report on its own.

What it won't do¶

Invent or estimate numbers. If the data isn't there, it says so.
Give investment advice, price targets, or buy/sell recommendations.
Swap one metric for another. It uses exact labels — e.g. EBITDA margin is not operating margin; free cash flow is not operating cash flow.
Read full filing text or run arbitrary database queries.

How it behaves¶

Single-company questions: it sets your active company first, then answers — so the page context follows the conversation.
Comparisons ("compare A vs B"): it does not change your active company.
Citations: it quotes the period, the basis/source, and units, and links to the relevant page when one exists.
Data-quality questions: it checks the quality notes and quotes the reason verbatim rather than guessing.
Clarifying questions: if your request is ambiguous, it asks one short question.

Reports from chat¶

Ask the agent to "create / generate / build" a report. It validates the request and shows a proposal card. The report is generated only after you click Generate report; in production this runs as a background job and then appears in My reports (/account/reports). The assistant never claims a report exists or invents a link before you generate it.

Report types it can propose include single-company snapshots (liquidity, performance, diligence-memo-lite, working-capital diagnostic), two-company comparisons (working-capital, profitability), and small industry-cohort snapshots.

Chat history¶

Your conversations are saved per account. In the AI drawer:

Chat history (clock icon) opens the list of your past conversations.
Click any conversation to reopen it and continue.
Use the delete control on a card to remove a conversation (with a confirm step).
The conversation you're currently in is excluded from the list.
New chat starts a fresh conversation.

Limits & quotas¶

Limit	Default	Notes
Per-hour messages	30	A short-term rate cap; you'll see a "per-hour limit" message if exceeded.
Tools per answer	8	The assistant calls at most a handful of tools to compose one answer.
Conversation memory	last 20 messages	Older turns drop out of the model's context.

Monthly plan quotas:

Plan	Messages / month	Tokens / month
Free	30	60,000
Analyst	3,000	6,000,000
Enterprise	Unlimited	Unlimited

When a monthly quota is reached, the input shows "Quota reached — upgrade to keep asking" and you'll be prompted to upgrade. Quotas reset at the start of each month.

Admins: these defaults are configurable in Admin → AI settings (per-hour cap, tool budget, history length, timeout). Setting the per-hour cap to 0 disables it. Monthly quotas are plan-based and enforced before the model is called.

When something goes wrong¶

The assistant currently uses a single model provider with no automatic fallback. If the provider can't be reached, you'll see a "provider request failed" message — retry in a moment. Other messages you might see: chat disabled, rate limit reached (per-hour), or quota reached (monthly).

Privacy¶

Your messages and the assistant's replies are stored with your account so chat history works and usage can be measured against your quota.
New chat clears the current view in your browser and starts a new conversation; to remove a past conversation from your history, use the delete control.
Administrators can review chat logs for support and abuse-prevention purposes.

Retention policy: TBD — to be finalized as a product decision and reflected on the privacy page.