AI Agents Overview

Waafir runs a workforce of specialist AI agents inside the data room. One summarises documents, one redacts sensitive information, one produces review reports, one proposes a folder structure, and one translates. Together they handle the repetitive reading and tidying that would otherwise consume a deal team's time.

This page is the model for the entire AI surface. Read it first; the individual feature pages cover each agent in detail.

The agents

Every organisation starts with the same five core agents configured. No setup is required before they operate; they are ready as soon as your data room exists.

You can adjust how each agent behaves — for example, refining its instructions or the model it uses — but these five core agents are always present and cannot be removed. They define the baseline of platform capability: you tune them, you do not lose them.

The roster does not stop at five. Some capabilities ship as agents that come with an add-on rather than the base plan — for example, the Investor Update Generator that drafts periodic investor reports, part of the Investor Materials add-on. It runs on the same framework and follows the same review and audit rules. AI-assisted pitch decks are coming soon to the same add-on. See Investor materials and reports for that surface.

Your agent settings also list two more built-in agents: the Prose Drafter, which runs only from the API (see Agents over the API); and the Fact Extractor, which Waafir uses internally and you do not run.

Custom agents that you define yourself are a preview and do not run yet; see Custom AI Agents for what the agent settings let you do today.

Automatic versus on-demand

The split in the table above is deliberate and is the most important distinction to understand.

• Read-only work runs automatically. Summarising a document does not change it and carries no risk, so it runs as soon as a file is uploaded, with no action required. Summaries appear on your documents shortly after processing completes.
• Work that changes a document, or what an investor sees, is on-demand and reviewed. Redaction, organisation, and translation never apply silently. You trigger the agent, it produces a proposal, and you confirm it before it takes effect.

This is the human-in-the-loop principle, and it is not optional. The platform is built so that the AI cannot make a destructive or investor-visible change without human approval.

A typical agent run

When you trigger an on-demand agent — redacting a document, for example — the flow is always the same:

1. You start the action from the document.
2. The agent performs its reasoning and produces a proposed result.
3. The platform presents that proposal in a review screen — a redaction diff, a folder preview, a language toggle.
4. You confirm or reject it.
5. The change is applied only on your confirmation.

If you reject the proposal, nothing changes. Each feature page documents the exact review screen for that agent.

When each agent runs

Each agent runs from its own product action: the Summariser when a file is uploaded, the others when you start them from a document or data room. The one exception is the Prose Drafter, which only an API call runs (below). You cannot set any agent to run on a schedule, or choose a different trigger for it. Whatever starts an agent, the review step is unchanged: any change that is destructive or investor-visible is a proposal a human approves.

Agents over the API

The Platform API can run one agent directly: POST /api/v1/agents/draft_prose/invoke runs the Prose Drafter, which writes prose grounded only in citations you send. It changes nothing in your data rooms; it returns text.

• Request: { "citations": [{ "id": string, "text": string }], "instruction": string }. Up to 50 citations, each up to 8,000 characters and 60,000 in total; the instruction is up to 2,000 characters. The draft attributes its claims to your citation ids and says so when the citations do not support an answer.
• Response: { agent, runId, suggestionId, content, usage: { inputTokens, outputTokens } }.
• Access: a personal access token with write scope, and AI features enabled for your organisation (on by default; an organisation admin can turn them off in Settings → AI & Integrations). Limited to 5 calls per minute and counted against your organisation's daily AI budget.
• Errors: draft_prose is the only agent name accepted; any other returns 404. If an admin has switched the Prose Drafter off in agent settings, the call returns 409. The full list is in the endpoints reference.

The run is recorded in the audit trail with counts only (how many citations, how long the instruction), not the citation text or the draft.

Audit trail

Every AI action is recorded: which agent ran, on which document, what it proposed, and who approved or rejected it. This audit trail is central to why the platform is safe to use with sensitive deal documents — there is always an answer to "what did the AI do, and who signed off on it?" When you need to demonstrate control over how a data room was prepared, this record is the source.

Next steps

• Custom AI Agents — tune the built-in agents' prompts and models; custom agents are a preview and do not run yet.
• Redaction — the Redactor agent in detail.
• Translation — the Translator agent in detail.
• Organise — the Organiser agent in detail.
• Readiness scoring — how the platform assesses whether a data room is ready for investors.
• AI Assistant — the conversational surface, which is not one of these agents.