AI on Hockystick

AI on Hockystick is an operator, not an oracle: it reads your data to answer your questions and drafts work you would otherwise do by hand — and anything that would become visible to another party stops and waits for your explicit confirmation, with no exceptions.

Philosophy

Deal flow runs on confidential information, so the AI layer is designed around containment first and capability second. Three commitments hold across every feature: the AI only sees data the requesting user could already see; it never takes a party-visible action without a confirmation click; and its scope is documented per feature — every AI page in these docs carries an explicit "does / doesn't" block.

Two agents, never one

Founders and investors talk to two separate agents with separate system prompts, separate tool sets, and separate contexts. The founder agent cannot see any investor's pipeline; the investor agent cannot see another investor's watchlist or a founder's private data. Where the two sides legitimately share context — inside a deal room — the agent works through the room's shared data and nothing else. The agents share infrastructure (the model router and rate limiting) but never context.

The confirm-first rule

Every AI-proposed action is classified by one question: would another human see a change because of this? If no — saving a draft, fetching a score, parsing a paste — it executes immediately. If yes — sending an invite, submitting a decision, approving access — the AI renders a confirmation card and waits for a click. There are no exceptions for "obviously safe" cases; the rule is structural.

AI features

Confrontational DD analysis

On investor request inside a deal room, the AI reads the actual extracted contents of every document and cross-examines them against every stated claim and metric. It is instructed not to summarize: it reports contradictions (statement vs. document), gaps (absent from both), red flags, and unverifiable claims — always citing the specific source, and explicitly stating its reasoning when it finds no contradictions. Findings include the exact question to ask and what a good answer looks like.

Where the line is

Honest update: the AI now goes further than drafting and summarizing — it critiques and cross-examines documents. What has not changed: it only reads data the requesting user could already see, every party-visible action still requires an explicit confirmation click, verdicts are conservative by instruction (it flags what it cannot verify rather than assuming it's fine), and no AI output — finding or verdict — ever auto-triggers a decision. Humans decide; the AI prepares.