How Endgame works.

Endgame turns a sales team's records — CRM, calls, emails, Slack — into cited facts, linked to the accounts, people, and deals they concern, before any question is asked. This is that mechanism, stage by stage, with the evaluation dataset as the running example.

Companion to Endgame vs Data 360Dataset Northwind (synthetic)Both systems identical data

01 / 04

01 The hard part

The facts live in the conversations. The fields drift.

The renewal date was in a customer's email. The economic buyer identified herself on a call. The last real customer contact was an email a CRM rollup had papered over. In each case the structured field held a stale or misleading value, and the truth sat in unstructured text.

A query surface over centralized records leaves gaps for the model to close at question time: rediscover the schema, choose the right tables, read raw text, and link what it finds — in seconds, on every question, with no memory of last time. Our evaluations, shared earlier in “Endgame vs Data 360,” showed what doing it live looks like: a stale field reported as the answer, a five-minute research run to find one date, and answers that varied from run to run.

“Data 360 successfully centralized the evidence. It did not turn that evidence into sales context.”

Codex — an OpenAI model, reviewing the evaluation results

02 / 04

02 The mechanism

Four stages, all before the question.

Unify the record

Connectors bring CRM objects and the engagement record — calls, emails, Slack — into one store. In the evaluation dataset: 308 accounts, 1,395 contacts, 535 opportunities, 2,399 engagements, identical on both sides.

Resolve it into a graph

Endgame resolves every engagement to the accounts, people, and opportunities it involves — including engagements that arrive with no CRM identifiers, the way real calls and emails do. The evaluation org resolved to 2,575 entities connected by 4,153 relationships.

Extract facts, each with a citation

An extraction pass reads every transcript, email, and thread and produces discrete facts — 17,822 in the evaluation org, each anchored to the exact source record it came from, and every one carrying at least one link into the graph. A fact in an answer traces to a specific email, call, or message.

Maintain perspectives with context agents

Scheduled agents read the graph and keep account perspectives current — a headline briefing and the account's active work streams, refreshed every half hour across all 215 accounts with open pipeline. The synthesis runs once, on a schedule, from the extracted facts — a specific claim checks against the facts behind it, and every user who asks gets the same answer.

At question time there is no archaeology left to do. Endgame's tools state what they contain, the graph carries the links, and the answer — with its citations — comes back in one turn. In the evaluation org, the production API returned 90 ranked facts for one account in 218 ms.

03 / 04

03 The evaluation, revisited

Each question turned on which layer did the linking.

Every evaluation question had a correct answer in the data and a wrong answer sitting in the obvious field.

Q1 · renewal date

The CRM close date said Sep 30. The customer's June 10 email corrected it to Aug 31.

Endgame had extracted the correction as a fact linked to the account — every model tier answered Aug 31 in one turn and flagged the stale field.

Q2 · last contact

The activity rollup read June 2026. The last real inbound was an email of April 3, 2025 — a year of silence.

Endgame's graph links each engagement to the account with its date and direction — the silence was directly visible.

Q3 · economic buyer

The contact roles omitted the buyer entirely. On a May 28 call she said: “I own the platform engineering budget.”

Endgame had extracted that statement as a cited fact — every tier named her and flagged the missing role.

04 / 04

04 Beyond three questions

The structure is what scales.

At the evaluation's scale, a capable model can compensate for missing structure by reading many records. At the scale of a real book of business, reading everything on every question stops being possible — the linking has to exist before the question. That precomputed structure is also what makes three properties hold:

Cited

Every fact in an answer names its source record — a claim can be checked against the email, call, or message it came from.

Consistent

The same question draws on the same extracted facts and agent perspectives — for every user, every session.

One turn

Answers arrive without a discovery phase, at every model tier — including the small, fast ones.

All figures are from the evaluation org described in “Endgame vs Data 360”; both systems saw identical data.