Golden Data

Issue #19 · September 16, 2026

The CRM stopped waiting for the rep.

Paul BrownWrites Golden Data from Ogden, Utah6 min read

On August 19 Salesforce published the Winter ’27 release notes, and the Pipeline Forecasting page grew a set of columns. Deal risks. Methodology scores. Contact insights. Activity heatmaps. And one called agent activity: what the software did on a deal, next to what the rep did. The grid I defend on Mondays is about to show the machine’s work as a peer of the human’s.

On February 18 Salesforce signed to buy Momentum, whose whole product is listening to calls, emails, and meetings and writing the result into Salesforce fields. The deal closed March 2. Salesforce’s May 11 release announcement explained why in one sentence: “Most conversation data never makes it into Salesforce, leaving deal signals incomplete and agents working blind.”

That sentence came from a company that has sold activity capture for years. Einstein Activity Capture syncs the rep’s email and calendar, and stores the captured email on AWS, outside the org. You cannot report on it or query it, and it deletes itself after six months, or twenty-four on the paid tier. The CRM had the emails the whole time. It kept them in a room its own forecast could not enter.

So the vendor bought a company to write down what its own product had been hearing. The native version is aimed at general availability at Dreamforce, which opened yesterday in San Francisco. Here is what is public, and what I would build first.

What’s actually shipping this week

Salesforce, August 19. Pipeline Forecasting adds deal risks, methodology scores, contact insights, agent activity, and activity heatmaps inside the forecast view. Call Coaching gives managers up to eight competencies and, for Agentforce for Sales users, an agent that scores every recorded call against them. Each is a machine writing what a rep used to type, or not type.

Salesforce, August 26. Q2 FY27: revenue $11.3 billion, up 11%. Agentforce ARR above $1.5 billion, up more than 240%. The footnote: “Effective Q2 FY27, Agentforce ARR includes our AI offerings, Slackbot and Headless 360.” Slackbot reached a million users five months after launch. The number got a wider definition in the quarter it got a bigger headline. Miguel Milano counted 2,000 new paying customers in production, up 70% on the quarter. Marc Benioff, on the call: “If your data’s not right, your AI is not right.”

Validity, August 25. The State of CRM Data Management in 2026 surveyed 500 marketing professionals across five countries. It is neither a seller survey nor Salesforce’s, which is why it is here. 21% say their CRM data is very well prepared for the AI they use or plan to use. 62% report losing revenue directly to poor CRM data quality. The agents are being handed the pen before the data is ready for it.

Clari + Salesloft, July 14. The independents are writing too. Conversation Intelligence went generally available with call summaries, action items, follow-up drafts, and “automatic CRM updates.” The independents mostly land on the Task. Salesforce’s capture lands on the Opportunity. In a forecast that is the whole difference.

HubSpot, September 4 and 8. The other CRM moved the opposite way. A rebuilt CRM index put Breeze “AI insights on columns” across Contacts, Companies, Deals, and Tasks, and four days later the 2026-09 API began enforcing admin-configured required properties on every write. One vendor widened who can write to the record. The other tightened what a write has to contain. UNBOUND opens in Boston on September 16, Dreamforce’s day two.

What I’d build in the revenue org this week

One build: a ledger of who wrote each opportunity field. Not the forecast. The provenance under the forecast. It serves the Monday call, when a manager asks why a Commit deal moved and the honest answer is a transcript.

Data shape: Salesforce OpportunityFieldHistory is the source: one row per field change with CreatedById, Field, OldValue, NewValue, CreatedDate. Join Task and Event on WhatId for the activity behind the write. Track what the capture writes: NextStep and the MEDDPICC-style custom fields. In BigQuery: sfdc.opportunity_field_history, sfdc.task, sfdc.event, sfdc.user, and a derived rev.field_writes, one row per field write, with a writer_kind column: rep, capture, agent, flow.

System shape: A nightly BigQuery job. writer_kind comes from the User row: integration users are capture or agent, the automated process user is flow, everyone else is a rep. A Looker explore on top with writer_kind as a dimension. No model in the path. The model already ran upstream, inside the vendor.

Latency budget: Hourly replication. The forecast is weekly. Momentum’s real-time write is the vendor’s latency, not mine.

Cost shape: Field history for a mid-size org is a few million rows a quarter. BigQuery cost is cents. Three engineer-days. The expensive line is the Momentum license, sold inside Agentforce 1 Sales Edition or as an add-on, priced by negotiation.

Eval shape: A backtest against last quarter. Take every opportunity in Commit at week six. Label it closed or slipped. For each, count writes by writer_kind in the prior 30 days. Did machine-written fields predict slippage better than rep-written ones? Lift over the base slip rate is the score. If they did, the forecast should weight them. If not, the capture is a note-taker, not a signal.

Instrumentation: Three weekly counters. Share of field writes by writer_kind. Share of Commit deals with no machine write in 14 days. Overwrite rate: a machine writes a field and a rep changes it inside 48 hours. The third is the disagreement rate the vendor does not publish.

Two weeks in, success is a Monday call where every Commit deal shows which fields a human last touched, an overwrite rate under 10%, and a backtest with lift. The credible failure is the first query: the capture writes as the rep’s own user, not an integration user, and writer_kind collapses to rep. Then your own field history cannot tell the machine from the person. That is also a finding. It goes in the renewal conversation.

Sources

Send me an email and we will talk. If something here landed close to what you're working on, the door is open. No calendar funnel, no pitch deck — I read every note that comes in.

Doing the work rather than deciding what to build? Crafting is the column for that chair.

Email meRead more Signal

← All Signal issues · Drafted with Claude · Edited by Paul Brown