Golden Data

Issue #18 · September 9, 2026

The Sales Forecast Got a Second Author

Paul BrownWrites Golden Data from Ogden, Utah5 min read

On August 26, Salesforce and Anthropic announced Claudeforce, and its first shipped piece is a plugin called Salesforce in Claude: 37 prebuilt sales skills that let a rep check a deal’s health, prep for a meeting, or update the pipeline without opening the Salesforce app. Pilot customers have it now. The open beta lands in September.

Read past the demo and into the rollout mechanics, and the tell shows up in one sentence from the announcement: an admin connects it once, and every seller on the team gets access from day one. No per-user setup. No new permissions model. No re-auditing account by account.

A security lead should read that twice. Sales Ops should read it three times, because the question is not just who gets to write to the pipeline. It is what those writes do to the forecast.

The teardown: Claudeforce

Give Salesforce its due first. Per-seat OAuth grants across a floor of two hundred reps is its own failure mode — half-configured, half-revoked, a spreadsheet somebody stopped updating in March. And every action still routes back through Salesforce itself, so the validation rules, permission sets, and workflow rules that already govern a write apply before Claude’s write lands. This is not a side door into the CRM. It is a new front door built on the locks that were already there.

The locks were built for human-speed, human-judgment writes. A rep clicking a dropdown thinks about the stage. An agent batching updates across a book of business does not think about anything. It applies a pattern. Every gap in the validation rule set that a human never found at one-write-per-minute, an agent will find at sixty.

Eval surface. What is public is a skill count, not a rubric: 37 skills, no confidence scoring shown in the demo material, no disclosed human-review gate on a write action. Deal-health review and pipeline updates are exactly the outputs a forecast call already trusts on faith. The real eval question was never whether Claude reads a clean Salesforce instance correctly. It is what it does when the account has three different close dates in three different fields, a next step that has not been updated in six weeks, a stage value that does not match the activity log, and an amount that was round-numbered to get past a required field. That is most accounts more than a year old. That is the data the tool will reason over on day one, and Sales Ops is the team that knows it.

Cost exposure. Claude runs here inside Salesforce’s security boundary through Amazon Bedrock, not the first-party Anthropic API most builders benchmark against — a different pricing and rate-limit surface. A skill like deal-health review plausibly resolves through several tool calls before it answers: read the opportunity, read related activity, read the Slack thread, then synthesize. At Anthropic’s own first-party per-token rates, a multi-step reasoning turn on a frontier model runs low-to-mid cents. The aggregate across a sales floor checking deals daily is the number nobody outside Salesforce has published, and it is on infrastructure a third party prices. No pricing has been announced. If you are the person who owns the ops budget, the line item is invisible until the invoice arrives.

Data shape. The onboarding dashboard is built in one pass from Salesforce, Slack, and any connector in Claude. The weakest joint is identity and permissioning across that union — a seller’s session now spans two products’ access-control models, and the plugin’s promise is that the combination behaves exactly like Salesforce’s model alone. Nobody outside Salesforce has tested that promise at scale yet.

What I’d ship instead

Same user moment — a rep asking what’s the health of this deal before a call — scoped to a stack a Sales Ops team actually controls.

I would not start with an open-ended reasoning skill. I would start with an eval set: fifty real opportunities pulled from the CRM, chosen for messiness rather than cleanliness — pushed close dates, stale next steps, stage-activity mismatches — each hand-labeled by a manager with the health verdict a human would give. I would not ship anything until the model matched the human call on at least forty-five of them.

System shape: a scoped tool that reads only the opportunity, its stage history, and its last three logged activities, not the full account graph, and returns a structured verdict — on-track, at-risk, stalled — with the two fields that drove it named, not a paragraph. Latency budget: under three seconds, because this runs before a call, not instead of prep. Cost shape: one bounded read plus one mid-tier model call per check, cached for the day so five reps looking at the same account do not repeat the retrieval five times.

The forecast gate. Before any agent writes to the pipeline, log every AI-authored field change separately from human-authored ones. The forecast process needs to distinguish between a rep who moved a stage because they had a conversation and an agent that moved a stage because the activity pattern matched. If you cannot tell the difference at forecast time, you have a forecast built on two different kinds of judgment and no way to weight them.

Instrumentation: log every verdict against what the deal actually does in the following thirty days, and recompute the eval score monthly against real outcomes, not just the original fifty. Two weeks in, success looks like managers stop double-checking the tool’s at-risk calls by hand. Credible failure looks like the model converging on at-risk for anything without a close-this-week date — a shortcut that clears a shallow eval and fails the room.

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