Sales Forecasting with AI: A Weekly Claude Agent That Builds the Bottoms-Up Forecast Brief

Write the evidence checklist: what a commit has to prove

Start from your own slipped deals. Pull the last two quarters of deals that were committed and did not close in the quarter, and for each, write down what was missing in the CRM the Monday before it slipped. The list is short and it repeats: no next step, or a next step dated in the past; a champion named in the notes but absent from the last three calls; no legal or procurement contact anywhere; the last activity was ours, eleven days ago, unanswered. Those absences are your checklist.

Write it as four checks a machine can run and a rep can argue with. Next step: a field with a date in the future and a verb in it. Champion: a contact tagged champion who attended a call in the last 30 days. Paper process: a contact from legal, security, or procurement, or a call mention of contract, redline, or vendor form. Activity: an inbound touch from the buyer within ten days. The checklist below is the one I start from, and the thresholds are what every team tunes against its own slips.

Decide what the checks mean. A committed deal with all four is a commit and the brief says so in one line. A committed deal missing one gets a flag naming what is missing. A committed deal missing two or more is a stage-evidence disagreement, and the brief puts it at the top of the deals to argue about. The agent never moves the deal. It says what it sees, and the manager and the rep decide on the call.

Connect the CRM and the call recorder, and snapshot last week

Connect HubSpot or Salesforce over MCP and confirm the agent can read every open deal with a close date in the quarter: amount, stage, forecast category, close date, next step and its date, owner, associated contacts with roles, and the activity log. The MCP directory lists what each connector exposes. Read only; the play needs no write access, and giving it none is the simplest way to keep the forecast a human decision.

Connect Gong so the agent can read the last two calls per deal. This is where the buyer's own words live: 'we are targeting a signature before the 30th', 'legal has it', 'I need to loop in my VP before we commit to a date'. The CRM says Negotiation; the call says the VP has not been met. The brief quotes both.

Keep two small files beside the run. Quota by rep for the quarter, as a CSV, so the roll-up can show coverage and gap. And last week's snapshot, which the run writes at the end of each week: deal, stage, amount, close date, forecast category. The Monday run compares the live pipeline to the snapshot and that comparison is the 'what moved' section, pulled in, slipped, closed, lost, new, without anyone maintaining a change log.

• Reads: open deals closing this quarter: amount, stage, forecast category, close date, next step (text + date), owner, contacts (role), activity log
• Reads: last two calls per deal: attendees, date, transcript
• Reads: files/quota.csv (rep, quarter quota) and snapshots/last-week.csv
• Writes: the forecast brief, one per-rep section, this week's snapshot, Slack posts. Never a CRM field

Write the brief so every flag cites its evidence

The brief is the checklist turned into a run, plus the rule that makes the forecast call shorter: every flag names the missing evidence and the source, and every quote names the call and the speaker. 'Commit, no next step since Aug 14 (CRM next_step_date), last buyer activity Aug 19 (email), champion J. Ortiz absent from Aug 28 and Sep 3 calls (Gong)' is a line a rep can check and a manager can act on. 'This deal feels soft' is the sentence the forecast call already had too many of.

Make the agent conservative about interpretation. A deal with a next step dated next week and a champion who missed one call is fine and should say so in one line. A deal that has all four checks but a buyer quote about 'the budget cycle in January' is a timing flag, and the brief should put the quote next to the close date and let the manager read it. The agent's job is to put the CRM field and the buyer's sentence in the same line. The judgment stays with the people on the call.

The roll-up is arithmetic and the agent should show it. Commit by rep, best case by rep, weighted pipeline as a third column nobody is asked to believe, quota, gap to quota from commit, and coverage (open pipeline over remaining quota). Then last week's commit beside this week's, with the deals that moved it listed by name and amount. When the manager picks the number on the call, the brief has already shown what the number would have to assume.

Every Monday at 6:00am {{TIMEZONE}}, build the forecast call brief for {{QUARTER}}.

Read from {{CRM}}: every open deal with close date in {{QUARTER}}. For each: amount, stage, forecast category (commit / best case / pipeline), close date, next step text and date, owner, contacts (name, title, role), activity log for the last 30 days.
Read files/quota.csv (rep, quota) and snapshots/last-week.csv.
For each deal in commit or best case, read the last two calls via {{CALL_RECORDER}} (attendees, date, transcript).

Run the evidence checklist (checklist/forecast-evidence.md) on every commit and best-case deal:
- NEXT STEP: next_step_date is in the future and the text contains an action. Cite the field.
- CHAMPION: a contact tagged champion attended a call in the last 30 days. Cite the call date.
- PAPER: a legal / security / procurement contact exists, or a call mentions contract, redline, MSA, vendor form, security review. Cite the source.
- ACTIVITY: an inbound buyer touch (email, meeting held, reply) within 10 days. Cite the date.
A check passes only with cited evidence. Quote the buyer's words on timing or budget where the calls have them (call date, speaker).

Flag disagreements:
- commit missing 2+ checks, or best case with all 4 checks
- close date within 14 days and PAPER fails
- stage Negotiation with no proposal or pricing activity in 21 days
- close date moved 2+ times this quarter

Write brief/{{DATE}}.md:
1. Roll-up by rep: commit, best case, pipeline, quota, gap from commit, coverage. Team totals. Last week's commit beside this week's.
2. What moved since snapshots/last-week.csv: closed won, closed lost, pulled in, slipped (with old and new date), new to commit, dropped from commit. Deal, owner, amount.
3. Deals to argue about, largest first: each flagged deal in 3 lines: the claim (stage, amount, date), the evidence (checks passed and failed with sources), the buyer's words if any.
4. Clean commits: one line each, so the call can skip them.
5. Per-rep summary: one line, commit vs quota, count of flags.

Write snapshots/{{DATE}}.csv for next week. Post section 1 and 3 to #{{FORECAST_CHANNEL}}; post each rep their own deals (sections 3-4 filtered) as a DM.

Rules:
- Read only. Change no CRM field, stage, date, or category. Never state a forecast number of your own; show the roll-up and the evidence.
- Fields and quotes, never adjectives. 'No next step since Aug 14' beats 'the deal is stalling'.
- If a source is unavailable, say which section is missing and why.

Run it Monday at 6am and rebuild the forecast call around it

Schedule the job as a scheduled Claude Code run for Monday at 6am, so the brief is in the forecast channel before anyone opens a spreadsheet. Each rep gets a DM with their own flagged deals and clean commits, which means the rep walks into the call having already fixed the next-step date that was blank, or ready to explain why the champion missed two calls. Half the flags resolve before the meeting starts, and that is the point.

Change the agenda to match the brief. Five minutes on the roll-up: where commit sits against quota, what moved. Forty minutes on the deals to argue about, largest first, each with its evidence on screen. Ten minutes on the number: the manager picks commit and best case for the team, on the record, with the brief as the paper trail. The first two weeks the call will still drift to 'how do you feel about Meridian'. Point at the evidence line and ask the question the evidence raises instead.

Run the first brief by hand with the sales manager and read every flag against their read of the deal. Where the agent flagged a deal the manager knows is solid, look at the evidence: usually the champion is real but untagged in the CRM, or the paper process lives in an email thread the agent cannot see. Both are CRM hygiene, and the hygiene play fixes them at the source. Where the manager names a deal they are worried about that passed clean, find the check the list lacks.

Track the roll every week: slips, pull-ins, and the number you called

The 'what moved' section is the manager's view of the quarter as a film instead of a photo. A deal that slipped from Sep 12 to Sep 26 to Oct 3 is a pattern the brief shows in one line, with the three dates, and a close date that has moved twice is a flag on its own. Pull-ins get the same treatment: a deal that jumped from pipeline to commit in a week should carry the evidence that justified it, and if it does not, the brief says which check it skipped.

Keep the number the manager called each week beside the roll-up. The brief records it: week 6, commit called at $1.9M against a CRM commit of $2.05M, with the two deals the manager pulled out named. By week 12 the sequence of called numbers against the closed result is the most honest record of forecast accuracy the team has, and it cost nobody a spreadsheet.

Watch coverage as a trend, not a snapshot. Three-times coverage on week two with a commit at 40% of quota is normal. Three-times coverage on week ten with commit at 60% means the pipeline is padded with deals that will not close this quarter, and the brief can say so: it already knows which of those deals have no next step and no paper process.

Calibrate the checklist against what closed

At quarter end, lay every deal that was in commit at week six beside what happened to it. A good checklist has most of the deals that closed passing all four checks at week six, and most of the deals that slipped failing at least two. Where a clean commit slipped anyway, find the evidence that was visible and unchecked, often the buyer's own timing quote, and add it. Where a flagged deal closed on time, find the check that was too strict for your sales motion, and loosen it. The checklist is a file and the calibration is a one-hour meeting with the data on the table.

Do the same by rep. Some reps commit early and land what they commit; some commit everything and land half. The brief's per-rep line, over twelve weeks, shows that pattern without anyone saying it out loud, and the manager can coach to it: 'your commits at week six passed two checks on average, the team passed three.' That is a conversation about evidence, which is a conversation a rep can act on.

Then share the calibration with the team. A checklist that has been tuned against two quarters of real slips is a credibility asset. Reps stop treating the flags as an accusation and start treating them as the list of things to fix before Monday, which is what they were for.