Product · AI champion · AI-enabled PMO

Action items, extracted and graded.

A read-only agent reads meeting notes, email, calendar and the Monday board, writes every commitment it finds with owner, date and the line that carries it, and asks a person to grade each one. Accuracy is measured before anything is automated.

0commitments found, first run
0sources read: Otter, Outlook mail, calendar, Monday
0 dayswindow, Aug 17 to Aug 31
0%precision on the first 20 graded rows
StatusRead-only. Running weekdays 7:25.
Proof dataDallas's own records
ByDallas Andrews, AI champion, Product
Grades so far20 graded, precision 95%
In one screen
  • What runs. Every weekday morning the agent reads the four sources for the last window and appends what it finds to a register: one row per commitment, with owner, due date, and a verbatim quote as evidence.
  • How accuracy is proven. Each new row also lands in a validation queue with a blank verdict. A person marks it correct, wrong owner, wrong date, not an action, or duplicate. The next run reports precision from the graded rows.
  • What it does not do. Send, remind, create tasks, or edit any source. Follow-up automation is a separate decision, taken after the precision number is known.
  • Open with Product. Which project records and whose notes count as "the PMO's". The first run proves extraction on Dallas's own records only.
Mechanism

Four sources in, two files out.

Meeting notesOtter transcripts and action items67 found
EmailOutlook inbox and sent, requests and commitments13 found
CalendarPrep commitments and deadlines in invites2 found
Project recordsMonday.com AI Initiatives board, Status Notes and updates3 found
pmo-action-extractor

Extract, attach evidence, assign confidence, dedup against the register, mint an id.

register · validation queue · daily log
id, run_date, action, owner, due, due_basis, source_type, source_ref, source_date, evidence, confidence, status  (+ verdict, verdict_note in the queue)

Grading, five verdicts

correctRight action, owner and date
wrong_ownerAction real, owner misread
wrong_dateAction real, date misread or invented
not_an_actionNot a commitment
duplicateAlready in the register

Precision = correct / graded. Reported in every daily log with the graded count, so the number is always the current one.

First run, Aug 31 2026

85 commitments from four sources.

Window Aug 17 to Aug 31. Read: 19 Otter meetings (17 transcripts fetched for the verbatim line), 185 inbox and 82 sent messages listed with 13 bodies read, 34 calendar events, the 68-item Monday board. 5 items have no named owner and 48 carry no stated date; the log names the 14 of those that read as near-term commitments. Requests completed inside the window were left out on purpose and listed in the log.

ActionOwnerDueSourceEvidenceConfidence
Launch multi-channel sequences to the roughly 300 identified back-office contacts by the all hands.Dallas2026-09-16otter: Back Office Lead Expansion Strategy"Is it unreasonable to target at least launching a multi-channel sequence for these 300 contacts we have identified by the all hands?"medium
Bring the campaign framework and lead the new-campaign build-out discussion at the Sep 3 GTM Campaign Alignment meeting.Dallas2026-09-03calendar: GTM Campaign Alignment Meeting"If anyone has a campaign framework you are currently working with, please provide to this team so we can be consistent as we bring new campa"medium
Develop the read-only action extraction agent for the AI enabled PMO goal.Dallas2026-12-31monday: AI enabled PMO 12664120782"Develop an agent that ingests notes, email, calendar, and project records to extract action items with owners and dates, starting read only "medium
Update the product team's percent complete on the Monday AI Initiatives board at least once a month.Dallasnone statedotter: AI Champion Role Transition Overview"at the end of the month, just kind of moving the percent complete on the goals because those are rolling up to the company scorecard."high
Build pre-approved outreach templates with marketing so campaigns can go out without per-piece approval.Dallasnone statedotter: AI Projects and Clay Integration Discussion"I'm just going to create templates for different situations, and once I get approval for the way those look like, then it's just going to be"medium
Build the AI adoption tracking dashboards and technical deliverables for the side quest before year end.Dallas2026-12-31otter: AI Change Management Side Quest Kickoff"Dallas is going to be really our guru on the technical side to build some dashboards and tracking that we can use to then fire or trigger so"high
Six of the 85 rows, all owned by Dallas, verbatim from automation/pmo/action_register.csv. Otter speaker labels are generic in several meetings; where a name was inferred the row is medium confidence and the log says so.

Where the number comes from

20 of 85 rows graded on Aug 31 by Dallas: 19 correct, 1 wrong date, 0 wrong owner, 0 not an action, 0 duplicates. All 14 graded meeting-note rows and all 10 high-confidence rows were correct; the miss was a medium-confidence email row with no date stated. The figure moves as more rows in automation/pmo/validation_queue.csv carry verdicts; the daily log reports it each run.

Open with Product
  • Scope. The goal says notes, email, calendar and project records. Whose, and which boards, is Product's definition of the PMO; the agent reads one person's records today.
  • Grading. 20 rows sit at 95%. How many more before the number is trusted, and who grades beyond Dallas, is the decision that gates automation.
  • Follow-up. Reminders or task creation stay off until the precision figure is known and agreed.
Next: grade the remaining 65, carry the precision line to the AI champions standup.