Pipeline updates an agent proposes and the deal desk approves

Most pipeline rot is not laziness — it is latency. The signal lives in a thread, a call recording or a redlined order form, and by the time someone transcribes it into the opportunity record, the forecast has already been wrong for a week.

An agent is very good at reading that signal and drafting the update. It should not be the thing that commits it. ActionsLedger routes each proposed amount, stage or close date through policy and a named approver, then merges with a separate credential and records the basis.

[ 01 / 03 ] · WHERE IT BREAKS // without a review layer //sha 4b227w

Forecast moved on a hallucinated date

"End of quarter" from an email becomes a specific close date in the wrong fiscal calendar. Syntactically valid, materially wrong, and it lands in the board number.

CONTROL → EVIDENCE + HUMAN GATE

Discount and terms drift

The agent copies a discount from an earlier draft of the order form. Nobody notices until billing invoices at the wrong rate.

CONTROL → APPROVAL BY AMOUNT BAND

Stage inflation

Optimistic reading of a thread pushes deals to a later stage. Conversion rates per stage stop meaning anything, so pipeline coverage math breaks.

CONTROL → STAGE RULES AS CODE

Approvals with no paper trail

A non-standard term was approved in a DM. Six months later, at renewal, no one can point to who agreed to what.

CONTROL → HASH-CHAINED AUDIT

[ 02 / 03 ] · THE REVIEW LOOP // propose, verify, merge //sha 6b86n2

Signal in, change set out

The agent reads the thread, the call summary and the order form with read-only access, then emits a change set: amount, close date, stage, next step, competitor, each with the message ID or document span it came from.

Policy prices the risk

Deterministic rules score each change: size of the amount delta, whether the close date crosses a quarter boundary, whether terms are non-standard, whether the stage jump skips a gate.

The right human, not any human

Low-impact edits go to the rep; amount bands and non-standard terms route to the deal desk; anything crossing a quarter boundary shows the fiscal calendar it was resolved against before approval.

Merge and lock the record

The merge worker writes the approved fields only, then appends diff, evidence, verdict, approver and timestamp. Quarter-end close windows make the affected fields immutable, and late proposals are re-evaluated at merge time.

[ 03 / 03 ] · POLICY // what auto-merges, what waits //sha d4735e

Revenue fields, and who clears them

Next step, notes, competitorAuto-merge with a citation. Low blast radius, high hygiene value.

Amount, under thresholdRep approval. Show delta, percentage and the evidence span.

Amount, over thresholdDeal desk approval. Requires a document, not a summary.

Close date across a quarterDeal desk approval. The fiscal calendar used is shown in the diff.

Discount and non-standard termsDeal desk approval, with the order form span attached.

Closed-won / closed-lostHuman only, and immutable once the close window ends.

deal desk — approval routing

agent                 read-only: thread, call, order form
rep                   approves low-band edits
deal desk             approves amount / terms / date
merge worker          sole writer, approved fields only
close window          fields immutable after cut-off

re-evaluated at merge time, not approval time
stage jumps that skip a gate fail closed
every merge appends: diff · evidence · verdict · approver

Accurate pipeline, without an agent holding the write key.

The agent drafts from real evidence, the deal desk approves what matters, and every number in the forecast has a traceable basis.

Start for free All use cases