Skip to contentSkip to main content
← Back to all posts

Weekly Decision Review: Use AI Change Signals Without Re-reading Every Note

Run a 30-minute weekly AI decision review with a decision register, change-signal triage, source checks, and approval-first follow-ups.

Share by email

Projects rarely lose their direction in one dramatic moment. More often, a customer changes a requirement, a meeting introduces a dependency, or somebody edits a source note without updating the decision that depended on it. By Friday, the team may have accurate notes and still be working from an outdated plan.

A weekly AI decision review asks a different question from a weekly status report: Which decisions still hold, what evidence changed, and which follow-ups need an owner? You need a bounded comparison against trusted decisions, not another summary of every document.

This practical guide uses Strata for source-backed decisions, Pulse for change signals, and Forge for reviewed next actions. It includes a portable template, a 30-minute routine, and an acceptance test you can also use without Phasoric.

1. Separate decisions, evidence, and tasks

These three objects are related but not interchangeable:

  • Decision: A reviewed choice that governs future work, such as "Start with three pilot customers."
  • Evidence: A meeting note, customer statement, experiment, or requirement supporting or challenging that choice.
  • Task: Something a named person must do, such as "Update the rollout checklist by Thursday."

New evidence does not automatically cancel a decision. Completing a task does not prove its rationale remains sound. Give each meaningful decision an owner, confirmation date, rationale, source link, review condition, and status.

Strata's decision-memory model preserves reviewed decisions and their evidence, and distinguishes confirmed choices from unreviewed candidates. Later source changes can raise review signals without rewriting history.

Illustrative scenario: A company limits a pilot to three customers because support can onboard only three weekly. New staffing might allow six, but testing and contractual approvals could still constrain the rollout. An AI agent may notice the changed assumption; a human should decide whether the pilot size changes.

2. Start with a small, portable decision register

Choose one active project and record three to ten consequential decisions. Do not start by importing every historical chat. A usable Markdown record looks like this:

## DEC-014 — Pilot release size
Status: Confirmed
Owner: Product lead
Confirmed: 2026-11-02
Decision: Start with three pilot customers.
Rationale: Support capacity and incomplete migration checks.
Evidence: [Kickoff notes](../meetings/kickoff.md)
Depends on: Staffing, migration checklist
Review condition: Staffing change or checklist completion
Last reviewed: 2026-11-02
Replaces: None

The example is synthetic; replace the dates and link with real evidence. Keep the original source. A generated explanation can link to it, but an AI summary must not become independent proof of its own claims.

"Confirmed" does not mean "permanent." A source can become stale without conclusively disproving a decision. Likewise, a decision can remain active while overdue for review. Preserve the distinction between decision lifecycle and evidence trust.

3. Establish a baseline and compare changes

A weekly review needs two states: the last reviewed baseline and the current project. Record active decision IDs, unresolved questions, overdue commitments, and evidence sources at each checkpoint.

Phasoric's Pulse change-intelligence guide describes per-lens baselines for decision edits, supersession, source-note changes, overdue tasks, conflicting statements, unprocessed captures, and trust-review deadlines. Rather than asking an assistant to repeat the entire project, the reviewer can inspect specific differences.

A manual alternative works too: maintain a dated weekly-review Markdown file with active decisions and evidence paths. Next week, list only what was added, modified, reaffirmed, superseded, or left unresolved. Do not confuse a newly captured document with a newly approved decision.

4. Prioritize consequences, not recency

Use a simple triage matrix:

SignalFirst questionReview action
Source changedDoes it affect a confirmed rationale?Open the affected passage and its decision.
Decision supersededIs downstream work still following the old rule?Identify tasks relying on the predecessor.
Apparent contradictionAre scope and effective date the same?Preserve both claims and request review.
Commitment overdueDoes the owner or deadline remain valid?Reconfirm ownership and timing.
Detached evidenceWas a source moved, changed, or deleted?Repair provenance or mark unsupported.
Trust review dueIs the source still current enough?Recheck primary evidence.

Prioritize changes involving security, customers, payments, production releases, and contractual obligations. A punctuation edit in a meeting note should not outrank a newly revoked access rule.

For each significant item, ask "why is this here?" and open the source. Pulse supports inspecting signal factors and navigating to the underlying record. A signal is an investigation lead, not authority to rewrite policy.

5. Configure a focused Pulse lens

Open Pulse and choose a decision-oriented lens such as Decision Review or Project Health, both described in the product guide. Begin with decision changes, tensions, overdue commitments, and changed or stale trust records. Keep news and exploratory research in a separate lens so they do not drown out operational risks.

Pulse's deterministic on-device signals are distinct from hosted background automation. Scheduled briefings have plan and environment requirements; a local-only vault does not become readable by a hosted process merely because a schedule exists. Start with a deliberate manual review and expand automation only after confirming source availability and permissions.

6. Give every signal a review disposition

Choose one of five outcomes:

  1. Reaffirm: Keep the decision and record the evidence reviewed and next checkpoint.
  2. Revise supporting details: Correct a source, owner, or explanation without pretending the governing choice changed.
  3. Supersede: Approve a new decision and link it to the old one. Preserve historical rationale.
  4. Defer: Record what evidence is missing, who must obtain it, and when to revisit.
  5. Reject: Identify an irrelevant or false tension and explain briefly why it needs no action.

An apparent contradiction may really be two different contexts: "Provider A for prototype" and "Provider B for production" can both be true. Check the scope and effective period before declaring a conflict.

An assistant can propose a disposition; an authorized human should confirm a consequential change. Treat unreviewed meeting notes as candidate evidence, not standing policy.

7. Convert decisions into reviewed actions

A beautiful digest without named ownership accomplishes little. Convert confirmed review outcomes into small, verifiable tasks, document updates, or drafts.

The Forge workflow guide documents local note updates, checklist extraction, decision-register notes, calendar-file exports, and follow-up drafts. Its process is proposal, preview, edit, approve, then apply, with an audit record.

Before approval, check target, effect, source, and reversibility. Does the proposed change touch the correct project? Is the diff exactly intended? Is the evidence still current? Is undo supported for this type of action?

A launch postponement could produce a revised checklist and a draft customer message. Generating that draft is not sending it. External actions depend on a supported executor and explicit authorization; a generic connector proposal does not imply direct provider execution.

8. A 30-minute weekly review

Minutes 0–5 — Scope: Choose one project, its decision register, and the previous review date. Keep unrelated vaults outside the review.

Minutes 5–12 — Delta scan: Inspect the relevant Pulse lens or compare the register with recent source revisions. Identify evidence gaps explicitly.

Minutes 12–20 — Decisions: Review the primary sources for the three highest-consequence signals. Reaffirm, supersede, or defer. Name owners for unresolved questions.

Minutes 20–26 — Actions: Prepare tasks, revisions, or follow-up drafts. Inspect proposed effects in Forge where supported, or review a manual diff.

Minutes 26–30 — Record: Save the dated review, the approved actions, what remains open, and the next baseline. Other assistants should retrieve this reviewed record rather than an unfiltered chat transcript.

Aim for an output that fits on one screen: decisions reaffirmed, decisions changed, highest risks, approved actions, and evidence still required.

9. Test your review process before trusting it

Use disposable synthetic notes for six acceptance checks:

  • Change one source and confirm that its affected decision can be located.
  • Add two apparently conflicting statements with different scopes; ensure no automatic supersession.
  • Remove an evidence source and verify that missing evidence is flagged, not reconstructed.
  • Change a task owner and inspect whether the proposed action targets the correct person.
  • Deny a Forge proposal and verify the underlying record remains unchanged.
  • Start a fresh assistant session and ask which decision is current, why, and what it replaced. Verify the answer against original sources.

If a check fails, fix the source mapping, authority boundary, or approval process before increasing automation. A polished briefing is not proof that a workflow is safe.

What success looks like

The measure of a weekly AI decision review is not how many notes were summarized. It is whether a team can tell what changed, explain which decisions still apply, and catch stale assumptions before they affect work.

Start with one project, one baseline, and one weekly checkpoint. Use Strata, Pulse, and Forge to keep evidence, attention, and approved action connected while preserving human control.

Frequently asked questions

How is a weekly decision review different from a status report?

A status report summarizes activity. A decision review tests whether confirmed decisions still hold, which evidence changed, and which reviewed follow-ups have owners.

Does Pulse automatically change a decision when its source changes?

No. Pulse surfaces reviewable source-linked signals. A reviewer decides whether to reaffirm, revise, defer, or supersede the governing decision.

Can Forge send follow-up emails without approval?

A draft or connected-service proposal is not a sent email. Forge requires review and explicit approval, and external execution depends on a supported executor.