Skip to main content
AGENCYSCRIPT
CoursesEnterpriseBlog
đź‘‘FoundersSign inJoin Waitlist
AGENCYSCRIPT

Governed Certification Framework

The operating system for AI-enabled agency building. Certify judgment under constraint. Standards over scale. Governance over shortcuts.

Stay informed

Governance updates, certification insights, and industry standards.

Products

  • Platform
  • AI Scripts
  • Certification
  • Launch Program
  • Vault
  • The Book

Certification

  • Foundation (AS-F)
  • Operator (AS-O)
  • Architect (AS-A)
  • Principal (AS-P)

Resources

  • Blog
  • Agency Archetype Quiz
  • Free Live Training
  • Build AI Agents Masterclass
  • Build with AI Challenge
  • OS Plugin Install
  • Verify Credential
  • Enterprise
  • Partners
  • Pricing

Company

  • About
  • Contact
  • Careers
  • Press
© 2026 Agency Script, Inc.·
Privacy PolicyTerms of ServiceCertification AgreementSecurityCookies

Standards over scale. Judgment over volume. Governance over shortcuts.

On This Page

Stage One: ProposeWhat the Assistant Does HereWhen to Widen or Narrow ItStage Two: ReviewThe Human GateCalibrating Review DepthStage Three: ActWhere Authority LivesReversibility as a DialStage Four: AuditClosing the LoopFeeding Audit Back Into ProposeApplying the Whole LoopA Worked ExampleWhen to Skip a StageAssigning Owners to Each StageFrequently Asked QuestionsWhy four stages instead of a simpler human-in-the-loop rule?Which stage do teams most often get wrong?How does the audit stage differ from the review stage?Can a small team really run all four stages?When should propose be deliberately narrowed?Does the loop ever fully stabilize?Key Takeaways
Home/Blog/The Plan-Track-Adjust Loop for AI-Run Projects
General

The Plan-Track-Adjust Loop for AI-Run Projects

A

Agency Script Editorial

Editorial Team

·March 11, 2018·6 min read
ai project management assistantsai project management assistants frameworkai project management assistants guideai tools

Most teams adopt an AI project management assistant by switching features on and seeing what sticks. That works until it does not, and when it fails there is no structure to diagnose why. What follows is a named model that gives the adoption a shape you can reason about, repeat across projects, and hand to a new team member.

Call it the Propose-Review-Act-Audit loop, or PRAA. It is deliberately small, four stages, because a model you cannot remember is a model you will not use. Each stage has a couple of components and a clear rule for when it applies. The loop is not linear in the sense of one-and-done; it cycles, with the audit stage feeding back into how you scope the next propose.

The value of a named model is that it turns vague debates into specific ones. Instead of arguing about whether to trust the assistant, you ask which stage a given task sits in and what the rule for that stage says. The model below is the structure behind the practices and checklists discussed elsewhere in this cluster.

Stage One: Propose

What the Assistant Does Here

In the propose stage, the assistant generates candidates: a draft status update, a risk flag, a suggested reprioritization. Crucially, nothing it produces here takes effect. Propose is where the assistant's speed is an unambiguous asset because the cost of a wrong proposal is zero until a human acts on it.

When to Widen or Narrow It

Widen propose for high-volume, low-stakes drafting where iteration is cheap. Narrow it where the assistant lacks context it cannot acquire, like client promises living in email. The scoping logic here mirrors the boundary argued in Best Practices That Hold Up When AI Runs Your Projects. One subtlety of the propose stage is that its volume sets the load on every stage downstream. An assistant configured to propose aggressively floods the review stage, and an overloaded review stage is where rubber-stamping begins. Tuning propose is therefore not only about output quality; it is about keeping the downstream human work at a volume people can actually sustain. A propose stage that generates more than your reviewers can genuinely read has already defeated the model, no matter how good each individual proposal is.

Stage Two: Review

The Human Gate

Review is where a person evaluates the proposal. The component that makes review work is traceability: the assistant must show the evidence behind its proposal so the reviewer can check it quickly rather than re-deriving it.

Calibrating Review Depth

Not every proposal needs the same scrutiny. A channel digest gets a skim; a client-facing email gets a careful read; a reprioritization gets a deliberate decision. Match review depth to stakes, which is the same calibration explored in How to Decide Between Competing AI Project Management Approaches. The failure mode to watch for in this stage is rubber-stamping, where review degrades into a reflexive click because the proposals have been right lately. Calibrated depth is partly a defense against that drift: by reserving careful reading for high-stakes proposals, you keep reviewers' attention where a miss would actually hurt and let the low-stakes ones flow. A review stage that treats everything identically either bottlenecks the trivial or waves through the dangerous.

Stage Three: Act

Where Authority Lives

Act is the only stage where something changes in the real world, and the rule is simple: a human authorizes the act, even when the click is automated afterward. The authority line you draw, which actions may proceed without human authorization, is the single most consequential configuration choice in the model.

Reversibility as a Dial

For genuinely reversible, low-stakes actions, you can let the assistant act autonomously and review after the fact. For irreversible or client-facing actions, authorization comes first. Reversibility is the dial that tells you which mode to use, a distinction the failures in Where Teams Go Wrong Trusting an AI to Run Projects make vivid.

Stage Four: Audit

Closing the Loop

Audit is where you check, after the fact, whether the assistant's proposals and acts were sound. The components are a scheduled spot-check of outputs against their source and a record of corrections that reveals systematic weaknesses.

Feeding Audit Back Into Propose

The audit stage is not a dead end; it reshapes the propose stage. If audits reveal the assistant consistently mishandles a task, you narrow propose for that task. This feedback is what makes PRAA a loop rather than a pipeline. The metrics that power a good audit are detailed in Reading the Numbers That Show an AI Assistant Is Working.

Applying the Whole Loop

A Worked Example

Take weekly client updates. Propose: the assistant drafts from ticket activity. Review: the manager reads carefully because stakes are high and corrects tone. Act: the manager sends. Audit: monthly, the team checks whether drafts have drifted in accuracy. Each stage has a clear owner and rule, and the task moves through them every week. Now contrast that with a low-stakes task like generating a personal daily to-do list from assigned tickets. Propose stays the same, but review collapses to a glance, act becomes autonomous because the action is trivial and reversible, and audit shrinks to an occasional sanity check. The same four stages, calibrated differently, fit both tasks. That portability across stakes is the test of whether a model is genuinely reusable or just a description of one workflow dressed up as a framework.

When to Skip a Stage

You almost never skip review or act, because those are where accountability lives. You can compress propose and audit for trivial, internal tasks where the cost of error is a shrug. Skipping is a deliberate choice tied to stakes, never a default.

Assigning Owners to Each Stage

A model only works when each stage has a name attached. Propose is owned by whoever configures the assistant's scope; review by the person closest to the stakes; act by whoever holds authority for that decision; audit by whoever watches quality over time. On a small team these can be the same person wearing different hats, but the hats still matter, because a reviewer thinks differently than an auditor. Writing down who owns each stage for each recurring task turns PRAA from a diagram into an operating procedure. When a proposal goes wrong, you can point at the exact stage and owner where the chain should have caught it, which is the diagnostic payoff a named model exists to provide.

Frequently Asked Questions

Why four stages instead of a simpler human-in-the-loop rule?

Human-in-the-loop tells you a person is involved but not where or how. PRAA names the four distinct jobs, generating, evaluating, authorizing, and checking, so you can calibrate each independently. The granularity is what makes the model diagnostic when something goes wrong.

Which stage do teams most often get wrong?

The act stage, by letting the assistant authorize its own actions. Collapsing act into propose is exactly how a context-blind reprioritization reaches production. The authority line in act is the model's most important single rule.

How does the audit stage differ from the review stage?

Review happens before an act, on a single proposal, to decide whether to proceed. Audit happens after, across many acts, to detect patterns and drift. Review protects the individual decision; audit protects the system over time.

Can a small team really run all four stages?

Yes, in compressed form. A small team's review might be a glance and its audit a monthly five-minute check. The stages do not require ceremony; they require that each job has an owner and a rule, however lightweight.

When should propose be deliberately narrowed?

Whenever the assistant lacks context it cannot acquire from the data, like commitments made verbally or in side channels. Narrowing propose there prevents confident proposals built on missing information from ever reaching a reviewer who might rubber-stamp them.

Does the loop ever fully stabilize?

The configuration stabilizes, but the audit stage keeps running because model behavior can shift under vendor updates. PRAA is a steady-state operating loop, not a setup phase you finish and abandon.

Key Takeaways

  • The Propose-Review-Act-Audit loop gives AI assistant adoption a repeatable, diagnosable structure.
  • Propose costs nothing until acted on, so widen it for low-stakes drafting and narrow it where context is missing.
  • Review depends on traceability, and its depth should match the stakes of each proposal.
  • Act is the only stage that changes the world; the authority line and reversibility dial govern it.
  • Audit closes the loop by feeding accuracy findings back into how you scope the next propose.

Search Articles

Categories

OperationsSalesDeliveryGovernance

Popular Tags

prompt engineeringai fundamentalsai toolsthe difference between AIMLagency operationsagency growthenterprise sales

Share Article

A

Agency Script Editorial

Editorial Team

The Agency Script editorial team delivers operational insights on AI delivery, certification, and governance for modern agency operators.

Related Articles

General

Rolling Out AI Hallucinations Across a Team

Most teams discover AI hallucinations the hard way — a confident-sounding wrong answer makes it into a client deliverable, a legal brief, or a published report. The damage isn't just to the output; it

A
Agency Script Editorial
June 1, 2026·11 min read
General

A Model Behind an API Is Only Potential

Large language models don't do much on their own. A model sitting behind an API is potential, not capability. What converts that potential into something useful—something that drafts, classifies, summ

A
Agency Script Editorial
June 1, 2026·11 min read
General

Case Study: Large Language Models in Practice

Most teams that fail with large language models don't fail because the technology doesn't work. They fail because they treat deployment as a one-time event rather than a discipline — pick a model, wri

A
Agency Script Editorial
June 1, 2026·11 min read

Ready to certify your AI capability?

Join the professionals building governed, repeatable AI delivery systems.

Explore Certification