Pocket PMO
Sign In
Back to blog

AI agents in project delivery: approval-first in practice

August 11, 2026
AI agents in project delivery: approval-first in practice

There are two versions of AI agents in project delivery. In the first, an agent watches your delivery data, notices things a busy human misses, and hands you a drafted action. In the second, the agent emails your sponsor at 3am about a risk it invented.

The technology behind both is nearly identical. The difference is entirely a design decision about who is allowed to act.

Table of contents

What an agent is, in delivery terms

An agent is a background process with a defined remit that reads the same data you do and produces a proposal. Not a chat window you have to remember to open. The distinction matters: a chatbot answers when asked, an agent notices when you are not looking.

Useful remits on a real project are narrow and specific:

  • Watch for milestones whose dependencies have slipped and propose a revised date.
  • Watch for RAID items with no update in three weeks and propose a chase or a closure.
  • Watch for workstreams that have gone quiet and propose a risk.
  • Watch for budget variance beyond a threshold and draft the escalation.

Each of those is a job a good PM does when they have time. The point of an agent is that it happens on the weeks you do not.

The approval-first pattern

Approval-first means the agent may propose, never act. The mechanics:

  1. The agent detects a signal and produces a proposal with the evidence attached.
  2. The proposal lands in a review queue owned by a named human.
  3. The human approves, edits or rejects, with a reason on rejection.
  4. Everything is written to the audit log: what was proposed, what evidence, who decided, when.

This costs a click per proposal and buys three things you cannot get any other way. Your governance stays defensible, because every change has a human decision behind it. Your data stays clean, because plausible-but-wrong proposals get filtered by someone who knows the project. And your team keeps trusting the tool, which is the asset that takes longest to build and least time to destroy.

The test for any vendor is a single question: can the AI send, log or change anything without a human clicking? If the answer is yes, you are buying a governance liability.

Where agents genuinely help

In practice, the highest-value agent work is the low-judgement, high-frequency work:

  • Drafting the status report from live plan and RAID data.
  • Proposing RAID entries from real signals rather than intuition.
  • Chasing owners with a drafted, human-approved message.
  • Spotting drift early, particularly the quiet kind that nobody escalates because no single week looks bad.
  • Preparing the meeting, with an agenda built from open actions and decisions due.
  • Summarising for a sponsor, turning delivery detail into three paragraphs and a recommendation.

None of that replaces the PM. All of it removes the reason PMs work Sunday evenings.

Where agents should never be trusted

Be equally clear about the boundaries:

  • Closing risks. Only a human can judge that a risk has genuinely gone away.
  • Committing dates externally. A proposed replan is a proposal until the owner agrees.
  • Anything that reaches a stakeholder unreviewed. Email, invitations, shared reports.
  • Interpreting people signals. An agent can see that a workstream has gone quiet. It cannot tell you why, and the why is usually the thing that matters.

Rolling agents out without losing trust

Run it in three stages.

Shadow. Agents produce proposals, the PM keeps working as normal and compares. Two to three weeks. You are measuring how often the proposal would have needed editing.

Assist. Proposals become the starting point. The PM reviews and approves rather than drafting. Measure time saved and, more importantly, whether quality holds.

Standard. Agents run across the portfolio, with the review queue as part of the weekly routine and the audit log as the evidence for assurance.

Do not skip shadow mode. It is the stage where a badly tuned agent gets caught before it has cost you credibility with a sponsor.

Related reading

Frequently asked questions

What is an AI agent in project management?

A background process with a narrow remit that monitors live delivery data and proposes actions, such as a revised milestone date or a new risk, for a human to approve.

Is it safe to let AI update a project plan?

Only as a proposal. The plan is a governance artefact, so the change should be reviewed and approved by the owner and recorded in the audit log.

How is an agent different from a chatbot?

A chatbot responds when you ask. An agent watches continuously and raises something when a signal appears, which is where most of the value sits.

What should I measure during a rollout?

Edit rate on proposals during shadow mode, time saved per report during assist mode, and report quality throughout. If edit rates stay high, the agent remit is too broad.