How I use AI agents to run program status reporting

The problem: status gets assembled, not written

Most of a TPM's reporting time doesn't go into thinking. It goes into collecting. Before every status meeting I was opening Jira filters, scanning Confluence pages, chasing owners for updates, counting open P1s and rebuilding the same tables. Across a portfolio of programs spanning more than 15 teams, that was close to two hours every reporting cycle, and the work was out of date the moment it was done.

The judgment part of the job, deciding what's actually at risk and what leadership needs to decide, was getting the leftover time.

Start with one source of truth

AI can't fix scattered information, so the first step had nothing to do with AI. Every program got a Confluence program page with the same structure: scope, milestones, owners, decisions, and a risk register with a mitigation plan and an owner for every risk. Each page links to the program's Jira epics and filters.

That consistency is what makes automation possible. When every program reports the same way, an agent can read all of them the same way.

What the agent actually does

I connected an AI agent to Jira and Confluence through MCP integrations, and gave it context about how our programs work: what "at risk" means, which fields matter, and what each report should contain. Then it runs on a rhythm.

• Daily risk report. For programs flagged at risk, it scans for aging P1s, blocked dependencies, items past their due date and scope added mid-sprint. It drafts a list with an owner, an ETA and a suggested escalation for each item.
• Weekly management report. It compiles RAG status by scope, schedule, quality and budget, pulls the week's highlights, and lists the decisions leadership needs to make.
• Dashboards. It refreshes the program health views: milestone adherence, P1 trend, dependency closure and sprint health. Those five dashboards replaced the tables I used to rebuild by hand.
• Intake analysis. New requests get classified, linked to the right epics, and flagged when they carry hidden dependencies.

The goal was never to automate the status meeting. It was to walk into it already knowing what's wrong.

Where the human stays in the loop

AI drafts are fast, and occasionally confidently wrong. A few rules keep the reporting trustworthy.

• I make the RAG call. The agent can suggest amber. Only the TPM declares it, because a color is a commitment to leadership.
• Every number traces to a source. Reports link back to the Jira filter or Confluence section each figure came from, so anyone can check it.
• Escalations are human decisions. The agent surfaces candidates. I decide who to call and what to ask for.
• Approved tools only. Program data stays inside the company's sanctioned AI tools and systems. Nothing goes into a consumer chatbot.

The results

The biggest change didn't show up in a metric. Risks now surface days earlier, because the daily scan doesn't get skipped on a busy week. Status meetings went from reading tables aloud to discussing decisions.

If you're a TPM starting out with this

1. Standardize first. One program page template, and one definition of each status, applied everywhere.
2. Automate the report you hate most. For me that was the daily risk scan. Get one report right before adding the next.
3. Treat prompts like templates. Write down what "good" looks like for each report, and refine it the same way you'd refine any process.
4. Keep the judgment. AI gives you back the hours. Spend them on the conversations a dashboard can't have.

The examples here describe my approach in general terms. Specific programs, data and internal tools are left out on purpose.