I replaced 3 weekly rituals with one AI-powered workflow — here's what actually changed

Six months ago, I spent roughly 3 hours every week on three rituals that felt essential: updating my program health dashboards, drafting post-meeting Slack summaries, and synthesizing what happened across a dozen programs before Monday started. I now spend about 20 minutes on all three. Here's what that actually feels like — and what I had to build to make it real.

Most articles about AI productivity stay in the abstract: "delegate tasks to AI," "automate your workflows," "use ChatGPT for meeting notes." They skip the part where that advice is just as much work as the original thing, if not more.

Making AI genuinely useful for program management requires real setup time upfront. You have to define exactly what you want the output to look like. You have to be specific about how decisions get made. You have to build in the right constraints, so the output is trustworthy, not just plausible-looking — which is a very different bar.

I manage 8 to 10 concurrent regulatory compliance programs. Each one has stakeholders across legal, policy, product, and engineering. Each has a weekly health status, a running set of action items, and decisions that need to be captured and tracked. Missing any of these — even once — has real downstream consequences: a legal review that waits a week because no one saw the approval come through, a PM surprised by a blocker that was sitting in my notes and never made it to them, a status update that contradicts what I said in a steerco the week before.

I wasn't looking for an AI assistant. I was looking for a workflow that didn't let things fall through the cracks.

Here's the concrete version.

Automated meeting processing. I run 8 to 12 in-scope meetings per week — 1:1s with engineering leads, weekly PM syncs, compliance escalations, accessibility reviews. For each one, I set up a scheduled job that fires at the end of each working day: it pulls the Zoom transcript, searches for post-meeting signals in the relevant Slack channels and Outlook, and writes a structured notes file directly into my Obsidian vault. Decisions, action items, program signals — all in a consistent format I can scan in under two minutes. No manual note-taking during the meeting. No scrambling to reconstruct context two hours later. The file is just there when I need it.

Monday morning synthesis. Every Monday, instead of starting the week with a vague sense of "what was I even tracking," I run a digest job that sweeps the previous week's meeting notes, six Slack channels, and Outlook signals from the programs I own. It generates a prioritized briefing: what's urgent, what's at risk, what needs a follow-up, and what may have quietly resolved itself. I get the full picture in one document before my first meeting starts. Last week it surfaced a pipeline failure blocking a product launch that no one had escalated directly to me — caught it Monday morning, five minutes before a stakeholder would have messaged asking for status.

Health dashboard automation. My program health dashboards used to take 30 to 45 minutes per week to update manually — pulling status from different sources, writing summaries, updating fields across multiple Airtable tables, and then pushing to a live dashboard. Now I say something like "x is blocked on the y pipeline, update health" and Claude reads the relevant signals from the week, proposes the status, the summary, and the next steps, waits for my approval, then writes to Airtable, updates my Obsidian file, and syncs to the GitHub Pages dashboard my team references. Start to finish: three minutes.

What made all of this work wasn't the model. It was the specification.

I spent time writing precise instructions — in a file called CLAUDE.md — that define exactly how each workflow runs: which meetings are in scope, how to format action items (no owner prefix, no embedded dates), what level of detail belongs in a status summary versus a next-steps field, when to ask me before writing versus when to proceed autonomously. It's closer to writing a junior TPM's operating guide than to writing prompts. Every time an output was wrong, I asked: is this a model failure or a specification failure? Almost always, it was the specification.

Start with output, not input.

When I first started building these workflows, I thought the hard part was getting the model to understand what I was saying. It wasn't. The hard part was figuring out exactly what I needed as output — with enough specificity that I could tell whether the result was right or wrong, not just good-sounding.

The most useful thing I did was audit one week's worth of my actual work and ask: what am I producing, for whom, by when, and what information has to go into it? Once I had that map, building the automation was mechanical. Before that, it was just experimenting.

I'd also start smaller. The Monday digest took three iterations before the format was right. The meeting notes took five. The health dashboard took seven — because the field requirements across Airtable tables are specific, and any ambiguity in the instructions produced plausible-looking but wrong outputs. Each iteration was about 20 minutes. But I ran all of them before I trusted the output with anything real.

The time I save now that the workflows are stable is the easy part to see. The time I put in upfront is the part people underestimate.

AI didn't make me faster at doing the same things I was doing before. It changed which things I have to do myself. I spend less time on the work that was high-volume and low-judgment. I spend more time on the things that actually require a person: deciding what a risk actually means for a program, figuring out how to communicate a delay to a VP without starting a panic, knowing when a decision that looks technical is actually political.

That reallocation is real. But it only happened because I was willing to do the upfront work of defining what "good" looked like before I handed anything off.

If you're a program manager wondering whether AI is worth the investment: the answer depends on whether you're willing to be specific about what you need. If you are, the return compounds fast. If you're not, you'll generate a lot of output that looks right and isn't.

What's one thing you do the same way every single week?