AI in the Chief of Staff Workflow: What Actually Works (and What Produces Slop)

The leadership offsite ran four hours. Granola caught every word. You got back to your desk, hit "summarize," and pasted the output into Slack. The CEO read it once and never opened a CoS summary from you again.

Here's what happened. The transcript captured the words. It missed the part where the VP of Sales went quiet at minute 34 and never came back. It missed the CFO saying "interesting" three times in a row, which everyone in the room understood meant no. It flattened the sixty-second pause after the product roadmap slide into a tidy bullet: "Team aligned on Q3 priorities."

The team was not aligned. You knew that. The AI did not.

This is the central problem with AI in the Chief of Staff workflow. The job is one part mechanical compression and three parts political pattern-matching, and most CoS who adopt AI tools confuse the two. They use the model where it shines and also where it fails, and the failure is invisible until the CEO stops trusting their read on the room.

Let's draw the line clearly.

Why This Matters Now

The CoS is the most over-leveraged role in the company. You're three people stitched together: the COO who hasn't been hired yet, the EA who's actually a strategist, and the diplomat who keeps the leadership team from imploding. Every minute you spend transcribing a meeting or formatting a board memo is a minute you're not making the strategic call only you can make.

AI is the multiplier. Used correctly, it gives you back ten to fifteen hours a week. Used incorrectly, it produces output the CEO ignores by paragraph two and quietly downgrades your credibility while you think you're being efficient.

The CoS who can tell the difference between "AI compresses this" and "AI corrupts this" gets the COO seat in eighteen months. The one who can't gets replaced by the AI.

Where AI Helps (Use It Aggressively)

Meeting transcript synthesis

Granola or Otter for capture. Neither is magic, both are good enough. Otter is better in noisy rooms, Granola is better when you want a clean structured output and don't need word-perfect quotes. Pick one and stop debating it.

The capture is the easy part. The hard part is what you do with the transcript. Don't paste it into Claude and ask "summarize this meeting." That gets you generic bullets nobody reads. Use a prompt like this:

Read this transcript. Extract: (1) every decision actually made, (2) the owner and date for each, (3) every disagreement that was not resolved, (4) every commitment that was implied but not explicit. Format as four lists. Do not soften disagreements. Do not interpret silence as agreement.

That last line is the one that matters. Default LLM behavior is to smooth conflict into consensus. Your job is the opposite. Surface the conflict so the CEO can act on it.

Briefing memo first drafts

Claude or ChatGPT for the skeleton. You then rewrite it in the CEO's voice, which the model cannot replicate no matter how much context you feed it.

A working pattern: feed the model the raw inputs (the WBR data, the Slack threads, the transcript, last week's memo), ask it for a structured first pass with a specific format ("three sections: what's working, what's broken, what I'm asking you to decide"), and treat the output as a wireframe. Maybe 20% of the language survives the rewrite. The point isn't to ship the AI's draft. It's to skip the blank page.

Board narrative drafting

The model writes the safe version. You add the uncomfortable truth.

A board deck is two documents in one: the version that explains what happened, and the version that signals where the company actually is. The first is mechanical and a model can do 80% of it. The second is judgment, and if you let the model write it, you'll produce a deck that makes the company sound like every other company. Boards notice that. They start asking sharper questions because the narrative feels generic.

Use AI for the chart explanations, the appendix tables, the historical context. Write the "what we got wrong this quarter and why" section by hand.

Scheduling automation

Reclaim, Motion, or just a clean Cal.com setup. The mechanical Tetris of moving meetings is the lowest-leverage thing a CoS does and the highest-leverage thing to automate. If you're still manually booking the CEO's 1:1s, you're paying yourself $150K to do what a $20/month tool does better.

The exception: any meeting involving conflict, performance issues, or external stakeholders the CEO has a relationship with. Those you book yourself, because the framing of the invite is part of the work.

Stakeholder pre-reads

A board member sends a 40-page deck the night before a meeting. The CEO will not read it. You won't either, but you need to walk in informed.

Feed it to Claude with a prompt like: "Summarize this for someone who has 8 minutes. Lead with what they're trying to convince us of, then their three strongest arguments, then the weakest claim that's hiding in the data." That last clause is the unlock. Default summaries flatten everything to equal weight. You want the asymmetric read.

Where AI Breaks (Do It Yourself)

Judgment calls

Which initiative to kill. Who to promote. Whether the new hire is going to work out at month three. These are not text problems. They're pattern-matching across humans, history, and context the model doesn't have and shouldn't.

A CoS who outsources judgment to a model is producing analysis that sounds rigorous and is hollow. The CEO can smell it. So can the leadership team.

Leadership conflict mediation

When the VP of Engineering and the VP of Product are in a slow-burn fight over roadmap ownership, the work is not text. The work is reading the room, knowing which one is actually fighting about something else (usually compensation, scope, or feeling unheard), and brokering a conversation that lets both save face.

You cannot prompt your way to that. The model will give you a "framework for resolving cross-functional conflict" and it will read like a Harvard Business Review article from 2014. Useless.

CEO trust calls

What to escalate, what to absorb, what to handle without telling the CEO at all. This is the most sensitive judgment in the job and the model has no business near it. The model doesn't know which board member the CEO is trying to manage, which co-founder relationship is fragile this quarter, which investor email is actually a soft threat.

If you let AI help you decide what to surface, you'll surface the wrong things. The CEO loses trust in your filter. Once that's gone, you're a glorified scheduler.

Strategic narrative

The story behind the numbers is yours to construct. Not the chart explanations; those are mechanical. The story. Why we missed this quarter. Why the new market is harder than we thought. What changed in the second half that made the original plan wrong.

A model can produce a competent narrative. It cannot produce a true one, because truth in this context is a function of what the company can stomach hearing, what the CEO is ready to act on, and what the board is actually worried about. You hold those three variables. The model holds none of them.

The Slop Patterns to Avoid

Pasting raw transcripts into Slack. The transcript is input, not output. Anyone can run Granola. Your value is the synthesis, not the recording.

Using AI-generated bullets in CEO-facing memos without rewriting. Tells. Every time. The CEO knows your voice. AI bullets are flatter, hedgier, and use the word "leverage" as a verb. If you ship them unedited, you train the CEO to skim your work.

Letting AI smooth out the conflict. This is the worst one. A leadership meeting where Engineering and Sales fought is a high-signal event. The AI summary that reports "team discussed Q3 priorities and aligned on next steps" deletes the entire signal. Your memo should say: "Engineering and Sales disagreed on the Q3 capacity allocation. Engineering is asking for two weeks to scope the tradeoff. Recommend we don't decide until they come back." That is what a CoS sounds like.

Treating AI as a peer. It's a junior analyst with infinite patience and zero context. Brief it accordingly. Vague prompts produce vague output. The CoS who writes good prompts gets useful drafts. The CoS who types "summarize this meeting" gets slop.

A Working Tool Stack

Skip the marketplace tour. Here's a version-agnostic stack that works.

• Transcripts: Granola (default) or Otter (for noisy rooms). Pick one.
• Drafts and synthesis: Claude (better at long-form, judgment-adjacent reasoning) or ChatGPT (better when you need fast structured output). Most CoS will use both.
• Scheduling: Reclaim or Motion for autopiloted calendar; Cal.com for external bookings.
• Search across your tools: Notion AI if your org lives in Notion, otherwise Glean or Mem.
• Research: Perplexity for fast factual questions with sources. Skip ChatGPT search for anything you'll cite.

That's it. If you need more than five tools to run the AI side of your workflow, you're collecting tools instead of doing work.

A 30-Day Plan to Integrate AI

Week 1: Audit. Track every task you do for one week. Tag each as mechanical (transcribe, schedule, summarize) or judgment (escalate, broker, narrate). Don't change anything yet. The audit is the work.

Week 2: Layer in transcripts and drafts. Pick Granola or Otter. Run it on every meeting where you'd normally take notes. End each day, feed transcripts to Claude with the structured prompt above. Start producing first-pass briefing drafts in Claude before rewriting. Goal: cut your memo turnaround time in half.

Week 3: Add scheduling and research. Set up Reclaim or Motion for the CEO's calendar. Move all stakeholder pre-reads into the Perplexity-then-Claude pipeline. Goal: stop being the person who manually books meetings.

Week 4: Measure. Hours recovered per week. Briefing memo turnaround. And the test that matters most: what does the CEO say about your output? "This reads like you" is the goal. "This sounds different lately" is the warning. "I didn't read this" means you've already failed.

After 30 days, you should have 8-12 hours back per week. That time goes to the judgment work: the conversations, the pattern-matching, the calls only you can make.

Optional: The ACE Framework Lens

If you want a structured way to map AI to your workflow, the ACE Framework (Ingest, Analyze, Predict, Generate, Execute) is a useful overlay:

• Ingest: transcripts, Slack, deck uploads. AI does this well.
• Analyze: synthesis, theme extraction, pattern surfacing. AI helps; you verify.
• Predict: calendar conflicts, meeting prep gaps, follow-through risk. AI handles the mechanical version.
• Generate: first-draft memos, chart explanations, pre-read summaries. AI produces; you rewrite.
• Execute: escalation, conflict brokering, the call. Still you. Forever you.

The pattern: AI compresses the first four layers. You own the fifth. If you flip that, you're done.

Measuring Success

Three metrics, in order of importance.

1. CEO comments on your output. "This reads like you" beats every productivity dashboard. If the CEO starts forwarding your memos verbatim to the board, you're winning. If they stop opening them, you've shipped slop and need to roll back.

2. Briefing memo turnaround time. Target: 60 minutes from meeting end to memo delivered. Pre-AI, this was 3-4 hours for most CoS. With the right pipeline, it's under an hour.

3. Hours recovered per week. Track honestly. Most CoS who integrate AI well recover 8-15 hours. If you're recovering 2 hours, you're not using it aggressively enough on the mechanical work. If you're recovering 25, you've probably outsourced something you shouldn't have.

The Real Test

A leadership meeting wraps. The CFO said "interesting" twice. The VP of Engineering looked at the floor when the Q3 plan came up. The CEO ended five minutes early.

You have two choices.

The first choice: paste the Granola summary into the CEO's Slack with a tidy bullet list. Three days later, the same conflict erupts in writing and the CEO asks why you didn't flag it.

The second choice: send a short memo that reads, "The plan didn't land. Engineering has concerns they didn't surface. CFO is signaling no without saying it. Suggest we reconvene with just the three of you Thursday." No bullet list. No transcript. Just the read.

The first one is what AI produces. The second one is what you produce. The whole job is knowing which one to ship.

Learn More

• Briefing Memo Templates the CEO Will Actually Read
• Running the Executive Operating Cadence
• How a CoS Builds Trust with the CEO
• Business Operations Manager Job Description