
How AI Gives CEOs Their Time Back
Three good decisions
In a September 2018 interview with David Rubenstein at the Economic Club of Washington, Jeff Bezos argued that a senior executive is paid to make a small number of high-quality decisions, not thousands of them every day. For him, he said, roughly three good decisions in a day was enough, as long as each was as good as he could make it. He also said things are different at a startup. (Source: the Economic Club’s edited transcript of the interview, pages 19–20.)
That’s one person’s account of how he chose to work, not a target for anyone else. The idea I want to borrow is narrower. If only a few decisions really matter, the work most worth getting help with is making those few ready.
What decision readiness means
A decision is ready when four things are true:
- It has an owner. One named person makes the call and is accountable for it.
- The relevant inputs are gathered. The facts, numbers, and opinions the decision actually depends on are in one place.
- Each input has an identifiable source. There’s a record you could open, ideally with a date or version, rather than a sentence that sounds right.
- The uncertainties are explicit. Anything missing, stale, contradictory, or inaccessible is written down instead of smoothed over.
If one of those is missing, the decision isn’t ready, however polished the briefing looks. Sometimes the right outcome of preparation is to defer the decision until a specific gap is closed.
What AI can prepare, and the access it needs
An AI system can help assemble that preparation. A few illustrative examples of what you might ask for:
- The point you need to land in a meeting. Give it the meeting notes, the proposal, and any earlier correspondence, and ask for a draft of the message and the decision being asked for. (I’ve written more about the questions good meeting preparation should answer.)
- A summary of the discussion so far. If the relevant Slack threads are exported for it, or the tool has been given permission to read those channels, it can summarise what was said and by whom. It can only summarise what it can see. Private messages, channels it can’t read, and conversations that never got written down are missing from its picture.
- Patterns across records you supply. Hand it several months of notes about, say, delayed renewals, and ask what reasons recur. Treat any pattern as a lead to check, not a finding.
Each example depends on the same conditions. The records have to be supplied, or accessible with permission. The output only covers those inputs, so a summary of three threads says nothing about the fourth. And what comes back is a draft. Ask it to cite the record behind each claim and to list what it looked for but couldn’t find. That makes the next step possible.
What a person checks
An AI draft becomes judgment only after a person has checked it. That check is specific:
- Open the cited source for every claim the decision depends on and confirm it says what the draft says.
- Check dates and versions. Is the record current enough for this decision?
- Look at what’s missing: the gaps the draft listed, and any source you expected it to use that it didn’t.
- Write down who checked and any discrepancy that’s still unresolved.
The checker doesn’t have to be the decision owner, but their name goes on the check. An unresolved discrepancy isn’t a failure of the process. It’s exactly the information the owner needs before deciding, or before choosing to wait. A separate article looks at the role of AI in leaders’ decision-making.
A week of decisions
Here’s the worksheet. Copy it into whatever you write in; it’s plain text and needs nothing else.
DECISION PREPARATION WORKSHEET
1. DECISION
Decision to make:
Owner (who decides):
Decide by:
2. INPUTS
For each input the decision depends on:
- Input:
Source (see 6):
Status: checked / unverified / missing / stale / contradictory / no access
3. PREPARER
Who or what prepared the draft:
Records it was given or allowed to access:
4. AI DRAFT
Summary or recommendation:
Claims the decision depends on, each with the source the draft cites:
Gaps the draft says it couldn't fill:
5. HUMAN CHECK
Checked by:
Date checked:
Claims confirmed against their source:
Unresolved discrepancies:
Ready to decide? yes / defer until:
6. SOURCES
For each record used:
- Record:
Date or version:
Where to find it:
An illustrative example
The decision, company, people, and records below are invented to show how the fields fit together. None of it is a real case, and the source records don’t exist.
DECISION PREPARATION WORKSHEET (illustrative example; all records fictional)
1. DECISION
Decision to make: Renew the analytics vendor contract at the new
price, or run a trial of an alternative first?
Owner (who decides): Head of operations
Decide by: Friday, before the renewal notice period ends
2. INPUTS
- Input: Renewal price and seat count
Source: S1
Status: checked
- Input: How many people actually use the tool
Source: S2
Status: contradictory (S2 shows more active users than S1 has seats)
- Input: Whether the team is happy with the tool
Source: S3
Status: stale (survey is over a year old)
- Input: Remaining budget for the year
Source: none
Status: no access (finance sheet wasn't shared with the preparer)
3. PREPARER
Who or what prepared the draft: AI assistant, prompted by the
chief of staff
Records it was given or allowed to access: S1, S2, S3
4. AI DRAFT
Summary or recommendation: Renew. Usage is steady and the price
increase is within the range of last year.
Claims the decision depends on:
- New price and seat count (cites S1)
- Usage is steady (cites S2)
- Team satisfaction is good (cites S3)
Gaps the draft says it couldn't fill: No budget figures available.
5. HUMAN CHECK
Checked by: Chief of staff
Date checked: Wednesday
Claims confirmed against their source: Price and seat count match S1.
Unresolved discrepancies:
- S2 lists more active users than S1 includes seats; the draft
didn't notice. Renewing as quoted may not cover everyone.
- "Last year's range" isn't in any supplied record. Removed.
- S3 predates the current team; satisfaction is unverified.
Ready to decide? Defer until finance confirms budget and the
vendor confirms how extra users are billed.
6. SOURCES
- Record: S1, vendor renewal quote (fictional)
Date or version: version 2, dated 3 March
Where to find it: shared contracts folder
- Record: S2, usage export from the vendor's admin page (fictional)
Date or version: exported 1 March, covering February
Where to find it: attached to the renewal thread
- Record: S3, team tool survey notes (fictional)
Date or version: last updated over a year ago
Where to find it: operations wiki
Notice what the check did. It confirmed one claim, caught a contradiction the draft missed, removed a claim with no source, and turned “no budget figures” from a footnote into a reason to wait. The draft was useful; it just wasn’t the decision.
Try the worksheet on your decisions for one working week. My editorial guess is that the first one takes 10 to 15 minutes when the inputs are already to hand. That’s an estimate, not a measurement, and it will take longer if you have to go looking. At the end of the week, look back at section 2 and note which inputs had no source anyone could check. Those are the inputs to find a source for first.
Want help turning ideas like these into reality? Let's talk.