Your 90-Day AI Plan

What every executive should know

1. 1. How to choose three, when six ideas look equally good

   The five-question test from Briefing 5 filters out bad ideas. It doesn't rank good ones, and most leadership teams walk out of a planning meeting with more qualified ideas than they have room to pilot at once. When that happens, three tie-breakers matter more than enthusiasm:

   • Visibility: a small, visible win that people notice beats a bigger win nobody sees, since momentum for the next 90 days depends on the first one being obvious, not just real.
   • Reversibility: if this goes sideways in week two, how hard is it to unwind. Favor the idea you could quietly stop without anyone outside the pilot noticing.
   • What it teaches you: some pilots mainly produce a result, others also teach your team how to run the next one better. Given a close call, pick the one that teaches you the most about doing this again.
   • None of these three override the five-question test. They're what you reach for once several ideas have already passed it.

2. 2. The trap in "Do Now": picking safe instead of picking right

   The most common way a 90-day plan quietly fails is three good-sounding, low-stakes Do Now items, chosen specifically because they're easy and can't embarrass anyone, while the one idea that would actually move the business gets bucketed into Wait because it's the uncomfortable one.

   A rough test: if none of your three Do Now items would have happened anyway, without an executive briefing series prompting them, you probably picked the right three. If all three are things a mildly motivated employee could have started on their own last month, you picked the safe three instead.

3. 3. What a real plan looks like, filled in

   A blank template invites vague answers. Here's what one business's finished plan actually looked like, a 40-person accounting firm, three weeks into its planning meeting:

   • Do Now #1: AI-drafted meeting prep summaries for client calls. Owner: Jenna (Senior Associate). Due: Week 2. Success measure: Prep time drops from 45 to 15 minutes across 10 meetings.
   • Do Now #2: AI first-pass on engagement letters, partner-reviewed. Owner: Mark (Ops Manager). Due: Week 3. Success measure: Zero client-facing errors across the first 15 letters.
   • Do Now #3: Meeting notes converted to action items automatically. Owner: Priya (PM). Due: Week 2. Success measure: Used in 90% of internal meetings by week 6.
   • Pilot: AI-assisted first-draft tax memo summaries. Owner: David (Tax Manager). Due: Day 45 review. Success measure: Time saved per memo, zero factual errors across 20 memos.
   • Policy action: Finalize the two-page AI acceptable-use policy. Owner: Dawn (Managing Partner). Due: Day 30. Success measure: Signed by all partners, distributed to staff.
   • Security action: Remove departed-employee access from shared drives. Owner: IT provider. Due: Day 21. Success measure: Access list matches current roster, zero exceptions.
   • Training action: 30-minute session on the policy and account types. Owner: Dawn. Due: Day 35. Success measure: Full attendance, staff can name two prohibited data types.

4. 4. Why that plan works

   Notice what makes this a real plan instead of a wish list: every success measure has a number or a plainly checkable fact in it. Nothing says "improve efficiency" or "raise awareness." A measure you can't fail to check in thirty seconds at the review meeting isn't a measure yet.

5. 5. Where these plans actually fail

   Every one of these is a planning failure, not an AI failure. They'd sink a hiring plan or a budget the same way. AI just gives them a new place to happen.

   • Padding: an eight-, ten-, or twelve-item plan because trimming felt uncomfortable. More items doesn't mean more progress. It usually means nothing finishes, because attention splits too thin to check on any of it.
   • The vague success measure: "improve efficiency" or "increase adoption" can't be failed, which means it can't really be checked either. If two people would disagree about whether it happened, it isn't a measure.
   • The orphaned owner: someone's name goes on a line item without them being told, or without the authority to actually do the work. The plan looks complete. Nobody's actually accountable for anything on it.
   • The plan that only exists in the meeting where it was written: nobody sends it to the people on it, so three months later half the room has forgotten it exists.

6. 6. How to actually run the 90-day review

   "Show up to the review date" isn't a process. A review that produces a decision, not just a status update, follows a short structure:

   • Go through all seven lines, not just the pilot. An unchecked security action is exactly as much of a finding as a failed pilot.
   • For anything not done, ask why in one sentence, without assigning blame. Under-resourced, deprioritized on purpose, or the owner never actually had authority to act are three very different problems with three different fixes.
   • Apply Briefing 5's measurement lens to the pilot specifically: time, quality, adoption, cost of mistakes. A pilot with a real answer to all four gets expanded. One with no answer to any of them isn't a failure yet, but nobody can tell, which is its own finding.
   • Draft the next 90-day plan in the same meeting. Scheduling a separate planning session almost always means a gap of weeks or months where nothing moves. Momentum here matters more than a tidy calendar.

Action plan

This is the last briefing in the 3rd Element Executive AI Briefing Series. Briefing 1 asked who's accountable for the results of AI already used inside your business. This plan is where that question gets a real answer.