The window that decides the next two years
Every AI program has a first 90 days. The decisions made and the patterns set in that window shape what’s possible for years afterward. Programs that get the first 90 days right usually compound; programs that don’t usually struggle to recover.
Most companies waste the first 90 days on strategy theater — decks, all-hands speeches, vendor selection processes, internal town halls. None of that work ships AI. By day 91, the team has talked about AI more and built less of it than they planned.
Below is the 90-day playbook we use to start a program well. It’s organized by two-week segments, with specific outputs each segment. The whole arc produces a program that has shipped one production thing, established the operational foundation for everything else, and has a year of credible roadmap.
Days 1–14: Listen and assess
The first two weeks are about understanding what’s actually true, not what the executive team thinks is true.
What to do:
- Sit in the executive team’s weekly meetings.
- Sit in the engineering team’s standups.
- Read the data warehouse schema.
- Read the security policy.
- Inventory every existing AI surface (vendors, pilots, internal tools).
- Talk to the head of engineering, the head of ops, the head of customer success, the head of finance.
- Take no positions. Make no commitments. Just understand.
What to ship: the inventory document and a written 5-page assessment of where the program stands today, with no judgment yet on what to do next. This document becomes the baseline that everything subsequent builds on.
Common mistake: committing to specific initiatives in week one to demonstrate momentum. The cost of these commitments compounds; once the program has direction, redirecting is expensive.
Days 15–28: Define the program
With two weeks of listening done, the next two weeks are about setting the program’s shape — not yet the specific roadmap, but the structure that the roadmap will fit.
What to do:
- Set the strategic priorities (3-5 statements, one sentence each).
- Decide the team shape (federated vs CoE; how many engineers; who they report to).
- Decide the platform strategy (multi-provider, self-host posture, gateway architecture).
- Decide the governance posture (the five governance documents).
- Schedule the meeting cadence (monthly roadmap review, quarterly planning).
What to ship: a 1-page program charter that defines the priorities, team shape, platform strategy, and governance baseline. Distributed to executive team and relevant operating leaders.
Common mistake: spending these weeks selecting vendors. Vendor selection comes later, after the program has a clear posture. Picking vendors before posture leads to vendor-driven architecture.
Days 29–42: Build the foundation
Now the engineering work starts. The first build is not a customer-facing feature. It’s the foundation that all subsequent features will run on.
What to do:
- Stand up the model gateway with at least two provider integrations.
- Set up observability (decision traces, cost meter, basic eval harness).
- Configure the first set of credentials and access controls.
- Create the deployment infrastructure for AI services (containers, networking, secrets).
- Pick the first internal-only AI initiative for weeks 43–70.
What to ship: a working internal “hello world” through the gateway. Cost is metered. Observability is captured. A real model call goes through the entire stack and produces a logged output. The platform is alive.
Common mistake: trying to ship a customer-facing feature in this window. Customer-facing features fail when the foundation isn’t there. Build the foundation first.
Days 43–56: Ship the first internal thing
The first production AI work should be internal — high tolerance for imperfection, fast iteration, learning the operational patterns. The customer-facing work waits.
What to do:
- Build the first internal initiative end-to-end. Knowledge agent, drafting assistant, internal research tool — whatever fits the company.
- Use the full operational stack (eval suite, monitoring, governance checks).
- Roll out to a small group of internal users.
- Gather feedback. Iterate.
What to ship: an internal AI tool used by 10-20 employees daily. Real, in production, with cost and quality both tracked. Not a demo.
Common mistake: picking an over-ambitious first initiative. Pick something small. The point is to run the entire workflow end-to-end and learn. Ambition comes in the next initiative.
Days 57–70: Operationalize and learn
Two weeks of running the first thing in production. Most of the program’s institutional learning happens here.
What to do:
- Run the internal tool. Triage feedback. Improve it.
- Conduct the first post-mortem on whatever breaks (something will).
- Refine the eval suite with real production examples.
- Document the operational patterns that worked and didn’t.
- Begin scoping the next two initiatives (one slightly more ambitious internal, one customer-facing).
What to ship: a “patterns we’ve learned” document that captures the operational lessons. This document becomes the institutional knowledge for all subsequent work.
Common mistake: declaring victory after the first thing ships. The first thing is the foundation, not the achievement. Treat it as the start, not the finish.
Days 71–84: Plan the year and scope the next initiatives
With foundation in place and one thing running, the program now plans real. The strategic priorities from week 4 become a concrete roadmap.
What to do:
- Build the quarterly roadmap for the upcoming quarter and rough sketch for the next 3 quarters.
- Scope the next 2-3 initiatives in detail.
- Run the first quarterly planning meeting with the executive team.
- Set the budget for the next quarter.
- Identify any additional hires needed for the planned work.
What to ship: a 1-page quarterly roadmap with named initiatives, owners, kill criteria, cost bands. The same shape the program will use indefinitely from this point.
Common mistake: committing to too many initiatives in the first quarter. The mature shape is 3-4 initiatives. The temptation is to commit to 6-8 to show ambition. The temptation produces stalled initiatives.
Days 85–90: Communicate and commit
The final week is about externalizing the program to the wider organization.
What to do:
- Present the program charter and roadmap to the broader executive team.
- Send a written communication to the company about what the program will and won’t do in the next quarter.
- Set up the monthly status update format.
- Schedule the next quarter’s review meetings.
What to ship: the public version of the program. Everyone in the company knows what’s happening, what’s coming, who’s responsible.
Common mistake: treating this as a big-bang launch. The program has been working for 90 days. The communication is acknowledgment, not announcement.
What this produces by day 91
By the end of the first 90 days, a well-run program has:
- A 1-page program charter, alive and current.
- A 1-page quarterly roadmap with 3-4 specific initiatives.
- One internal AI tool in production with real users.
- A working platform (gateway, observability, eval harness).
- The governance baseline in place.
- A team that knows how to ship and operate.
- Executive alignment on direction.
- Patterns documented for the next initiatives.
This is dramatically more than most companies have after 90 days. Programs that have this foundation move faster for years; programs that don’t are often still talking about strategy at month 12.
What this doesn’t produce
Worth being clear about what isn’t done by day 91:
- The customer-facing AI feature isn’t shipped yet. That’s quarter 2 work.
- The full team isn’t hired yet. Hiring continues over months.
- The governance posture isn’t audited. That’s quarter 2 or 3.
- The vendor relationships aren’t all renegotiated. That’s ongoing.
The first 90 days set the trajectory. The actual ambitious work happens in months 4-12 against the foundation.
The fractional CAIO pattern
If the program is being started without a full-time CAIO, the first 90 days is a natural fit for a fractional CAIO. The work is concentrated, requires senior judgment, and benefits from someone who’s done it before. After 90 days, the program either has the foundation to hand off to a less-senior owner, or it makes sense to extend the fractional engagement.
The cost of fractional leadership in the first 90 days is typically a fraction of a full-time hire and produces dramatically better foundation than a self-started program. The math favors fractional in the startup window.
The take
The first 90 days of an AI program set the trajectory for years. The playbook above — listen and assess, define the program, build the foundation, ship the first internal thing, operationalize, plan the year, communicate — produces a program that compounds. Programs that start without this discipline spend month 12 building what they should have built in months 1-3. The 90-day window is bounded. The compounding effect is unbounded.
Starting a program well is one of the most common shapes of Fractional CAIO engagements. If you’re a few weeks into a new program and want a second opinion on direction, schedule a call.
