The “add AI” mandate: a CTO’s first 90 days

A version of this lands on most CTOs’ desks in 2026: “We need to add AI to the product.” It comes from the CEO or the board, it’s light on specifics, and it expects a roadmap before anyone knows which bets are real.

The trap is committing to a grand vision and a big build. The move is to sequence it so early, cheap wins fund and de-risk the expensive ones. Here’s the 90-day shape we recommend.

Days 1–15: Get to numbers before you get to roadmap

You can’t sequence what you can’t see. Two audits, both read-only, both fast:

Cost and usage. Instrument every LLM call — tokens, latency, cost, per route. Within days you’ll know what AI actually costs you today and where the waste is. This number anchors every later decision.
Readiness. Score your codebase and your team’s AI-tooling maturity. Are your engineers already shipping with agents? Is the codebase one an agent can work in? This tells you how fast you can realistically move.

You now have evidence instead of vibes — and something concrete to bring back to the board.

Days 15–45: Ship a win that pays for itself

Resist the urge to start with the moonshot. Start with the engagement that returns money or velocity in weeks, because that’s what buys you the room to do the bigger thing.

For most teams that’s LLM cost optimization: a documented 30–60% reduction in four to six weeks. It’s measurable, it’s defensible to finance, and the instrumentation you build for it becomes the foundation for everything else — including evals.

The second-best starting point is closing the evals gap. If you already shipped AI features blind, making quality measurable de-risks every change you make next.

Days 30–60: Make the build-vs-buy calls honestly

By now you have data, so the build-vs-buy conversation can be grounded instead of vendor-driven. A few rules that hold up:

Buy the commodity, build the moat. Use off-the-shelf models and tooling for undifferentiated work. Build only where AI touches your actual product edge.
“Why not just hire?” is a fair question — answer it honestly. You should build an internal AI platform team. Outside help is for shipping now and setting the patterns your future hires inherit, then transitioning out.
Avoid lock-in by default. Stay model- and vendor-agnostic so today’s cheapest-and-best choice can change next quarter without a rewrite.

Days 45–90: Stand up the durable layer

With a win banked and the calls made, invest in what makes the next year of AI work cheap instead of heroic:

Evals in CI so quality is a gate, not a surprise.
Observability for cost, latency, and agent behavior.
Safe internal access — an MCP layer — if your engineers need agents to reach company systems.
Self-hosted inference if data residency or regulation demands it.

None of this is glamorous. All of it is what separates teams that compound their AI investment from teams that keep re-litigating it.

What to tell the board

Not “we’re transforming with AI.” Say this instead: here’s what AI costs us today, here’s the win we’ll bank in six weeks, here’s the durable platform we’ll build with what it saves, and here’s the honest plan to staff it internally. Architecture, tradeoffs, and numbers beat a transformation narrative every time — especially with the people writing the check.

If you’re holding the mandate and the constraints, that’s exactly the conversation we have with CTOs. Bring both; we’ll give you a straight read.