Strategy as a Sequence of Bets

Overview

The reference architecture says what shape things must be. It says nothing about order, and order is where strategy lives.

A technology strategy is not a list of things you will build. It is a small number of bets, stated as bets, each with the condition that would make you change your mind.

Why “bet” and not “plan”

A plan implies you know what will happen. In a domain where the available components change every few months, that implication is false, and everyone in the room knows it — which is why detailed multi-year technology plans are read as theatre.

Calling a move a bet does three useful things at once. It states honestly that the outcome is uncertain. It makes reversal a planned act rather than an admission of failure. And it forces you to write down what you expect, which is the only way anyone can later tell whether the bet paid off.

The cost is that it feels less authoritative than a plan, and some audiences want authority. The counter is that a bet with an explicit reversal condition demonstrates more competence than a plan with none — as long as you say so out loud rather than apologising for the uncertainty.

The format

Five fields. If a move cannot be written in this form, it is not yet a strategic move.

The bet — what you are committing to, in one sentence.
Why now — the driver or constraint that makes this the right moment.
What we expect — the observable outcome, with a number where one exists.
What would falsify it — the specific signal that says this was wrong.
Cost to reverse — cheap, moderate, or expensive, and why.

The fourth field is the one people leave blank, and it is the one that makes the document worth keeping. A bet with no falsification condition cannot be lost, only abandoned quietly — which is the organisational failure mode Module 1 described from the other direction.

Order by reversal cost

The ordering principle is simple and it is not “most valuable first”.

graph LR
  A["Cheap to reverse<br/><i>decide now,<br/>learn from it</i>"] --> B["Moderate<br/><i>decide when the<br/>first bet reports</i>"]
  B --> C["Expensive to reverse<br/><i>defer until there<br/>is evidence</i>"]
  C --> D["Irreversible<br/><i>decide once, with<br/>everyone present</i>"]

  classDef cheap fill:#7D9A4E,stroke:#5F7A38,stroke-width:2px,color:#ffffff
  classDef mid fill:#E8A33D,stroke:#C4832A,stroke-width:2px,color:#ffffff
  classDef dear fill:#D9695B,stroke:#B4483A,stroke-width:2px,color:#ffffff
  class A cheap
  class B mid
  class C,D dear

Four tiers by cost of reversal. Sequence runs left to right — not because the left is more valuable, but because deciding there buys information cheaply.

The reasoning: a cheap-to-reverse decision made now produces evidence at low cost. An expensive-to-reverse decision made now spends optionality you may need. So you front-load the moves that teach you something and defer the ones that lock you in until the earlier bets have reported.

Two honest limits on this principle. Some expensive decisions cannot be deferred — a residency constraint forces the serving topology before you have learned anything, and waiting does not help. And deferral is not free: a decision held open has a coordination cost, because teams downstream are blocked or hedging. Deferring everything is a recognisable failure mode of its own.

What this is not

It is not a feature roadmap with technology names in it. The test: every entry should be a choice you could have made differently, not a thing that obviously had to happen. “Build the retrieval path” is work. “Start with hosted inference rather than self-hosting, accepting vendor coupling to buy six months” is a bet.

It is also not a document that predicts the market. Resist writing bets about what the industry will do; write bets about what you will do and what would change your mind.

Three bets, with their reversal conditions

Bet 1 — hosted inference inside the region, not self-hosted. Why now: the enterprise contracts forbid data leaving its region, which rules out the cheapest hosted options but not all of them; self-hosting would demand GPU operations capability that 140 engineers fully committed to the migration do not have. What we expect: p95 time to first token under 1.5 s and blended cost under the $1.54 per-interaction ceiling. What would falsify it: provider cannot contractually guarantee in-region processing, or cost per interaction lands above $2.00 at realistic volume. Cost to reverse: moderate — the reference architecture keeps the provider behind an interface, so switching is weeks, not quarters.

Bet 2 — ship deflection first; defer the acting agent to the second release. Why now: the deflection case covers the 38% documented questions and needs no action surface, so it reaches customers while the oversight boundaries for credits are still being agreed with Lena Fischer and Dev Raghunathan. What we expect: measurable deflection on real traffic inside one quarter, and a golden-set baseline before any action is automated. What would falsify it: Priya Shah’s two-quarter revenue commitment turns out to require the agent actions to be sellable, in which case the order inverts and the compliance conversation becomes the critical path. Cost to reverse: cheap — it is a sequencing choice, not a structural one.

Bet 3 — small model first, escalate on low confidence. Why now: it is the largest single lever on the cost ceiling, and the ceiling is the CFO’s driver. What we expect: most documented questions handled by the cheap tier at a fraction of the per-interaction cost. What would falsify it: the cheap tier’s correctness on the 300-question golden set falls more than 5 points below the larger model’s, or the confidence signal proves too poorly calibrated to route on. Cost to reverse: cheap — routing is policy, and the fallback is to send everything to the larger model.

Two of the three are cheap to reverse and sit first. The expensive one is bet 1, and it is first anyway — because the residency constraint does not permit deferral. That is the limit of the ordering principle, stated rather than hidden.

Feeds the workbook

Your 12-month roadmap should read as a sequence of bets in this form, not a list of deliverables. The next two topics handle the two forces that will reshape the order: capacity, and the people who disagree about it.

→ The bets are ordered by what is cheap to undo. Now impose the constraint that actually governs: there are no free engineers. — Sequencing with no spare capacity