career bridge
Build or buy
Cost three routes over three years, price the option nobody prices, and name the number that would change your mind.
You know the estate and what the use cases keep asking for. The next decision is the expensive one, and it is usually made badly, in a meeting, on the strength of whoever spoke last. Build the thing, buy it, or assemble it out of what the department already pays for.
What you do
Make the decision properly and write it so it survives the next change of mind. Pick the capability that repeated most across your use cases: retrieval over internal documents, a gateway in front of several models, an evaluation harness. Then cost three routes honestly across three years rather than one. Build, meaning engineering time to a first version, plus the part nobody counts, the maintenance and the being-on-call that follows forever. Buy, meaning licence plus integration plus the cost of leaving, which you should assume you eventually will. Assemble, meaning what the department already owns that does most of this badly, because that option is real far more often than anyone admits and it is always the cheapest to test. Then write the decision record: what you decided, the options you rejected, the assumptions it rests on, and above all what would change your mind. Name the specific number or event that would make you revisit it. A decision without that line hardens into dogma within a year, and whoever inherits it cannot tell reasoning from preference.
Done when
- One repeated capability is chosen, with the use cases that share it named.
- Three routes are costed across three years, including maintenance and exit, not just first delivery.
- The assemble-from-what-we-own option is genuinely priced rather than waved away.
- A decision record states the decision, the rejected options, and the assumptions underneath it.
- A specific number or event is named that would cause the decision to be revisited.
What you end up with
An architecture decision record on one build-or-buy call, costed over three years, naming the specific thing that would change your mind.
If you get stuck
The common failure is costing the build at time-to-first-version and the purchase at the licence fee, which flatters building and always has. Add the maintenance and add the exit. The other failure is skipping the assemble option because it is unglamorous and nobody gets promoted for it. It wins more often than people expect, and finding that out costs you an afternoon rather than two quarters.