06

What AI changes in the weights

The criteria for the decision are the same as before: cost, speed, fit, risk, differentiation. What changed is how much each one costs and how much each one is worth. Some costs dropped sharply. A few things became much more valuable. Scrub from 2022 to 2026 to watch the weights move.
202220262026.0
Build+305
net shift in its favour
Buy+60
net shift in its favour
LLM+195
net shift in its favour
▼ Costs that collapsed
Cost to write code
2022: 90▼ 30

Coding agents write, test and refactor. For clear specs, implementation is no longer the bottleneck.

Cost of maintenance
2022: 80▼ 45

Dependency upgrades, framework migrations and bug triage become agent work. Ownership stays with you.

Cost of integration glue
2022: 70▼ 25

Mapping schemas, writing adapters and syncing systems gets cheap, so mixing vendors and your own code is easier.

Moat of thin SaaS
2022: 75▼ 30

A CRUD app with a nice UI can be rebuilt in days. Vendors whose moat is trust, data or network keep their value.

Cost per LLM call
2022: 85▼ 20

The price for a given capability tier keeps falling fast. Workloads that didn't pay off last year can this year.

▲ Things that got more valuable
Value of differentiation
2022: 55▲ 90

When everyone can buy or prompt the commodity, only the unique core sets you apart.

Value of proprietary data
2022: 45▲ 90

Models are commodities. Your context isn't. Data and domain knowledge decide who has the better AI.

Value of verification
2022: 40▲ 85

Tests, evals and review capacity become the bottleneck. Knowing it's correct matters more than writing it.

Value of clear specs & judgment
2022: 50▲ 90

When anything can be built, deciding what to build and what 'done' means is the scarce skill.

Value of trust & certification
2022: 60▲ 80

SOC 2, licenses, SLAs and a track record can't be generated. Regulated vendors get stronger.

The new rule of thumb
Build more than you used to

The line between “core, so build it” and “commodity, so buy it” has moved. Internal tools, glue and thin workflow apps are now cheap enough to own. Pricing engines, ledgers and your data model are still the parts that matter most.

Buy trust, not features

Pay vendors for things AI can't produce: licenses, certifications, networks, proprietary datasets and someone who carries the pager. Be more sceptical of seat-based feature bundles.

LLM where the spec is fuzzy

Use models for unstructured input and the long tail, and wrap them in evals and deterministic checks. Once a behaviour is stable and high-volume, turn it into code.

  • what didn't changeSomeone still owns it. AI makes code cheaper to write, but every line still needs an owner, an on-call rotation and a security review.
  • what didn't changeCorrectness where money moves. A ledger has to balance every time. Probabilistic output needs deterministic checks around it.
  • what didn't changeKnowing what to build. Agents make execution cheap, so choosing the right problem is where most of the value is now.