Work Decomposition Has Become the New Model Strategy

The clearest August 23 signal came from Digg AI and Digg Tech picking up a new round of discussion around the business ceiling on frontier-model pricing. Ramp’s August 12 AI Index update provides the primary quantitative signal: businesses are still increasing AI spend overall, but Fable 5 uptake fell below expectations and cheaper model-serving routes continue to gain ground. Simon Willison’s August 23 note on the Financial Times reporting connected that market signal to real product behavior by calling out how expensive frontier usage is being constrained by what teams will actually pay for in practice. Drew Breunig’s August 23 essay then makes the builder-level lesson explicit: the free lunch is over, which means teams now have to think carefully about what work goes where. The useful synthesis is that model strategy is moving up the stack. The important design decision is no longer only which model is best, but how the workflow itself is partitioned by value density.

A lot of teams still treat model choice as a single top-level decision: pick the smartest model available, ship the feature, and wait for future price drops to fix the economics. That worked when capability jumps were arriving fast enough to mask sloppy workflow design. It works much less well when a premium model is great but not great enough to justify being the default for every subtask. The failure mode is subtle: teams measure “it works” while margins erode, latency drifts up, and users stop seeing enough quality lift to justify the cost. Builders who decompose the job can keep frontier reasoning where it changes the answer and move rote drafting, transforms, classification, and cleanup to cheaper paths. Builders who do not will mistake model quality for workflow quality.

Refactor one AI feature around task classes instead of one default model. Trigger: choose a workflow with mixed cognitive load such as coding, support drafting, research synthesis, GTM prep, or document operations. Context: break the workflow into moments that actually change outcome quality, such as planning, ambiguity resolution, tradeoff analysis, or architecture choice, versus routine steps like extraction, formatting, rote edits, and status updates. Tools: pair a premium model with a cheaper execution model, keep a structured handoff format between them, and log accepted-run cost by substep rather than only by full request. Verifier: use tests, grounded citations, schema checks, rubric review, or downstream state confirmation so cheaper execution cannot silently degrade the final artifact. Budget: set ceilings by step class, not just per request. Artifacts: store the decomposition map, route decisions, verifier outcomes, accepted-run cost, and the classes of work that still need premium reasoning. Stop condition: promote a cheaper route only when it clears the verifier within budget across real traffic, and move work back up-market when failure or cleanup cost exceeds the savings.