Subject: Q1 2027 planning: three tests before a specialist-model commitment
Preview: Prepare for changed obligations, uncertain economics and a continuing need for buyer-owned evidence.
Issue: Planning draft for Q1 2027, based on information checked 26 October 2026. Refresh before circulation in 2027.
The central view
The strongest purchase is a bounded commitment tied to evidence, not a belief that specialist models always become cheaper or more capable. Plan for several possible outcomes: retain the current API, improve the surrounding workflow, route part of the traffic, or deploy an adapted model that passes the same requirements.
This is the desk’s planning judgement. It is not a market-growth forecast or investment recommendation.
Test 1: will the governance plan match the actual product use?
CPPA’s final-rule announcement identifies January 2027 as the compliance start for its specified significant-decision ADMT requirements. Colorado’s current law also has January 2027 requirements. These are separate regimes with different scope; they are not a universal fine-tuning approval checklist.
Preparation: map the product’s actions, affected people, jurisdictions and processing. Give counsel concrete workflows and supplier responsibilities, not just a model name. Update contracts and evidence schedules where required. Verify final rules and any intervening legal developments before launch.
Test 2: does the economic advantage survive plausible downside cases?
Use current published prices or actual quotes and keep them dated. Then stress-test lower volume, a cheaper API alternative, more human review and a required redundancy layer. Those changes are scenarios, not claims about what prices or demand will do.
Preparation: separate operating savings from recovery of the initial investment. If the cheaper path fails quality or the downside case needs unbudgeted engineering, shorten the commitment or buy more evidence before proceeding.
Test 3: can the team keep evaluating after launch?
The 2026 evaluation guidance emphasises that automated tests, monitoring and human review serve different purposes. A final pilot score cannot replace an operating process when documents, policies and models change.
Preparation: assign the regression-suite owner, incident owner and authority to roll back. Define triggers for reassessing the baseline. Retraining is one possible response to a diagnosed problem; replacing a prompt, corpus or provider may be more appropriate.
What would change this outlook?
An actual buyer-controlled win with stable whole-life economics would justify a larger commitment. A stronger API baseline, lower demand, new legal constraints or loss of an operational owner would justify pausing or narrowing it. Write these conditions into the decision memo before award so reassessment is an expected part of the purchase.
Suggested call to action: Take the readiness quiz with product, engineering and finance; resolve the hard gates; then circulate one common pilot brief to potential partners. Keep the option not to build explicit.