Microsoft Copilot

Frontier assistants are not a headcount decision, they are a delivery-cost decision. The gap between agencies that route work through a standardised assistant stack and those that do not shows up as margin per retainer, not as a hiring freeze. Model choice now moves both invoice cost and output quality on the same client job, so the routing decision is a pricing decision.

Frontier assistants change month to month, so any agency workflow hard-wired to one model version carries a hidden rework tax: prompts, evals, and client-facing claims that must be retested every release cycle. The framework asks a simple question before you standardise: how much of this deliverable survives a model swap? Workflows that survive are the ones worth productising into retainers; workflows that break are experiments, not delivery lines. OpenAI's October 2, 2026 guide splits the GPT-6 family into three tiers for prototyping, feature development, and multi-step orchestration, which means a single client workflow can now span multiple models with different cost and quality profiles. Anthropic's Claude 5.5 release drew independent scepticism about announcement hype, and a September 2026 technical analysis found top benchmark scores do not reliably predict production performance. Agencies that map churn exposure per deliverable can quote model-agnostic retainers instead of absorbing every release as unpaid rework.

Frontier assistants change month to month, so any agency capability claim has a shelf life measured in weeks, not quarters. The Capability Drift Window is the gap between when a model ships and when your client-facing promises stop matching what the tool actually does. Agencies that treat model releases as procurement events (buy once, standardise, move on) accumulate silent drift: the retainer scope describes work the assistant could do six months ago, while competitors quote the current ceiling. The fix is a scheduled re-scoping cadence tied to release events, not annual reviews. When Anthropic shipped Claude 5.5 in September 2026, the practical question for agencies was not whether to switch, but which existing deliverables could now be quoted at a higher margin or lower hour count. Forrester's Q3 2026 research puts workflow integration, not model capability, as the bottleneck, which means drift shows up in delivery process before it shows up in output quality.

Seat cost per deliverable measures what an agency actually pays for each unit of client work produced on a frontier assistant, not what the subscription line item says. A $30 seat looks cheap until three people touch the same retainer deliverable and the true cost lands at $90 plus review time. The framework forces a simple calculation: monthly seat spend divided by completed, client-accepted deliverables, tracked per role. It matters because model releases reset the math every few months. OpenAI's October 2, 2026 GPT-6 guide splits the family into three tiers for prototyping, feature development, and multi-step workflow orchestration, and routing a job to the wrong tier inflates cost without improving the output a client sees. Agencies that track this ratio can move copy work to a cheaper tier, keep orchestration on the expensive one, and defend retainer margin when a client asks why the AI line item grew.

Route each deliverable type to the model tier that matches its complexity, then standardise only the interface your team uses daily.

Map the deliverable to the workflow step before committing seats, then buy the assistant that survives your own review process.