Ollama

Frontier model pricing is falling fast enough that the infrastructure layer, not the model itself, now determines whether an AI retainer holds its margin. Agencies that route client work through a gateway with caching, fallbacks, and per-client cost telemetry can reprice deliverables every quarter; agencies hard-wired to one provider absorb every price move and outage as a margin event. The strategic work is orchestration, not model selection.

Concentration Risk Ledger is a framework for tracking how much of an agency's delivery capacity depends on any single model provider, region, or price tier. The unit of analysis is not the vendor relationship but the retainer: for each client engagement, list which workflows break if one provider raises prices, degrades quality, or restricts access. Forrester warned in October 2026 that AI supply chains hide single points of failure in plain sight, and the same week Anthropic cut Claude Haiku 5.5 to $0.10 per million input tokens while OpenAI shipped GPT-6 to 1.2 billion weekly users, both reminders that pricing and capability floors move fast. An agency running every client summarization job through one API has an unpriced liability. The ledger converts that into a number: percentage of monthly delivery hours exposed, and the cost of a routing layer that reduces it.

Provider Concentration Ratio is the share of an agency's live client delivery that depends on a single model vendor, measured in billable projects rather than API keys. A retainer running three client builds on one provider's API looks diversified on an invoice and is not: a pricing change, a deprecation, or a guardrail shift hits all three at once. The ratio matters because infrastructure cost is a pass-through that agencies rarely reprice mid-contract, so a vendor move lands directly on margin. The practical test is to count distinct providers behind each deliverable and flag any client where one vendor carries more than half the inference load. Routing layers such as OpenRouter, Portkey, and Helicone exist to make that count adjustable without rewriting application code. The same discipline applies to compute: industry executives expect only 7 to 10 of hundreds of AI cloud providers to survive consolidation, so GPU capacity booked with a single neocloud carries the same concentration risk as a single-model stack.

Inference Cost Floor is the practice of tracking the lowest available price per million tokens for a capability tier, then treating every drop as a trigger to re-price client retainers rather than a windfall to bank. Agencies that price AI work on today's model economics get undercut the moment a cheaper tier ships, because the client's procurement team reads the same launch posts. Anthropic's Claude Haiku 5.5 arrived at $0.10 per million input tokens with a 1 million token context window, which resets what high-volume document summarization and campaign analysis should cost a client. The framework has three moves: benchmark your current blended cost per deliverable, set a review cadence tied to model releases, and pre-agree with clients that savings split rather than vanish. Agencies running fixed-fee AI retainers without a floor review are quietly donating margin every quarter.

Treat provider portability as a delivery requirement: put a gateway or routing layer in front of every model call, and price the migration path into the retainer before the first token is billed.

Stand up a gateway layer that logs, caches, and routes requests across at least two providers before any single-model contract becomes load-bearing for a client retainer.

IF your agency runs more than two client AI workloads in production and any single provider exceeds roughly 40% of inference spend, THEN build a routing layer that abstracts model calls behind one interface. IF client work is confined to one deliverable type, one model family, and under $2,000 monthly inference, THEN integrate the provider API directly and revisit the decision when either number doubles.

These three solve different layers of the same problem: routing and visibility, provider abstraction, and private inference. The lock-in risk named in this category is real, and the practical hedge is to keep the orchestration layer separate from any single model provider so a pricing change or capability shift becomes a configuration decision rather than a client-facing rebuild. Agencies should pick the layer that matches their current constraint, then revisit when retainer volume or client data rules change.

A productized engagement that puts a routing and failover layer between client applications and frontier model providers, so agencies can swap models on price or capability shifts without rewriting delivery code. The offer converts a one-provider dependency into a governed, observable, multi-vendor stack the client keeps paying a retainer to maintain.