Will Your AI ROI Case Survive at 5x? The Stress Test Your Board Will Eventually Run
Per-token AI prices are falling — roughly 30–50% per year, with frontier output prices down close to 90% since 2023. So why is your total AI bill going up?
The answer is the token-efficiency reversal: reasoning models and agentic workflows consume far more tokens per task, so consumption grows faster than unit prices fall. The unit gets cheaper while the task gets hungrier. That means the financial risk to your AI program is not a sudden price shock. It is under-modeled consumption — a heavy-tailed user base and agentic use cases that quietly multiply token volume long before any list price moves.
This is how AI programs fail the board-level cost review: not because token prices spiked, but because total spending was never modeled across its true cost layers or forecast against real consumption. Cost forecasting, not price speculation, is the governance priority. Here is the test.
The 2x/3x/5x effective-cost stress test
Model your AI program at 2x, 3x, and 5x today's modeled spend. At each scenario, ask one question: does the ROI case remain positive?
USDM recommends this stress test for every AI business case. The scenarios are driven primarily by consumption growth — more tokens per task as reasoning and agentic use expand — not by list-price increases. It is a resilience test, not a price prediction.
Business cases are 2–5x sensitive to changes in effective token cost, whether from consumption growth or price moves. A cost-resilient design keeps ROI positive across the range. If your case turns negative at 2x, that is a modeling problem you can find in one meeting — before your board does. Passing the test requires two inputs most cases skip: the full cost stack and a real consumption forecast.
Why the license line lies: the five layers of AI TCO
The most common estimating error is treating the license or per-token line as the cost of AI. In practice, it is the smallest of five layers — and in regulated life sciences, the lower layers routinely exceed it by a multiple. A defensible TCO model accumulates all five over a defined horizon, typically three years.
| Layer | What it covers | Why it gets missed |
|---|---|---|
| 1 — Platform & licensing | Per-seat subscriptions, platform and tenancy fees, consumption (API/token or credit spend) | It doesn't — this is the visible line, and the smallest |
| 2 — Data & integration | Pipelines, connectors, RAG corpus preparation, vector and storage infrastructure | Booked as an IT project, not the AI case |
| 3 — People & enablement | Implementation and solution engineering, administration, training, change management | Usually the largest hidden layer |
| 4 — Governance & compliance | Validation and qualification, risk intake, audit trail, monitoring, periodic re-validation | The life sciences premium — omitted until Quality asks |
| 5 — Run & sustain | Ongoing operations, drift monitoring, vendor management, incident response, decommissioning | The cost that never stops — and never makes the pitch deck |
Layers 4 and 5 are not optional polish in a regulated setting; they are the cost of being allowed to run the system at all. A business case that omits validation, audit trail, drift monitoring, and re-validation is not cheaper — it is incomplete, and it will be repriced upward the first time Quality or an inspector asks how an AI-assisted decision was controlled. The upside cuts the other way, too: governance built once amortizes across every subsequent use case, so marginal cost falls as the program scales.
Forecast consumption by persona, not average
Consumption is the one cost that is not a fixed contract — and the layer teams most often underestimate. A single blended per-user number hides the truth, because usage follows a bell curve with a heavy right tail: a small group of power users drives a disproportionate share of spend, so the mean sits well above the median. Budgeting on the average understates the spend driven by the tail.
Break the population into personas and sum them:
- Engineers and data scientists run coding agents over large data volumes and multi-step tasks — and can consume 20–40x the tokens of a general user. As a heavy-tail benchmark (not a typical cost), intensive coding-agent users can reach roughly $150–250 per user per month, while most users stay far lower.
- Power business users — analysts and operations — do document analysis, drafting, and extraction at moderate volume.
- General business users treat AI as super-powered search: quick questions, short summaries.
For the practitioner building the model, the forecast is bottom-up: users × working days × interactions per day × [(input tokens × input price) + (output tokens × output price)], less caching and batch savings. One structural fact shapes the result: output tokens are priced roughly five times as much as input tokens, so output volume — not input — usually drives the bill.
The three levers that keep the run-rate inside the envelope
Once the forecast is honest, the levers are concrete:
- Prompt caching can cut cached-input cost by up to ~90%.
- Batch processing saves ~50% on non-urgent work.
- Model routing — simple tasks to a low-cost tier, frontier models reserved for hard reasoning — spans a 5–25x cost range.
Because consumption is spiky, budget 30–50% of headroom over the point estimate and set per-team usage alerts. Readers who want to run their own numbers can model all five layers and a persona-based consumption forecast with the USDM AI TCO Calculator.
Model-agnostic architecture is financial risk mitigation
The most effective structural protection against AI pricing risk is architectural, not contractual. Model-agnostic design means the underlying model can be swapped without full revalidation or workflow redesign: abstract model-specific behavior behind a consistent interface, build prompt libraries that are model-portable, and maintain validated performance benchmarks that can be applied to alternative models.
For the CTO, the stakes are long-dated. Every architectural decision made today carries a 3- to 5-year consequence, and model lock-in — deep dependency on a single vendor's proprietary features — is the most common and most costly architectural mistake in enterprise AI. The organizations with the most strategic flexibility in 2028 will be those that made model-agnostic design a non-negotiable requirement in 2026.
The value side of the same equation
None of this is an argument against the investment — it makes the investment defensible. On the return side, well-implemented AI workflow tools recover an average of 11 hours per knowledge worker per week (Glean Enterprise Productivity Research, 2025) — but that benefit evaporates if AI literacy is insufficient for effective use and governance. The value case and the governance case are the same case, not competing budgets.
Run the test before your board does
Defensible AI ROI is designed, not discovered: model all five cost layers, forecast consumption by user persona, stress-test the case at 2x, 3x, and 5x effective cost, and keep the architecture model-agnostic so the case survives whatever vendors and models do next.
Cost clarity is one of the highest-value governance decisions a CxO can make in 2026 — and one component of the broader governance architecture in the CxO Guide to Sustainable AI. If you cannot yet run this stress test with confidence, start by establishing the baseline.
Take the first step with the USDM AI Governance Readiness Assessment — a current-state gap analysis, maturity scorecard, and peer benchmarks that pinpoint where to begin.
Sources: Adapted from the USDM white paper The CxO Guide to Sustainable AI (July 2026). Every figure and framework in this post — the token-price and consumption dynamics, the 2x/3x/5x effective-cost stress test, the five-layer TCO model, persona-based consumption forecasting, the three cost levers, and model-agnostic architecture — traces to that source. The 11 hours recovered per knowledge worker per week is cited to Glean Enterprise Productivity Research, 2025. The $150–250 per user per month figure is a heavy-tail benchmark for intensive coding-agent users, not a typical or average cost.