When Does a Cheaper AI Model Cut Workflow Costs?

· Decision Intelligence
China Strategic Signals

Before switching models, measure what it costs to get a case to the right team on time, including retries, escalations, and human review.

Imagine an AI contract renewal meeting where the finance team sees a lower model bill. The support team sees more tickets that need a second look. Supervisors are spending more time reviewing borderline cases, and a few urgent requests have landed in the wrong queue. Before signing the renewal, the CFO and CIO need an answer both teams can verify: After those extra checks and corrections, does it cost less to route a ticket correctly and on time?

A Capgemini Research Institute survey published in November 2026 shows why model costs are under scrutiny. Of the 500 senior executives at large companies surveyed in September, 59% said their AI spending over the previous year was significantly higher than expected or budgeted. Nearly three in five surveyed organizations planned to expand their use of open-weight models. For a team considering a switch, the savings still have to be measured on its own workflow.

Where a cheaper model can earn its place

Airbnb’s engineering team compared models on a classification task already used in production. The task was to flag content that violated its pricing and payment transparency policy. On 10,000 weighted test examples held out from training, a small, fine-tuned model reached an F1 score of 0.902, close to the best proprietary models tested. At thousands of requests per minute, Airbnb estimated that serving the classifier cost about 1% as much as using the leading model APIs. At low volumes, some of those APIs were cheaper than renting a dedicated GPU. The estimates cover model serving, and the economics depended on Airbnb’s traffic and pricing assumptions.

AT&T illustrates a different choice. It used three models to classify customer service call summaries, sending the hardest cases to a larger model. An AT&T data scientist said at a conference that the new setup cost about 35% of what AT&T had been paying to use ChatGPT, with 91% relative accuracy. A day’s summaries took under five hours to process, down from 15. The Business Insider report does not explain how relative accuracy was measured or how much labor went into reviewing and correcting classifications. Before making a similar switch, check which kinds of tickets the cheaper setup sends to the wrong queue. A ticket sent to the wrong general queue may be easy to reroute. An urgent ticket sent to the same queue may miss its deadline.

Measure the work that passes review on time

Start with a representative batch of tickets and compare three setups: the current workflow, the cheaper model on its own, and the cheaper model with a route to the current model or a human reviewer. Set the acceptance rules before running the test. A ticket might count as complete when it reaches the correct queue by the deadline, with a separate limit on missed urgent cases.

For each setup, divide the total cost of handling the batch by the number of tickets routed to the correct queue by their deadline. Count what the team spent on every ticket, including those that failed review, arrived late, or remained unresolved. Include model calls, infrastructure, retries, escalations, human review, and rework. Spread the cost of labeling, fine-tuning, and switching over the volume you expect to handle, and include ongoing maintenance. Report the on-time acceptance rate and the rate of critical errors alongside the cost. That way, a cheaper average cannot hide a deterioration in service quality.

The definition of success can change which option looks cheapest. In Microsoft’s ThinkingBox evaluation, GPT-5.6 Sol had a slightly lower estimated inference cost per successful attempt than GPT-5.4. When the researchers measured cost per task that passed all 20 runs, GPT-5.4 was cheaper. That benchmark used simulated workflows and excluded human labor and downstream failure costs. Its relevance to procurement is straightforward: choose the reliability standard the job requires before ranking models by cost.

Calculate how much escalation the savings can absorb

Consider an illustrative case. The current model costs $10 per incoming ticket. The cheaper model costs $4 on the first call, with another $1 per ticket for switching and ongoing maintenance. That leaves $5 per ticket to cover extra work. If an escalation adds $25 on average, both options have the same on-time acceptance rate, and all other costs are equal, the cheaper setup breaks even at a 20% escalation rate. Add $2 in extra review costs per incoming ticket, and the break-even rate falls to 12%. If the setups route different shares of tickets correctly and on time, compare their costs per ticket routed to the correct queue by its deadline.

At a renewal or rollout review, approve the switch for the ticket categories where the cheaper setup meets the quality and deadline requirements and still costs less after escalations and review. If the savings hold only for routine tickets, move those tickets and keep the existing route for harder cases. Recheck the decision when ticket volumes, model versions, or review workloads change enough to affect the comparison.

Sources

Section image

Turning Global AI Signals into Business Decisions

InfoAI helps decision-makers identify what is changing in AI, evaluate the evidence behind those changes, understand what they mean, and determine which business assumptions or decisions may need to be reconsidered.