The team picks a model in January and runs everything through it for the rest of the year. It classifies tickets, drafts emails, extracts invoice fields, and answers customers with the same engine. It works, so nobody touches the configuration until the quarterly bill arrives and someone asks why it costs that much.
Capgemini Research Institute published its World Cloud Report for Financial Services 2027 on November 5, based on a survey of 1,100 executives in financial services. One number captures the shift: 90% of institutions use or plan to use intelligent model routing, meaning they send each task to the most cost-effective and accurate model for that specific task. When the tightest-margin industry in the economy puts routing at the center of its strategy, you should ask what they know that you haven't acted on yet.
One model for everything gets expensive
Your processes mix tasks of very different difficulty. Deciding whether an email is a question or a complaint is something a small model handles in milliseconds. Drafting the reply to that complaint, with the right tone and the order data in hand, sits at another level. And deciding whether a customer deserves compensation calls for reasoning and judgment.
If everything runs on the most powerful model, you pay reasoning prices for work a small model would do well. If everything runs on the cheap model, errors show up: flat responses, hallucinated order details, broken formats that someone fixes by hand afterwards. The Thoughtworks Technology Radar (volume 35, November 2026) sums it up well: a cheaper model used carelessly can cost more than an expensive one used well.
Agents multiply the problem. An agent loops through thinking, calling tools, and correcting itself several times per task. If every step burns the expensive model, the bill grows with each retry. That is where routing stops being a technical nicety and becomes the difference between a profitable agent and one you have to switch off.
How to build a simple route
First, catalog your tasks. Write down what your system does today and group the work into three tiers: mechanical tasks (classifying, extracting fields, formatting), drafting tasks (emails, summaries, replies), and judgment tasks (decisions, analysis, sensitive content). This table is the heart of the system, and you write it yourself, with your team, in an afternoon.
Second, assign a model to each tier and define escalation. Mechanical work goes to the small model. Drafting goes to a mid-tier one. Judgment goes to the powerful one. Add a fallback rule: when the small model lacks confidence in its answer, escalate to a stronger one. A well-tuned fallback is what keeps the savings from getting expensive in corrections.
Third, measure per task and revisit the table every few months. Providers change prices and ship new models without warning, and the route that is optimal today stops being optimal within a quarter. Keep the record of every completed task: cost, time, and whether someone had to fix the output. With that data, the review happens over numbers instead of opinions.
The most common mistake
Technical teams sometimes overbuild routing until it becomes unmaintainable. Rules for every corner, a model per language, logic only its author can read. Routing holds up only if anyone on the team can open the table and understand it in five minutes. Start with three tiers and don't add a fourth until the data asks for it.
The Capgemini numbers also suggest most companies haven't finished the journey: only one in ten surveyed institutions has reached full orchestration of their agent systems. You still have room to start now and learn ahead of your competition.
At Luxion Solutions we build agents and automations that reach production with the model route defined from day one: cataloged tasks, escalation rules, and metrics per completed task. If your AI bill doesn't match the value you get out of it, drop us a line and we'll take a look at your case.
