Don’t Use a Ferrari to Make a Grocery Run (or, Tips for Building Cost-Effective AI Models)

Let’s get one thing out of the way: AI is not optional unless your strategic plan includes gradual irrelevance. 

It’s changing how work gets done, who does it, and how fast they get replaced if they don’t figure it out. So yes — companies are right to push adoption. You want your people using it, learning it, breaking things with it.

But we’ve officially entered the phase where usage is being treated like value. Tokens, the fundamental word part units of language intelligence, are the new currency, and they’re being tallied. Prompts are being counted. Someone, somewhere, is absolutely putting “AI hours logged” into a performance review. And while the intention may be good, the execution can get expensive fast — one developer recently shared that his company accidentally burned through $18,000 a month in unexpected AI usage. 

But is this just the latest remake of a classic movie? Take another major disruption, electricity for example. Electricity didn’t just improve business — it redefined it. Factories untethered from steam, workdays extended past sunset, entire industries rewired around a new capability. It was genuinely revolutionary. But then the rush was on to apply it to everything, good and bad. Electricity gave us the electric light and the electric can opener. One transformed civilization. The other added a cord to a solved problem.

AI can be just as energizing, but having a grounded assessment can help direct it to its full potential. So how can AI be a gain instead of a drain? Let’s take a look at two cases.

Generative AI is incredibly good at helping people ask better questions. You can type in a messy, half-formed thought — “How are deposits trending?” — and it will translate that into something meaningful. Generative AI connects dots across systems, fills in gaps, and gives business users access to answers they didn’t have before. And it usually complements you for asking a smart question in the first place.

But consider how this plays out in the real world. A regional bank used generative AI to help summarize lengthy committee reports before board meetings. Executives loved it because it cut hours of reading down to a few minutes. But the bank noticed employees were repeatedly asking the AI for the same performance metrics and portfolio summaries, week after week. So, the bank converted the most common requests into permanent dashboard reports using standard database queries.

The result? Executives still got fast answers, but the bank reduced AI usage costs dramatically.

The lesson: Use AI to discover what matters. Then build a system so you don’t have to rediscover it every day. The first answer is the value. The hundredth time you ask the same question? That’s just a meter running.

Where else might token-overkill exist? Consider a new default assumption creeping into how companies think about AI: that the end state is an “agentic workflow” — systems that think, decide, and adapt at every step.

It sounds modern. It certainly demos well. And it has its place in certain situations. But think about how Anthropic and OpenAI market their latest models as having PhD-level reasoning. That's a fantastic asset to take advantage of. But nobody hires a PhD to be a filing clerk. You don’t bring in the most advanced reasoning system available and then ask it to perform routine, repeatable tasks that already have a known answer. That’s not leverage — that’s misuse.

On top of that, just like an absent-minded PhD professor, when the AI makes decisions, you can enter the “Transparency Gap,” as written about by Tristan Green in a recent GonzoBanker article (3 Strategic Blind Spots Emerging With AI Adoption).

What should truly transform with AI is how you build it. That’s where most organizations are leaving value on the table. Right now, a lot of teams are using AI as a worker — something that executes tasks over and over again. That’s useful. But the real leverage is that AI can finally make certain systems buildable. Workflows that were historically too painful to define, too complex to document, or too expensive to implement are now within reach. For example, with generative AI, you can:

• Turn policy documents into executable rules
• Map messy processes into structured workflows
• Build the scaffolding — forms, data models, integrations — that used to take months

In other words, AI doesn’t replace the system. It makes it possible to build the system you always wanted in the first place. That’s the shift.

And yes, AI should be part of the process as well — the messy part of a process like understanding the emails, the documents, the ambiguity. Let it do the part humans used to struggle with. But run it within a system that runs cleanly and predictably for all steps.

The pattern isn’t “AI everywhere.” It’s:

• AI where the work is uncertain, and
• Systems where the work is known

If everything in a daily workflow thinks for itself, it’s probably thinking too much — and costing you even more.

So, what can bankers do to keep the AI-juice flowing in the right direction? Start by having a plan for your vendors embedding generative AI into their platforms. Many AI platforms are optimized for maximizing large language model (LLM) involvement across an entire workflow. That can be powerful — but it can also be expensive, opaque, and difficult to govern if applied indiscriminately.

A more pragmatic approach should be sought by asking:

• How does this platform account for token usage and cost at scale?
• Does it allow AI to be used selectively within workflows?
• Can workflows be executed deterministically without requiring an LLM at every step?
• Is there a path from AI-assisted design to non-AI runtime execution where appropriate?

The vendors that get this right aren’t trying to electrify everything. They’re building systems that know when to use the power — and when to turn it off.

Because the goal isn’t to wire your entire business to a meter. The goal is to build something that runs — without charging you every time it thinks.

Judd Heckman is an AI architect at Cornerstone Advisors. Follow him on LinkedIn. John Meyer is a managing director at Cornerstone Advisors. Follow John on LinkedIn.