The model behind an AI agent is becoming the least interesting choice in the stack
Enterprise AI agents now run on several models at once. Learn why value is shifting to permissions, audit trails and workflow design, and what fintech teams should build.
Key takeaways
- Enterprise AI agents increasingly run on several models and choose the best one per task.
- As models become interchangeable, value moves to permissions, audit trails, cost controls and workflow design.
- Fintech and wealth teams must define who an agent acts for, what it can see and where its audit trail lives.
- Long-running agent tasks need review checkpoints that a person can understand and act on.
Enterprise AI agents are increasingly model-agnostic: the platform picks the best large language model for each job, so the model itself becomes a replaceable component. For businesses building AI agents, the lasting value now sits in the layers around the model, including identity, permissions, audit trails, cost controls and the handover between people and agents.
Google's new work agent runs on both Gemini and Claude models, picks the best model for each job, and can keep going on a task for hours or days. That detail tells us a lot about where enterprise AI is heading.
What it means when the platform picks the model
When a platform chooses the model for you, the model becomes a component, like a database engine or a payment rail, and the value moves to everything wrapped around it. Teams no longer win by picking the right model once. They win by building the system that decides how an agent is allowed to act, what it can reach and how its work is checked.
The questions fintech and wealth teams need to answer
For the fintech and wealth clients we build for, that wrapping is where the real questions sit:
- Delegation: who is the agent acting on behalf of?
- Data access: what can it see, and what is off limits?
- Auditability: where does the audit trail live when a task runs for two days across five systems?
- Cost and governance: who sets spending limits and approves actions?
Built-in cost controls and governance are now part of the headline pitch for enterprise agents, which suggests buyers are already asking these questions. In regulated industries, they need clear answers before an agent goes anywhere near customer money or personal financial data.
Designing AI agent workflows that run for days
We think this shift is healthy for product teams. There are fewer debates about which model is best this month, and more time for workflow design, permissions and what a good handover between human and agent looks like.
A task that runs for days needs checkpoints a person can actually review. In practice that means breaking long tasks into stages, showing what the agent has done and why at each stage, and making it easy for a person to approve, correct or stop the work. Good agent design is closer to designing a reliable operations process than to writing a clever prompt.
Where lasting advantage comes from
This raises a question without a settled answer. If every platform can swap models underneath, where does a company's lasting advantage come from: the data it holds, the process it has refined, or the trust it has built with the people relying on the agent?
Our view is that it is likely a combination of all three, and that each depends on engineering decisions made early: how data is structured and governed, how workflows are documented and measured, and how transparent the agent is with the people it works for.
Building AI agents with Gunpowder Innovations
Gunpowder Innovations designs, builds and runs software products, including AI-powered products for fintech and wealth businesses. If you are planning an AI agent and need help with permissions, audit trails, integrations or the product experience around it, explore our AI development services or talk to our team.
Frequently asked questions
What is a model-agnostic AI agent?
A model-agnostic AI agent can use more than one large language model and select the most suitable one for each task. The model becomes a replaceable component, and the agent's behaviour is shaped mainly by the workflow, permissions and data around it.
Does the choice of LLM still matter for an AI agent?
It still matters for quality and cost, but less than it used to. When platforms can swap models underneath, the bigger differences come from how the agent is governed, which systems it connects to and how its work is reviewed.
How do you audit an AI agent that runs for hours or days?
Break the task into stages, log every action with the data and permissions used, and add checkpoints where a person can review, approve or stop the work. The audit trail should live in a system you control, not only inside the agent platform.
What should fintech companies check before deploying AI agents?
Check who the agent acts on behalf of, what customer and financial data it can access, how actions are logged for compliance, how spending is capped and how a person can step in. These answers should be designed in before launch.
- AI
- AI Agents
- Enterprise AI
- Fintech
- Product Design
