What's the role of AI agents in fintech software engineering workflows?

Fintechs lead the rest of financial services on AI adoption. However, the IT teams that pull ahead don't use more AI. They run agents as an orchestrated team instead of one at a time.

Our engineers, here at The Software House, have been documenting that shift.

Here's what they've learned.

Are fintechs ahead on AI?

Here's the data behind that.

Cambridge's Centre for Alternative Finance surveyed the financial services industry for its 2026 Global AI in Financial Services Report.

It found fintechs leading incumbents by 47% to 30% in the adoption of advanced AI, and by 19% to 6% in reaching a transforming stage of adoption.

The Software House's CTO Adam Polak has been unpacking what that adoption looks like in practice across the issues of Effective Delivery – TSH’s newsletter.

He's been documenting how TSH's IT teams moved from single AI tools to coordinated agents running stages of the pipeline.

That’s exactly how top fintech companies are using AI.

One AI agent, or a team of agents?

IT teams adopting AI coding tools can hit a ceiling and assume that's as good as it gets

In The age of multi-agent AI engineering, Adam lays out why.

A single developer running one AI agent on one task at a time gets a real but capped productivity boost of 20 to 40%.

However, when the same developer can use one agent to command other, more specialized agents, the productivity boost is, at least in theory, unlimited.

Adam describes 3 levels of agent use as:

• sequencing task-specific agents by hand,

• letting an orchestrating agent choose which to run,

• and running agent groups in parallel.

Buying licenses for a coding assistant is level 1, and it looks nothing like the role AI agents play in a mature workflow.

For top fintech companies in particular, their AI use focuses on managing layers of AI agents.

What if AI agents write all the code?

Handing all the coding to AI agents sounds like a clear win, but TSH's first attempt ran into problems.

Review: Our first 100% AI-driven project covers the first time TSH let AI agents handle all the coding for a commercial client.

The project exposed 3 bottlenecks outside of coding:

• a backlog that didn’t have any more tasks,

• a slow code review process,

• and merge conflicts.

A single developer backed by AI agents closed tickets faster than the business analyst could write them or a reviewer could check them.

The backlog ran dry, reviews piled up, and merge conflicts became frequent.

The fix wasn't to slow the agents down.

Instead, the team extended automation to the rest of the pipeline with a backlog agent, automated code-quality checks alongside human review, and fewer issues left for manual review.

The lesson for fintech engineering is that a fast pipeline next to a slow, manual compliance process moves the bottleneck instead of removing it.

Luckily, in fintech, where structure matters greatly and compliance and review can't be skipped, embedding AI at every stage of development in a predictable way is actually a natural thing to do.

What does an AI agent workflow look like?

It helps to look at a real agent-driven workflow instead of treating "AI agents" as one thing.

Another issue of Effective Delivery by Adam Polak walks through Copilot Collections, the framework TSH built to standardize how engineers work with AI.

It's built around 6 agents, such as a Business Analyst that clarifies requirements, an Architect that drafts a solution blueprint, and a Code Reviewer that reviews before a human sees the pull request.

Each agent hands output to the next in 4 phases of Research, Plan, Implement, and Review.

Engineers using it saw a 30 to 40% productivity gain compared to working without it, and TSH uses it in about 70% of its projects.

The same principle of speed from structure applies to fintech delivery too.

When xpate needed to cut payout processing from days down to minutes, the TSH team didn’t focus on skipping steps.

Instead, it combined serverless design with built-in observability using a highly structural approach to problem-solving in development.

How do you prep engineers for AI agents?

Another Effective Delivery issue describes the adoption gap in AI TSH had to close first to make more advanced AI use possible.

At the outset, a mere 10% or so of engineers used AI for code in a meaningful way.

Switching from ask mode to agent mode was one of the main steps that took that share to around 70% eventually.

An IT team that hasn't made that switch isn't ready to discuss orchestrating agents, since most of its engineers work a level below where that conversation starts.

Is AI adoption enough for an advantage?

Fintechs are ahead on adoption, but adoption and maturity aren't the same thing.

• A single AI agent per task gives a real but capped boost,

• Speeding up coding without speeding up backlog, review, and QA moves the bottleneck,

• The biggest gains for fintechs come from structured, multi-agent workflows, not from more tools.

TSH's engineering team has been running this experiment on its own and working with fintech companies.