New: What is engineering drift costing your organization? Put a number on it in 2 minutes:
New: What is drift costing you? Find out in two minutes:
November 17, 2026

Context Before Scale: Preparing Banks and Fintechs for Agentic AI

Gina Dragulin | CMO at DesignVerse

Our days at CTO Craft Fintech in London gave me a lot to think about. One question from the floor stayed with me longer than any of the talks: in a regulated environment like banking, should leaders prepare for a future where AI writes, reviews, and deploys almost all of the code?

It's the question most engineering leaders in financial services are asking, and it came up again over the DesignVerse-hosted dinner with technology leaders from across the sector. The answer in the room wasn't a simple yes or no. For a regulated business, the goal isn't to be first with every new AI tool. It's to build an engineering environment where AI can operate safely, effectively, and at scale. That work takes far longer than most leaders expect.

My take: the race isn't to go first. It's to be ready first.

Because the future of AI engineering will depend less on the models organizations choose, and more on the quality of the engineering environment those models operate within.

Key Takeaways

  • The market has moved beyond AI coding assistants to agentic engineering, where AI agents take on work across the whole software development lifecycle.
  • AI adoption isn't productivity: strategy, training, and ways of working decide whether AI spend turns into value.
  • Regulated engineering teams don't need to be first with every AI tool, but they do need a new engineering operating model, and building it can't wait.
  • AI needs governed organizational knowledge, not just access to documents: architecture, standards, policies, past decisions, and ownership.
  • As AI writes more of the code, the competitive advantage shifts from producing code to preserving and governing the knowledge behind it.
  • In financial services, AI readiness sits alongside DORA, the EU AI Act, and existing accountability obligations.

The Market Has Moved Beyond AI Coding Assistants

Most AI engineering conversations are still about coding assistants. The market has already moved on. Broadly, it has gone through three phases:

  • AI experimentation. Individual developers adopting copilots, generating code, writing tests, and experimenting with prompts. This is what most teams spent the past year doing.
  • Enterprise adoption and governance. Approved tools, training, security controls, cost management, and productivity programs. This is where most organizations are now.
  • Agentic engineering. AI agents taking on increasingly complex development tasks: coordinating workflows, reviewing code, and interacting directly with development infrastructure. This is the emerging frontier.

The evidence for that third phase is building quickly. In May 2026, Gartner predicted that by 2027, more than 65% of engineering teams using agentic coding will treat IDEs as optional, shifting control, governance, and validation to automated platforms. McKinsey's State of AI survey from August 2026 found that about two in ten organizations are already scaling software coding agents (31% of larger enterprises), and that 32% have decided against buying at least one software product or feature because they could build it internally with agentic coding tools.

That changes the question regulated firms need to answer. It's no longer which coding assistant to approve. It's how to run an engineering organization where AI agents do a growing share of the work.

Adoption Isn't Productivity

The pressure to move first on AI comes from everywhere: boards, investors, vendors, and engineers who feel they're falling behind. But adoption and productivity are two very different things. It was one of the strongest themes at the event, and the wider research says the same.

Atlassian's State of Developer Experience 2026 report found that 87% of developers say AI helps them ship faster, yet only 17% of developers and 23% of leaders say their organization has a clearly communicated, actively implemented AI strategy across the software development lifecycle. Developers in organizations that do have one are nearly twice as likely to report less friction. Training matters just as much: after Indeed rolled out structured AI coding training, engineers who completed it cut their coding time by around 35%, while those who didn't saw little change. We dig into the numbers in The AI Productivity Paradox.

Buying licenses isn't a strategy. The organizations getting value from AI are the ones investing in the environment around it.

The Real Challenge: Preparing Engineering for Agentic AI

The discussions at the event went well beyond writing code faster. Speakers explored AI agents coordinating tasks, routing work to different models, supporting continuous code review, and removing friction across the whole software development lifecycle.

For regulated organizations, that introduces a different challenge. How do you give AI agents the knowledge they need, while making sure they operate within the architecture, standards, and controls the organization has approved?

That requires more than an AI coding assistant. It requires a new engineering operating model, built on a governed foundation that provides consistent context, makes organizational knowledge accessible, and keeps visibility and accountability across the lifecycle.

That's the problem we're building DesignVerse to address. It gives AI a shared, governed foundation built around an organization's own code, architecture, documentation, design systems, and engineering standards. Product and engineering teams work from the same organizational context, and the organization keeps control of how AI operates within its own infrastructure.

The underlying principle matters whether or not you use DesignVerse: AI needs to work from knowledge your organization owns, maintains, and trusts. Without that foundation, more AI autonomy simply raises the cost of getting things wrong.

And it holds whichever tools win. One speaker summed it up in one word: Betamax. Plenty of technologies looked inevitable and never made it. The tools will keep changing; the foundations that let you use any of them safely won't. In a regulated business, those foundations take months, sometimes years, to build, so don't wait for the market to settle. Get ready while it does.

What Regulated Organizations Need to Get Right

The speakers at CTO Craft offered something more useful than predictions: a readiness agenda. None of it depends on picking the winning tool.

Governance and Platform Controls

One line from the platform session stuck with me: the fastest platform isn't the one with the fewest controls, it's the one where controls are automated and consistent. As agents take on more autonomy, those controls need to extend beyond code generation into review, validation, and deployment. That means knowing what agents can access, what they can change, which systems they can interact with, how their work is validated, and who remains accountable. Think immutable audit trails, visibility into which model did what, and paved roads that give every team the same secure path, inside your own perimeter where needed.

Shared Engineering Context

Context isn't prompt engineering, and it isn't just connecting AI to your documentation. It's a governed, maintained understanding of the organization that every AI agent works from:

  • Architecture and system dependencies
  • Engineering standards and approved libraries
  • Design systems and product requirements
  • Security policies and compliance constraints
  • Historical decisions and their rationale
  • Ownership and responsibility

Without it, AI guesses. And it can't be an engineering-only concern: product, architecture, security, and engineering teams need to agree on one shared, owned context rather than a dozen private ones. (

Training and Ownership

Start with an honest picture of who uses AI, for which work, under whose rules, and with how much training. Then make ownership explicit. Atlassian found that 75% of developers now have more responsibility for validating the quality and safety of AI-generated output. That responsibility needs the skills to match, and every change that reaches production needs a person accountable for it.

Measurement and ROI

Productivity has to be measured by business outcomes, not engineering activity. More code, more PRs, and more releases don't necessarily mean more value. DORA metrics still matter, but on their own, they're not enough.

One case shared at the event made this painfully clear: four months of work, 38 contributors, hundreds of PRs, and a successful launch, followed by 14% adoption, no improvement in customer satisfaction, and support interactions up 8%. Shipped isn't the same as delivered.

As agents scale, leaders will face harder questions:

  • Does AI reduce the total cost of software delivery?
  • Does it improve quality and reliability?
  • Can we safely increase the volume of software we produce?
  • Can we reduce our dependency on external vendors?
  • Can we keep control as agents take on more autonomy?

AI spend will be scrutinized like any other line in the budget. These are the questions it will be judged on.

The Risk of Losing Engineering Knowledge

This was the quietest theme at the event, and I think it's the most important. Most AI engineering content focuses on hallucinations, security vulnerabilities, and code quality. The more fundamental question is this: what happens when engineers no longer understand the systems they're responsible for maintaining?

One scenario from the Q&A stayed with me. Fast-forward five years. Most of the code has been written by AI, the engineers who knew why the system works the way it does have moved on, and AI has become the only practical way to change it. In most industries, that's a maintenance problem. In banking, it's an accountability problem. I can't imagine a regulator accepting "the model wrote it" as an answer.

Much of what makes a system work lives in people's heads: the unwritten rules, the reasons behind past decisions, the exceptions everyone knows about but nobody wrote down. Agents make this more urgent, not less. An agent working without that knowledge repeats the same gaps on every task, at speed. And finding information is already developers' biggest source of friction, according to Atlassian's research.

My take: as AI takes on more of software creation, the competitive advantage shifts from the ability to produce code to the ability to preserve and govern the knowledge behind it.

That means designing ownership in, not assuming it:

  • Every system, standard, and decision has a named owner who understands it.
  • The reasons behind decisions are captured, not just the decisions.
  • Junior engineers still build judgment by working on real problems, not just reviewing AI output.
  • Accountability for what ships stays with people.

What This Means for Financial Services

In financial services, AI readiness isn't simply an engineering efficiency exercise. It sits alongside existing obligations around operational resilience, ICT risk management, third-party oversight, and accountability.

The EU's Digital Operational Resilience Act (also abbreviated DORA, but not to be confused with the engineering metrics above) has applied since 17 January 2025. Its rules on ICT risk management and third-party risk make auditability, infrastructure control, and dependence on outside technology providers board-level concerns, and AI tools and agents that interact with your systems are part of that picture.

The EU AI Act adds further obligations, depending on how AI systems are used. Its AI literacy requirement has applied since February 2025, while the rules for high-risk systems, which cover uses such as credit scoring, were pushed back to December 2027 by the Digital Omnibus. Most engineering uses of AI won't be high-risk in themselves, but the direction of travel is clear.

The challenge is to build an engineering environment where AI operates within controls that are demonstrable, repeatable, and auditable.

Readiness Can't Wait

The line from the event I've repeated most since: leaders should be neither blindly enthusiastic nor unnecessarily resistant.

For regulated engineering teams, that's a call to start the right work now, not to slow down. The organizations that come out ahead won't necessarily be the ones that adopted AI first or generated the most code. They'll be the ones that built it into a well-designed operating model, and kept the human judgment and knowledge to run it.

FAQs

What is agentic AI in software engineering?

Agentic AI in software engineering refers to AI agents that can plan and carry out multi-step development tasks with a degree of autonomy, such as writing and testing code, reviewing changes, and working with development tools and infrastructure. It goes beyond coding assistants, which suggest code while a developer stays in control of each step.

How can banks govern AI coding agents?

Start by defining what agents can access, what they can change, and which systems they can interact with. Validate their work through automated checks and human review, keep an audit trail of what each agent did, and make sure a named person stays accountable for every change that reaches production.

What does AI readiness mean for financial services?

It means having the platform controls, shared engineering context, training, ownership, and measurement in place to use AI safely at scale. It also means doing so in a way that stands up to obligations such as DORA's ICT risk and third-party requirements and the EU AI Act.

How can organizations measure the ROI of AI coding tools?

Measure business outcomes rather than activity: the total cost of software delivery, quality and reliability, customer and adoption outcomes, and reduced dependency on external vendors. Code volume, PR counts, and AI usage on their own say little about value.

How do you protect engineering knowledge as AI writes more code?

Give every system, standard, and decision a named owner, capture the reasons behind decisions, and keep junior engineers working on real problems. Accountability for what ships has to stay with people, even when AI writes the code.