Back to blog Łukasz Mokrzycki
Governance Risk and Compliance Engineer
Point of ViewNov 17, 2026
On 2 December 2027, the EU AI Act's full obligations for high-risk AI systems come into force. For banks and insurers, that includes AI used to make decisions about individual customers: who is approved for a loan, and how life and health insurance are priced.
However, not every AI system a bank or insurer runs will fall into that high-risk category, and the parameters for what does could also change before December.
With that in mind, preparing for 2027 isn’t a separate compliance exercise. It fits into the wider goal of building AI your team can understand, explain, and stand behind. The best practice for now is to apply the Act's core principles across every AI use case - from onboarding new customers to underwriting loans and reviewing claims.
In this article, I won't lay out every detail of the new obligations. Instead, I'll cover four practical ways your organization can build transparent AI systems that don't keep you up at night.
Most of these principles come back to one core idea: treat every AI system not as a compliance project you close out, but as a product you own and maintain.
What will the EU AI Act enforce in 2027?
Before we dig into how to start preparing, here’s a quick overview of what’s on the horizon for December 2027.
The AI Act sorts AI systems by risk. Obligations for the high-risk tier, set out in Article 6(2) and Annex III, come into effect next December.
Those requirements include a risk management system, data governance, technical documentation, automatic logging, human oversight, accuracy and robustness requirements, and a quality management system.
As of now, only two "high-risk" use cases directly impact banking and insurance:
- Point 5(b) covers "AI systems intended to be used to evaluate the creditworthiness of natural persons or establish their credit score."
- Point 5(c) covers "AI systems intended to be used for risk assessment and pricing in relation to natural persons in the case of life and health insurance."
But while these are the only sector-specific use cases listed today, the Commission's final guidance on how to classify high-risk systems isn't due until August 2027. That's why we'd recommend applying the following best practices throughout every AI system your team runs.
1. Only add AI where it creates value, and guardrail it with rules
Banks and insurers can limit regulatory exposure by only using AI where it adds real value.
Credit, claims and underwriting processes are a chain of steps, and not every step benefits from adding AI. A business rule that says "no account for applicants under 18" is transparent, cheap to run, and never drifts.
AI earns its place by automating the work that rules can’t handle: extracting figures from a scanned bank statement, or summarizing a complex claims file. In Taktile Labs' FinSpread-Bench, AI agents spread financial statements with 96.5% accuracy, against 89% for trained analysts. Deciding the borrower’s final credit limit stays a deterministic, rule-driven step.
Combining rules and AI can also help your team meet regulatory requirements by putting clear guardrails around what an agent is allowed to do.
After an agent runs, rules can check the output against policy, so the agent never has the final say. They can also set the conditions for human review, such as any claim above €2,000, or any application with a data inconsistency.
Because those rules live in the decision process rather than in the agent's instructions, the agent cannot work around them.
2. Run AI systems like evolving products, not one-off projects
Because of the pressure to deploy AI quickly, teams often focus on the initial build, without a plan for long-term maintenance. IDC predicts that Global 1000 companies will underestimate their AI infrastructure costs by 30% through 2027.
That gap is critical, because the EU AI Act demands continuous oversight, validation, and risk management of AI systems
Just like other core infrastructure, such as a CRM or data pipelines, a healthy AI system also needs the support of dedicated people, money, and time. As a fellow engineer at Taktile put it: we need to run AI-driven systems like products, not projects.
A project has a clear beginning and end. A product is never finished. It's supported by a dedicated budget, cross-functional team, and system for constantly adapting and improving.
Here's what this looks like in action:
We've seen a case where a new model version looked better on general benchmarks. But when tested within a deployed agent, credit decision accuracy dropped by roughly 30%. The model wasn't worse. The agent's prompts and architecture had been tuned to the previous version, and the new one behaved differently enough to change outcomes.
The team caught it because the business users who understood the KPIs had long-term ownership over and visibility into the system.
3. Bake human oversight in from day one, and treat it as a UX challenge
The EU AI Act requires that a human can genuinely understand, monitor, and override an AI system's decisions in real time. The AI system itself has to give the human enough clear information and control to catch and correct problems.
Once you're running your AI system like a product, you can look at these human oversight requirements as a UX challenge. It’s about designing an environment that's not only optimized for AI, but easy for humans to interact with and understand.
Is it easy for a claims officer to query in natural language why an AI agent made a certain decision? Can a credit leader see in a single audit trail every data source, rule, AI, and human-driven action that went into a loan approval?
A governance policy can document that this is needed, but it comes to life in the system design.
In fact, the Act is explicit that the system itself has to be designed for oversight, and it even names the trap to design against: automation bias, the habit of trusting the output because it came from the machine.
The best defense against that trap is to bring domain experts in early. The people who understand what a reliable decision looks like should be in the room when the system is designed, not just when it's audited.
4. For easier auditing, build AI into end-to-end workflows
KPMG found that 51% of banks were already piloting AI agents by mid-2026. Now, the AI conversation in banking and insurance is moving from individual productivity to rewiring entire end-to-end processes around AI agents.
As we move into this next phase, there’s a risk that institutions could inherit a hangover from the first one: AI scattered across separate tools. And when AI is disconnected, it’s harder to audit, document, and align with regulations.
For a regulated end-to-end decision like a loan approval or a claim, the goal is for every AI step to run in the same environment. When the whole claims process runs through one decision flow, every AI step is visible in the same place: which model ran, what data it saw, what it produced, and where a rule or a person took over. The audit trail is much easier to assemble, and simpler for both your team and regulators to understand.
Getting ready by December 2027
A lot could still change before the high-risk obligations apply, including which use cases end up in scope. But the steps above will put your team in a good position, however the final rules land.
A practical place to start is mapping where AI already sits in your credit, claims, and underwriting processes. Which steps actually need it, and where would a rule do the job better? Who owns each system once it's live? Could a claims officer or credit leader explain a specific decision to a supervisor using what's in front of them today?
This is the work that enables your team to use AI in the decisions that define performance, and stand behind every one of them.
Frequently asked questions #
Does the EU AI Act classify credit scoring and insurance pricing as high-risk?
Yes. Annex III of the EU AI Act lists two financial-services use cases as high-risk: AI systems used to evaluate the creditworthiness of natural persons or establish their credit score (point 5(b), with an exception for fraud detection), and AI systems used for risk assessment and pricing of natural persons in life and health insurance (point 5(c)). Business lending, non-life insurance, and claims validation are not listed in Annex III.
When do the EU AI Act's high-risk obligations apply to banks and insurers?
From 2 December 2027. The Digital Omnibus on AI (Regulation (EU) 2026/1744), published in the Official Journal on 24 July 2026, moved the Annex III deadline from 2 August 2026 to 2 December 2027. The requirements themselves did not change. Transparency obligations under Article 50 and AI-literacy obligations under Article 4 already apply to banks and insurers today.
What does the EU AI Act require for human oversight of AI in credit and claims decisions?
Article 14 requires that a high-risk AI system is designed so the people overseeing it can understand its capacities and limitations, monitor it for anomalies, interpret its output, stay aware of automation bias, decide not to use it or override its output in a given case, and stop it. Oversight measures must match the system's risk and level of autonomy. These are design requirements on the system, not only a governance policy.
Do banks and insurers need a fundamental rights impact assessment under the EU AI Act?
Yes, for high-risk systems. Article 27 specifically requires deployers of Annex III point 5(b) and 5(c) systems, meaning credit scoring and life and health insurance pricing, to complete a fundamental rights impact assessment before first use. It must describe the process the system is used in, the people affected, the risks of harm, the human oversight measures, and what happens when a risk materialises. The results are notified to the market surveillance authority.
Does the EU AI Act apply to AI systems that are not high-risk?
Partly. Article 4 (AI literacy) and Article 50 (transparency, such as telling customers when they interact with an AI system) apply to all AI systems in scope. The heavier high-risk obligations do not, but financial supervisors expect similar discipline anyway: EIOPA's August 2025 Opinion applies governance, documentation, and oversight principles to insurers' AI systems that sit outside the high-risk tier, under existing Solvency II and IDD rules.
Disclaimer
This article is for educational purposes only and does not constitute legal advice. The EU AI Act and its implementing guidance are still evolving, and how they apply will depend on your institution’s specific systems and circumstances.