AI in the financial sector

In April 2026, GPT-5.5 was the model to beat, and anyone who wanted to build seriously with it paid a hefty price in API costs. In August, an open model appeared that runs on a few powerful graphics cards in your own server room and comes surprisingly close in benchmarks. Five months.

That pace determines how you, as a financial organisation, should think about AI. Models are getting smarter, cheaper and faster, and at the same time riskier. Below you can read what these four developments mean for insurers, accountants and other financial service providers, and what you can put in place now.

Smarter: today's best model is middle of the pack in six months

If you want to see how quickly language models are improving, take a look at AI IQ. Since May 2026, this initiative has been converting the scores of frontier models on twelve public benchmarks into an IQ-like scale. It is not a real intelligence test. A single number hides, for example, that a model can excel at maths while performing modestly at legal reasoning. As a trend line, however, it works: the top shifts every few months, and the lead keeps changing hands between OpenAI, Anthropic and Google.

More striking is how small the lead is. Epoch AI calculated that since 2023, Chinese models have trailed the American frontier by an average of seven months, ranging from four to fourteen months. Recent releases such as Kimi K3 and Qwen 3.8 suggest the gap is narrowing further. Many of these models are also open-weight: you are allowed to download them and run them yourself.

For your IT strategy, this has a sober consequence. A multi-year contract with a single AI vendor is a bet on who will be in front two years from now. Our unique comparative AI model benchmark showed that one model wins at bug fixing and another at building a complete application. So set up your AI environment in a way that lets you switch models per task. Connecting through open standards such as MCP keeps your knowledge sources and integrations in place when the model changes.

Cheaper: the bill shifts from tokens to people

According to Epoch AI, the price of a fixed level of performance falls by a factor of 9 to 900 per year, depending on the task. That remains abstract until you see it reflected in your own budget.

Take Qwen 3.8 27B, released as an open-weight model in August 2026. In the benchmark mentioned above, it scored 24 out of 25 points on full-stack development, close to the most expensive commercial models. The difference: you can run this model on your own hardware. No data goes to an external party, and a small team can work with it at the same time.

It is not free. Hardware, power and management cost money, and a local model is slower than the large cloud models. On the other hand, you trade a variable per-token bill for an investment you can plan. And your data stays in-house, which for many compliance departments carries more weight than speed.

It can get even smaller. In September, PrismML released Ternary Bonsai 2 27B, a compressed version of the same Qwen model. It is nine times smaller, 5.9 GB, and runs on a laptop. The makers say it retains 98.2 percent of the benchmark score. That figure comes from the makers themselves, and the model gives up noticeably more on knowledge questions and reasoning than on arithmetic and code. The direction is clear, though: you do not need a frontier model to summarise a policy condition or classify an incoming letter.

So where does the money go? To setup, integration with your systems, management and supporting employees. In the boardroom, the question shifts from "what does a licence cost?" to "who will set this up securely and make sure people use it well?"

Faster: custom chips and models that only make decisions

GPUs were designed for graphics. They handle AI better than ordinary processors, but they are not optimal. That is where heavy investment is now going. In January 2026, OpenAI signed a multi-year deal with Cerebras for 750 megawatts of compute, built around one enormous chip that brings compute, memory and bandwidth together. In the Netherlands, Eindhoven-based EUCLYD raised more than 200 million euros in September, with Samsung as co-lead, for energy-efficient AI chips. Axelera AI, also from Eindhoven, likewise raised more than 200 million euros in February. If you take digital sovereignty seriously, that is good news: part of tomorrow's AI infrastructure may well be European.

Speed sounds like a technical detail, until an underwriting check is done in two seconds instead of two minutes. Then you can design a process differently.

For finance, a different kind of model may be even more interesting. Jev, from the American company TypeSafe AI, does not generate text. You give it a case file and a question with fixed answer options, and it returns a choice with a probability and a confidence score. Because the answer always falls within the given options, it cannot make anything up. TypeSafe claims this is 40 to 200 times faster and 40 to 400 times cheaper than a frontier language model. Those are the maker's own measurements, but the idea is catching on: open variants are already appearing.

Think of sanctions and PEP screening, risk assessment during underwriting, assessing the level of a set of financial statements, categorising transactions or prioritising alerts. All of these are bounded decisions made in large volumes. A calibrated probability is also easier to log, test and explain to a regulator than a freely worded paragraph.

More dangerous: 2026 was the year of the escaped agent

What was still a research-paper scenario in 2025 made the news this year. Three examples from recent months:

The same capability a model uses to find a vulnerability can help your own security team close it sooner. Still, as an organisation holding client files and financial data, there is one lesson to draw: give AI exactly the access a task requires and not a millimetre more. An agent that runs continuously and can reach everything is a door that stays open 24 hours a day.

Regulation is keeping pace, albeit more slowly than planned. The obligations for high-risk AI under the EU AI Act, including creditworthiness assessment and risk assessment in life and health insurance, were due to take effect on 2 August 2026. In May, the Council and Parliament reached a provisional agreement to postpone that date to 2 December 2027. If you use that time to deploy AI without logging, human oversight and documentation, you are only putting the work off.

What this means for financial and business services

Accountants, insurers and advisers earn most of their money by reading, assessing, summarising and recording. That is exactly where language and decision models are strongest. The pressure on these sectors is therefore greater than on most others.

Junior case work changes first. Processes that now run one after another, from application to approval, can partly run in parallel, with people handling the exceptions. But a chatbot alongside an unchanged process only saves a few minutes here and there. The real gain lies in the redesign.

For a 30-person firm, that could mean suddenly being able to handle the volume of a 50-person firm. An insurer that waits will see a newcomer bring products to market faster and more cheaply. Which applications deliver the most in insurance and financial services, and how Copilot and ChatGPT compare with an environment of your own, is covered in Applying AI in financial services.

What you can put in place now

Part of what is needed lies outside your control: global oversight of the most powerful models, and enough power, preferably green, to run them. But there are four things you can tackle yourself.

1. Make sure you can see what is happening

   You cannot steer what you cannot see. In almost every organisation, employees occasionally paste client data into a public chatbot using a personal account. In the financial sector, that is a direct GDPR and reputational risk. A central, approved environment with insight into usage and costs per department is the first step.

2. Use the right model for each task

   The most expensive model is rarely the best for every task. A summary can be handled perfectly well by a small or local model, while a complex analysis justifies a heavier one. An environment in which you can switch without redoing work protects you against price increases and lock-in.

3. Let guardrails grow with capability

   A chat that searches internal procedures requires fewer safeguards than an agent that writes to client files on its own. Define which sources AI may use, log what happens and keep a human accountable for decisions that affect clients. Look at jurisdiction too: data held by an American cloud provider falls under the CLOUD Act, even if the servers are in Europe.

4. Bring the organisation along

   The technology is the easy part. Change provokes resistance and takes effort. What works: AI ambassadors in each team, a shared prompt library, training on your own use cases and room to experiment, for example with an internal hackathon. Make it structural, and make it fun. People who experience AI as a helpful tool use it better.

In practice: an AI chat for accounting firms

You can see how this comes together in the solution we developed for MostEasy. It runs an AI chat for accountancy with Dutch hosting, with no dependence on the CLOUD Act and with client data that is never used to train models.

Employees choose the model that fits each task, the chat draws on pre-approved sources and it includes features built specifically for accountants. What it deliberately does not include: swarms of agents that run continuously and can reach everything. Control first, then autonomy.

Curious what this would look like for your organisation? Get in touch for an in-depth demo of the Compliant AI Chat. We will show you how it works with your own documents and use cases. 

In closing

AI is becoming smarter, cheaper, faster and more dangerous, all at the same time. For financial service providers that is uncomfortable, because their core work lends itself well to automation. It is also an opportunity. Organisations that get a grip on usage, costs, model choice and risk now will build a lead that will be hard to close two years from now. Those who wait for the dust to settle will find that it never does.