Future Labs

001

Live

Polaris View

Better, faster, human-led decisions.

The first product from Future Labs. It checks business loan proposals against lending policy, flags each risk with its source, and leaves every decision with the underwriter.

The question

Can AI do the groundwork on a loan decision, and show its evidence, without taking the decision from the underwriter or exposing customer data?

Built with

Azure OpenAI (GPT-4.1), ASP.NET Core, React, Azure SQL, OCR and extraction models. Hosted in Azure UK South.

Status

Live

The problem

Underwriters at UK lenders spend much of their time assembling information rather than judging it. A single business loan proposal can mean cross-checking credit reports, company records, guarantor details and affordability against a policy manual before any decision is made. Slow decisions cost deals.

But regulated lenders can’t hand those decisions to a black box. In the Bank of England and FCA’s 2024 survey, 46% of firms surveyed said they only partly understood the AI they use, and data protection was the biggest regulatory barrier to adoption. In January 2026, Parliament’s Treasury Committee flagged opaque AI decisions in credit as a risk to consumers.

Lenders need AI that does the groundwork, shows its evidence and keeps personal data under their control, without replacing the systems their teams already rely on.

Sources: Bank of England and FCA, Artificial intelligence in UK financial services 2024; House of Commons Treasury Committee, report on AI in financial services, 20 January 2026.

What we built

Polaris View is an assistant that sits alongside a lender’s existing systems. Before an underwriter opens a proposal, it has read the application, credit reports, company details and guarantor information, and checked them against the lending policy.

The underwriter sees a plain-English risk summary and each issue rated by severity, with the policy rule behind it and a link to the exact data that triggered it. They can ask questions about the proposal, view the reasoning, accept or decline, and rate each output.

How we built it

We treated Polaris as if a regulated lender were the client. That meant building to the governance, security and data standards we’d expect to be audited on, but in a synthetic test environment where we could try ideas, fail quickly and change direction without risk to anyone’s customers.

A data preparation pipeline uses OCR and extraction models to turn scanned documents and legacy system reports into structured data. Polaris runs in the lender’s own Azure environment in UK South, built on Azure OpenAI GPT-4.1, ASP.NET Core, React and Azure SQL.

Why we built it this way

The most important decision came first: Polaris would never make a decision itself. It prepares the case and the underwriter decides. Results always need human review, so the realistic gain is time on specific tasks, such as reaching a clear decline faster, not removing people from the process.

Next was trust. We restricted the model to the lender’s own approved policies and data, rather than its general knowledge, and required a citation for every statement.

What we found

Polaris has proved that AI can take on the heavy lifting of a credit decision without taking the decision away from the people accountable for it. That matters most where speed to the right decision wins business: underwriters start every case with the evidence already gathered, the risks ranked and the policy breaches flagged, so their time goes on judgement, not assembly.

Polaris flags policy breaches and risk factors on each proposal, from a credit score below the policy minimum to a guarantor under the required age, and links each one to the rule and data behind it.

For ClearSky, Polaris set the pattern for how Future Labs works: build to client-grade governance standards, test openly and keep what holds up.

Safety and governance

Polaris sends the model only derived, de-identified fields, such as postcode district instead of full address, and masks or drops the rest. No direct personal identifiers reach the model.

We followed Microsoft’s six principles of responsible AI, and our QA team tested outputs for bias, hallucination and correctness.

Every interaction is logged with the user, time and reasoning, so a lender can show how each decision was reached and who made it.

From lab to live

Polaris is now being deployed in live client environments, and it’s ready to be adapted for other lenders. An implementation starts with discovery to map the lender’s policies, data and systems, followed by a build tailored to their rules.

The same approach applies wherever experts review sensitive personal data to make decisions, including affordability checks and eligibility decisions in healthcare and energy.

Future Labs

Polaris View was the first product to come out of Future Labs, which ClearSky set up in summer 2025 to give its engineers time to test ideas too new to try on live client systems.

Our engineers work deep in regulated client projects, so they see early where new AI can help. Future Labs gave them the time and freedom to act on that, and Polaris View came first. It started in business lending, but it works wherever FCA-regulated firms need fast, human-led decisions.
Darren Auld
CEO

Common questions

How do you keep a human in control when using AI for credit decisions?

Polaris prepares each case and the underwriter makes the decision. It can’t operate without human review. Every flag shows the policy rule and the data behind it, and every interaction is logged with the user, time and reasoning, so a lender can show how each decision was reached and who made it.

How long does it take to implement an AI underwriting assistant?

We plan implementations in two phases. Discovery and data work takes 4 to 12 weeks: mapping your decision rules, auditing data sources and building the data pipeline. The build and integration phase takes 10 to 12 weeks. The assistant runs alongside your existing systems, so your team keeps working as normal.

Does our customer data leave our environment or train the AI model?

No. Your database, documents and business data stay in your own Azure tenancy, hosted in UK South. The AI model receives only derived, de-identified fields in an encrypted request, and that data isn’t used to train or improve the model.

Ready to turn frontier tech into reality?

Speak directly with our team in Edinburgh or Glasgow about your platform, data or AI initiative.

Georgina Hart, Head of Brand & Strategic Growth at ClearSky

Georgina Hart

Head of Brand & Strategic Growth

georgina.hart@weareclearsky.comEmail Georgina

Explore other Future Labs products

See what else we’re building and testing in Future Labs, our self-funded R&D.

Explore Future Labs
Sentinel, a ClearSky Future Labs prototype, shown as a floating 3D object

#003

In development

Sentinel

Human-in-the-loop telemetry.

Explore Sentinel

Aura, ClearSky's personalisation prototype, shown as a floating 3D object

#002

In research

Aura

Hyper-personalised customer journeys.

Explore Aura

Polaris View, ClearSky's AI decision prototype, shown as a floating 3D object

#001

Live

Polaris View

Better, faster, human-led decisions.

Explore Polaris