---
title: The Data Foundations of AI Success
description: "Discover the data foundations of AI success: how classification, metadata and managed RAG help banks and insurers build trusted, production-ready AI."
image: https://sedona.digital/hubfs/Data%20Foundations.jpg
---

[![](https://sedona.digital/hubfs/logo-sedona-digital-white.svg)](https://sedona.digital/)

- [What we do](https://sedona.digital/what-we-do) 
    - Strategy
      
          - [Technology Strategy](https://sedona.digital/technology-strategy)
          - [Customer Strategy](https://sedona.digital/customer-strategy)
          - [Data Strategy](https://sedona.digital/data-strategy)
    - Innovate
      
          - [Software](https://sedona.digital/software-services)
          - [Data & AI](https://sedona.digital/data-ai)
          - [Security](https://sedona.digital/security-services)
          - [Microsoft Cloud](https://sedona.digital/microsoft-cloud)
    - - [Our Microsoft Partnership](https://sedona.digital/our-microsoft-partnership)
          - [Speak to a strategist](https://sedona.digital/contact-us)
- [How we do it](https://sedona.digital/how-we-do-it)
- [Industries](https://sedona.digital/industries)
- [Insights](https://sedona.digital/insights)
- [About us](https://sedona.digital/about-us)

[Contact us](https://sedona.digital/contact-us)

- Blog

# The Data Foundations of AI Success

#### Sedona Digital

 November 19, 2026

This article is written and published in conjunction with Sedona Digital

When we look across our financial services customers and prospects, a few things separate the firms trying out AI from the firms putting serious AI applications into production. One of the biggest is how much they've invested in the data services and capabilities that sit underneath.

That gap is getting more important. The Bank of England and FCA's latest survey found that 86% of UK financial services firms now use AI, up from 75% in 2024. Generative AI now makes up 56% of all reported AI systems. The same survey found that 80% of firms have AI agents deployed, but only 2% give those agents broad autonomy. Most agents still recommend an action while a human decides and carries it out. [Bank of England, AI in UK financial services 2026.](https://www.bankofengland.co.uk/report/2026/artificial-intelligence-in-uk-financial-services-2026)

So adoption is no longer the problem. The harder question is how firms move from AI that supports decisions to AI they trust enough to rely on, and that comes down to data.

As an insurer or bank, you'll already have data platforms, engineering capability and governance processes in place. You'll probably have a few silos too, whether in enterprise applications or unstructured data repositories. Much of this can already support your AI applications or Copilot initiatives.

The challenge is rarely a lack of data. It's whether the right data is discoverable, trusted, properly governed and available in the right context when an AI application needs it. That's where many AI programmes stall. The model can usually produce an answer. What the organisation can't always do is say which sources informed it, whether it was allowed to use them, or whether the output is reliable enough for the decision being made.

## ![Two professionals reviewing a tablet in a modern office, illustrating collaboration on the data foundations of AI success.](https://sedona.digital/hs-fs/hubfs/AI%20in%20financial%20services.jpg?width=780&height=521&name=AI%20in%20financial%20services.jpg)

## Classification and Metadata

Sensitivity, confidentiality, integrity, quality and ownership are familiar concepts, and you've likely got them covered. But what about classifications based on usage, business process or specific AI use cases?

The same data asset can need very different classification levels depending on the process or agent using it. A document that's fine for internal research may not be appropriate for a customer-facing claims decision, an automated underwriting recommendation or an agent that can trigger an action. Classification needs to reflect context, not just sit on a repository as a fixed label.

This matters because classification can act as a proxy for regulatory impact. The same information might support a low-risk internal search in one workflow and a higher-impact customer decision in another. The AI system needs to understand that difference, and the controls around it need to reflect it.

Trust and confidence weighting matters too. A source that's current, authoritative and consistently used in a process should carry more weight than an old, duplicated document with unclear ownership. Equally, a highly relevant source may still need to be excluded if its usage rights, freshness or quality don't meet the threshold for the task.

Usage rights and quality thresholds can vary a lot, especially with unstructured and human-generated data. A policy document, claims note, email, call transcript or underwriting file can hold valuable context. It can also contain personal data, commercially sensitive material, inconsistent terminology or information that's since been superseded.

Regulators have been clear that data sits at the heart of AI risk. In the Bank of England and FCA's 2024 survey, four of the top five perceived AI risks were data-related: data privacy and protection, data quality, data security, and data bias and representativeness. [Bank of England, AI in UK financial services 2024](https://www.bankofengland.co.uk/report/2024/artificial-intelligence-in-uk-financial-services-2024)

Classification metadata gives AI agents the context they need to work with the accuracy, confidence and accountability expected of a human decision-maker. It helps an agent decide what to retrieve, whether it's appropriate to use, how much confidence to place in it and when to bring in a human.

Firms that skip this step risk deploying AI with low confidence, limited functionality and poor accuracy. They may also find themselves stuck in pre-production because risk and compliance can't sign it off, so the business value never arrives.

## Good Grounding and Intelligently Managed RAG

Do your applications enable LLMs to be properly grounded in subject-matter expert knowledge? Is that capability available in the user interface, and is it maintained with the right quality standards and guardrails?

Done well, a managed RAG capability can drive adoption, improve application performance and reduce running costs. It means an AI application answers from approved material rather than relying only on what's built into the model.

But grounding isn't just a technical integration. It's a knowledge-management discipline. RAG is only as good as the quality, currency, relevance and permissions of the content in the retrieval layer.

Subject-matter experts need an easy way to curate knowledge, store it properly and rerun processes with it. If an expert finds a better source, corrects a misreading or adds a new exception, updating the knowledge base and checking the impact should be straightforward.

These repositories also need ongoing care. Too little grounding data reduces accuracy. Too much, or outdated content, slows applications down, increases costs and introduces errors. A system that retrieves every possible document isn't more useful than one that retrieves the right document, with the right permissions, at the right point in the workflow.

The aim is a controlled knowledge supply chain. Firms should know where content came from, who approved it, when it was last reviewed, which version applies, whether it's permitted for the use case and how it influenced the response. That's what turns RAG from a search feature into a governed production capability.

## Building The Right Data Foundation

A strong AI data foundation doesn't need a full replatforming programme. For most financial services firms, the starting point is making better use of the data, metadata and governance already in place.

![Data foundations of AI success illustrated by a connected globe and icons for analytics, cloud storage, databases, security, and workflows.](https://sedona.digital/hs-fs/hubfs/Financial%20advisor.jpg?width=780&height=520&name=Financial%20advisor.jpg)

- **Define the use case and its decision impact.** A claims summarisation tool, an underwriting research assistant, a customer service agent and an internal policy assistant may all use similar content, but each needs different confidence thresholds, permissions and human oversight.
- **Build a practical metadata model.** Go beyond sensitivity and ownership to capture business domain, process relevance, source authority, version, review date, usage rights, quality status and permitted AI use cases.
- **Make the metadata operational.** It should shape retrieval, prompt construction, responses, confidence scoring, escalation and audit logging. Metadata that sits in a catalogue but doesn't affect how the system behaves won't give the business the control it needs.
- **Create feedback loops.** SMEs, reviewers, compliance teams and end users should be able to flag poor sources, outdated material, missing context or unsafe outputs, and that feedback should flow back into classification, curation and evaluation.

This turns data governance from a box ticked before a project starts into a live operational capability. It also builds the evidence trail firms need to show their AI is grounded in the right material and producing decisions that can be explained and challenged.

## Why Metadata Matters for Agentic AI

Agents can retrieve information, interpret a rule, choose a tool, make an interim decision and pass the result to another system or a human approver. Without clear metadata, an agent might have access to the right information but not the context to use it properly. With it, the agent can tell an approved underwriting guideline from an informal note, a current policy from a superseded one, and data that can support a recommendation from data that needs human validation.

Metadata also supports traceability. If an agent produces an unexpected outcome, the firm should be able to reconstruct the request, the sources retrieved, the rules applied, the model used and any human approval or override.

This is where the industry is heading. The 2026 survey shows that most firms keep agents in a decision-support role for now. Even counting planned deployments over the next 12 months, no more than 9% of firms expect agents to approve or decline customer outcomes, initiate payments or change limits and controls. [Bank of England, AI in UK financial services 2026](https://www.bankofengland.co.uk/report/2026/artificial-intelligence-in-uk-financial-services-2026) Bank of England Deputy Governor Sarah Breeden has also signalled that existing frameworks weren't built for autonomous agents. She said relying on a human in the loop for every agent action is unlikely to be realistic. [Reuters](https://www.reuters.com/business/finance/bank-england-sees-growing-risks-financial-stability-ai-2026-07-07/)

When the human in the loop can no longer check everything, metadata and traceability become the control.

## From Data Foundations to AI Value

The firms making the fastest progress aren't necessarily the ones with the most advanced models. They're the ones that connect trusted data, clear governance and reliable engineering into a repeatable production pattern.

Third-party AI is making this more pressing. Third-party implementations now account for 53% of high-materiality AI systems, up from 35% in 2024. [Bank of England, AI in UK financial services 2026](https://www.bankofengland.co.uk/report/2026/artificial-intelligence-in-uk-financial-services-2026) When more of the model sits outside your walls, the data and context you control become your main lever for accuracy and accountability.

Classification and metadata provide the context, and managed RAG provides the grounding. Together, they help an insurer or bank move beyond a promising proof of concept and deploy AI with the confidence a regulated environment demands.

Sedona builds consulting-led classification, metadata and managed RAG services, based on years of putting AI applications into enterprise production. We fit these foundations around your existing platforms, teams and governance model, so AI becomes safe, useful and scalable where your business already works.

If you're looking at how to get your AI initiatives into production, get in touch. We'd be happy to help.

## FAQs

##### What data foundations do AI applications need?

AI applications need data that's discoverable, relevant, current, permitted and trustworthy. That means more than a data platform. Firms need metadata that covers sensitivity, ownership, lineage, version, quality, usage rights and permitted AI use cases, plus a grounding layer that connects approved knowledge to the AI workflow. Without that, generative AI can produce plausible answers that can't be traced back to an authoritative source.

##### How should insurers manage RAG knowledge bases?

Treat them as live operational assets, not static document stores. Content should be curated by subject-matter experts, versioned, reviewed for currency, classified by sensitivity and use case, and removed or improved when it goes out of date. Firms should also test retrieval against realistic scenarios, checking source relevance, accuracy, completeness and whether the system escalates when the evidence is weak.

##### Why is Metadata important for AI governance?

Metadata gives the context needed to control and explain AI behaviour. It determines which data can be used, by which application or agent, for which process and to what confidence and quality thresholds. Captured alongside prompts, sources, model versions, outputs and human decisions, it creates an audit trail that lets firms investigate issues, respond to complaints and show that AI-supported decisions were properly governed.

<https://www.linkedin.com/shareArticle?mini=true&url=https://sedona.digital/insights/data-foundations-ai-success> <https://api.whatsapp.com/send?text=https://sedona.digital/insights/data-foundations-ai-success> [mailto:?subject=Check%20this%20out!&body=https://sedona.digital/insights/data-foundations-ai-success](mailto:?subject=Check%20this%20out!&body=https://sedona.digital/insights/data-foundations-ai-success)

![](https://sedona.digital/hubfs/overlay-bg-2.png)

Get in touch to find out more

[Connect](https://sedona.digital/contact-us)

![logo-sedona.png](https://sedona.digital/hubfs/logo-sedona.png.svg)

Award winning transformation partner specialising in Software/AI, Data/AI, Cloud Security and Business Change.

**United Kingdom** | 10 Lower Thames Street, London, EC3R 6EN

**USA** | 251 Little Falls Drive, Wilmington, New Castle, 19808.

[Sedona Home](https://sedona.digital/) [What we do](https://sedona.digital/what-we-do) [Industries](https://sedona.digital/industries) [Insights](https://sedona.digital/insights) [About us](https://sedona.digital/about-us) [Careers](https://sedona.digital/career) [Contact](https://sedona.digital/contact-us)

Copyright © 2026 Sedona Digital

| [Privacy Policy](https://sedona.digital/privacy-policy) | [Terms of Use](https://sedona.digital/terms-of-use) | [Sustainability](https://sedona.digital/sustainability) | [Modern Slavery Statement](https://sedona.digital/hubfs/POL%20Modern%20Slavery%20Act%20Statement.pdf)

```json
{
  "@context" : "https://schema.org",
  "@type" : "BlogPosting",
  "author" : {
    "@type" : "Person",
    "name" : "Sedona Digital",
    "url" : "https://sedona.digital/insights/author/sedona-digital"
  },
  "dateModified" : "2026-11-19T09:13:38.685Z",
  "datePublished" : "2026-11-19T08:57:18.000Z",
  "headline" : "The Data Foundations of AI Success",
  "image" : [ "https://sedona.digital/hubfs/Data%20Foundations.jpg" ],
  "mainEntityOfPage" : {
    "@id" : "https://sedona.digital/insights/data-foundations-ai-success",
    "@type" : "WebPage"
  },
  "publisher" : {
    "@type" : "Organization",
    "logo" : {
      "@type" : "ImageObject",
      "url" : "https://sedona.digital/hubfs/logotype%20(1)-1.png"
    },
    "name" : "Sedona Digital Ltd"
  }
}
```