Grounding

A language model can write fluent prose about your firm's capital requirements, even when the numbers it produces are entirely fabricated. Grounding is the discipline of anchoring AI outputs to verified sources: your data, your documents, your systems of record. Without it, generative AI in financial services is a liability, not an asset.

What is grounding?

Grounding is the process of constraining an AI model's outputs to information from authoritative sources rather than allowing it to generate from its training data alone. In a financial services context, this means connecting the model to your policy documents, transaction databases, regulatory libraries, and other verified sources so that its responses reflect your institutional reality.

The concept exists because large language models are fundamentally generative. They produce text that is statistically likely, not text that is verified. A model asked about your firm's IFRS 17 transition progress will generate a plausible answer whether or not it has access to your actual transition documents. Grounding ensures it only responds based on what it can retrieve from your sources.

Retrieval-augmented generation is the most common grounding technique, but grounding is broader. It includes tool use (the model calls an API to get a live balance rather than guessing), structured output constraints (the model must return data in a schema that matches your systems), and verification steps (a second model or rule-based system checks the output against source data). These techniques can be combined.

The landscape

Regulators have not yet used the term "grounding" in formal guidance, but the underlying principle is embedded in existing expectations. The PRA's SS1/23 requires that model outputs be "fit for purpose" and that firms understand the limitations of their models. An ungrounded language model that generates plausible but unverified regulatory analysis is not fit for purpose in any supervised context.

The EU AI Act's transparency requirements for high-risk AI systems implicitly require grounding. If an AI system must explain its reasoning to an affected individual (a borrower denied credit, a policyholder whose claim is queried), the explanation must trace back to specific data inputs and rules. An explanation generated from training data rather than the actual case file does not meet this standard.

Every major AI platform now offers grounding features. Azure OpenAI Service provides "On Your Data" connectors. Google's Vertex AI includes grounding with Google Search or custom data stores. AWS Bedrock offers knowledge base integration. The tooling exists. The challenge is configuring it correctly and validating that it works for your specific documents and queries.

How AI changes this

Grounded AI assistants are replacing the FAQ and the help desk for internal queries. A compliance analyst asking "what is our policy on gifts from clients above 500 pounds?" gets a direct answer citing the specific section of the gifts and entertainment policy, rather than a generic response. The answer includes the document reference, the section number, and the last-updated date. This is achievable today with well-configured RAG.

Grounding makes customer-facing AI viable for the first time. A chatbot that answers customer questions about their insurance policy, grounded in the specific policy wording and the customer's individual cover, is qualitatively different from one that generates generic insurance guidance. The FCA's Consumer Duty requires that communications are fair, clear, and not misleading. Grounding is how you meet that standard with AI.

Automated report generation with grounding produces audit-ready outputs. A quarterly risk report generated by a grounded system includes specific figures drawn from your risk systems, specific regulatory references drawn from your compliance library, and specific commentary drawn from your internal analysis. Every statement is traceable. The human reviewer verifies rather than drafts.

What to know before you start

Grounding reduces hallucination but does not eliminate it. A grounded model can still misinterpret a retrieved document, combine information from unrelated sources, or fill gaps in retrieved context with generated content. Test for these failure modes explicitly. Build evaluation datasets that include questions where the answer is not in the knowledge base, and verify the model says "I don't have that information" rather than guessing.

Source quality determines output quality. Grounding a language model on outdated, contradictory, or incomplete documents produces grounded answers that are still wrong. Before deploying a grounded AI system, audit the underlying knowledge base. Remove outdated documents. Resolve contradictions. Fill gaps. This preparatory work is often more valuable than the AI system itself.

Measure grounding fidelity, not just answer quality. A good answer that cites the wrong source is dangerous. A bad answer that correctly says "no relevant documents found" is safe. Build metrics that track whether citations are accurate, whether the answer is supported by the cited documents, and whether the system correctly abstains when information is unavailable.

Start with a high-stakes, well-curated knowledge base where accuracy is non-negotiable. Your regulatory library or your product terms and conditions are good candidates. These documents are maintained, versioned, and authoritative. Build the grounding pipeline, test it rigorously, and measure citation accuracy before expanding to less curated document collections.