
The most requested AI feature in banking CRM right now is some version of "ask the assistant about my client". A relationship manager types "what changed for Harbour Logistics since our last review?" and gets a concise, accurate brief.
The tempting implementation is a vector store full of CRM notes and transaction summaries, plus a chat model. In a bank, that implementation is a breach waiting to happen. A retrieval index that ignores entitlements will happily surface a politically exposed person's file, a colleague's staff account or another region's clients to anyone who asks the right question.
This post describes a pattern for a permission-aware assistant: one that can never see more than the person asking.
The rule that drives everything
The model is an untrusted component. It receives only data the requesting user is entitled to see, for the specific task, after redaction. Authorisation happens before retrieval, never after generation.
request (user, client, intent)
-> entitlement check (RBAC + ABAC)
-> scoped retrieval (filtered by policy, not by prompt)
-> redaction / minimisation
-> grounded generation (template + citations)
-> output checks (PII leakage, policy rules)
-> response + audit record
Step 1: resolve entitlements first
Banks typically combine role-based access (relationship managers see their book) with attribute-based rules (no access to staff accounts, extra approval for PEP clients, region restrictions).
def entitlements(user, client_id):
policy = pdp.evaluate(subject=user, resource=client_id, action="read")
if not policy.allowed:
raise Forbidden(policy.reason)
return policy.scopes # e.g. ["balances", "cases", "notes"], not "kyc_docs"
The policy decision point is the same one the CRM UI uses. If the assistant and the screen disagree about what a user can see, one of them is wrong.
Step 2: retrieval filtered by policy
Every indexed chunk carries metadata: golden client ID, data classification, owning line of business, region. Retrieval applies those filters as hard constraints in the query, not as instructions in the prompt.
chunks = index.search(
query=embed(question),
filter={
"client_id": client_id,
"scope": {"$in": scopes},
"region": user.region,
"classification": {"$lte": user.clearance},
},
k=20,
)
For structured facts such as balances, exposures and maturities, skip the vector store altogether and call the system APIs with the user's delegated token. Numbers should come from the source, not from an embedding of last week's summary.
Step 3: minimise before the model sees it
-
Replace account numbers and national IDs with stable tokens.
-
Drop fields the intent does not need. A meeting brief does not need full transaction narratives.
-
Keep the token map server-side so the UI can re-hydrate values for the user if required.
Step 4: grounded, templated output
For a meeting brief, constrain the model to a fixed structure and require each statement to cite a source record.
SYSTEM: Produce a client brief with sections:
1. Relationship snapshot 2. Recent activity 3. Open cases
4. Upcoming maturities 5. Suggested talking points
Every factual line must end with [source:<record_id>].
If information is missing, say "not available". Do not infer.
A post-processor rejects any line without a valid citation to a record that was actually retrieved. That one check removes most hallucinated facts.

A second use case: complaint classification
The same guardrails apply to service triage. Regulators in several markets define a complaint broadly; an email that says "this is the third time I have asked" may legally count, even without the word "complaint".
def triage(message):
redacted = redact(message.body)
result = classifier.predict(redacted) # labels + confidence
if result.is_complaint and result.confidence < 0.9:
route_to_human_review(message, result)
elif result.is_complaint:
open_complaint_case(message, result, sla=regulatory_sla(message.region))
else:
route_service_queue(message, result.category)
audit.log("triage", message.id, result)
Tune for recall on the complaint label. A false positive costs an agent a minute; a missed complaint costs a regulatory breach.
Step 5: audit everything
Log the user, client, intent, retrieved record IDs, model version, prompt template version, output and any output-check failures. Keep prompts and responses inside approved regions and excluded from vendor training. Model risk teams will ask for this; it is easier to build in than to retrofit.
Measuring the assistant
-
Citation coverage: share of factual lines with a valid source.
-
Edit and override rates: how often bankers change or discard the brief.
-
Time saved per meeting: measured, not estimated.
-
Leakage tests: red-team prompts that try to reach other clients should fail every time.
Where this fits
A permission-aware assistant only works on top of resolved client identity and timely data from core systems. The full TechCirkle guide to banking and CRM covers those foundations, and our LLM integration work applies this pattern to regulated environments.
Frequently Asked Questions
Why can't a banking CRM assistant just use a shared vector store?
Because a shared index without entitlement filters can return data the user is not allowed to see. Access must be enforced in retrieval, before the model receives anything.
Where should authorisation happen in an LLM pipeline?
Before retrieval, using the same policy decision point as the CRM user interface, so the assistant can never exceed the user's own permissions.
Should balances come from embeddings?
No. Fetch structured figures directly from source system APIs with the user's delegated access, and use retrieval mainly for unstructured notes and documents.
How do you reduce hallucinations in client briefs?
Use a fixed template, require a citation to a retrieved record on every factual line, and reject output lines that lack a valid citation.
Can an LLM identify regulatory complaints?
It can classify likely complaints well, but low-confidence cases should go to human review and the system should be tuned to avoid missing genuine complaints.
What should be logged for model risk review?
User, client, intent, retrieved records, model and template versions, output and any failed checks, stored in approved regions.