Compliance teams could not approve the tool because no one could reconstruct, months later, why a given answer had been produced from which documents at which version.
STUDY 03 / 30 · ANONYMISED · NDA · AI TOOLS · INDIA · SERIES B
AI TOOLS · INDIA Retrieval you can audit, for a regulator who will ask.
RoleLead designer + systems
Timeline12 weeks
Team2 designers, 1 PM, 2 ML engineers
VerticalRAG platform for regulated industries · BFSI + pharma
Made the retrieval chain a first-class object with a permanent, exportable trace - chunk, document version, embedding run, and the ranking that put it in context.
Compliance approval unblocked at three institutions. The audit export became a sales asset rather than a support burden.
3 weeksTIME TO COMPLIANCE APPROVALfrom indefinite
100%ANSWERS WITH FULL TRACEby construction
4 minAUDIT RECONSTRUCTION TIMEfrom ~2 days
THE ARGUMENT
Why the obvious solution was wrong.
The study matters because the product problem was reframed before the interface was polished.
Regulated buyers do not evaluate an AI product on answer quality. They evaluate on whether they can defend it to a supervisor eighteen months later, when the model has been retrained twice and the source document has been superseded. Every RAG interface we reviewed treated retrieval as plumbing - invisible, ephemeral, and unversioned. That is a product-blocking decision made by default rather than on purpose.
We built the trace as a permanent artefact attached to every answer: which chunks were retrieved, from which document at which version hash, under which embedding run, with the ranking scores that selected them and the ones that just missed. It renders as a readable chain a compliance officer can follow without an ML vocabulary, and exports as a signed PDF. The near-miss chunks matter more than expected - reviewers wanted to see what the system nearly used, because that is where their judgment about coverage lives.
THE INTERFACE CRAFT
The interaction, rendered as a working product surface.
The specimen below is code-native and uses the study's own design logic. The client interface remains protected.
AI TOOLS
CONFIDENCE 82%
Action preview ready
DESIGN DECISIONS
Positions we would defend.
Each decision names the principle and the product consequence, not a stylistic preference.
Retrieval is not plumbing
In a regulated context the retrieval chain is the product's defensibility. Hiding it is a commercial decision, not a design simplification.
Version the corpus, not just the model
A correct answer from a superseded document is a finding. Document version belongs in the trace at equal prominence.
Design for the person who was not there
The trace's reader is a supervisor eighteen months later with no context. That reader, not the daily user, set the information architecture.
PRODUCT LEADER READOUT
What transfers, and what should remain specific to this product.
A case study is useful when its operating principle travels without turning the original interface into a template.
Read the operating condition
For RAG platform for regulated industries · BFSI + pharma, the transferable lesson is not a copied screen. It is the condition the interface had to make legible: Answer, then the chunks that produced it, then the documents and versions those came from, then the run that indexed them. One scroll, no ML terminology. Rebuild that visibility for your own roles, risk, terminology, and operating cadence.
Protect the design rule
In a regulated context the retrieval chain is the product's defensibility. Hiding it is a commercial decision, not a design simplification. Keep that rule in the acceptance criteria, component states, and production QA record so later visual cleanup cannot erase why the interaction exists.
Measure behaviour after ship
The evidence record is 3 weeks for time to compliance approval, from indefinite. Recreate the baseline and outcome window before rollout, segment the result by role and context, and state clearly what the measure cannot prove.
RESEARCH RECORD
The work behind the interface.
These artefacts connect the final interaction back to the evidence and product model that produced it.
Regulator interview series
Six interviews with former supervisory staff across RBI-regulated and pharma contexts on what a defensible record actually contains.
Trace-comprehension test
Tested the chain with 9 compliance officers who had no ML background; three vocabulary revisions before comprehension hit ceiling.
Corpus versioning model
Designed the document-version and embedding-run relationship with the ML team before any interface was drawn.
Export format review
Audit export reviewed by two Big Four assurance practices for filing suitability.
“Compliance stopped being the blocker. The trace is now the first thing we show them.”
Head of Platform, RAG infrastructure · under NDA
NEXT STUDY · 04 / 30 · AI TOOLS