CSM Workspace
Playbooks
AI Product Design
Document Intelligence & Human Validation
An engagement drifting toward a custom build closed on native capability plus one missing platform primitive that is now live in the product.
A global investment bank's engagement had no agreed scope and a requirements register growing faster than anyone could resolve it. Both sides were frustrated and the work was heading for a custom build. Nobody had mapped what the customer needed against what the platform already did, so every unmet need looked like something to build. It was a renewal year in a strategic licensing area.
MY ROLE
UX Advisory Lead
TIMELINE
3 weeks, joined mid-engagement
SCOPE
No agreed scope
A requirements list is not a scope
The register I inherited had hardened into positions, and the room had deep domain command with incomplete visibility into what the platform already did. Triaging the list would have relitigated all of it. I started from the journey instead, identified pain points as the customer described them, and mapped each against existing capability. Most items resolved as already served or as misapplied capability they owned. A small number were real.
Journeymap with identified UX focus
I designed against sensitivity, not extraction
Reading the document was never the hard part. A wrong match was, because in this data many legal entities map to a single client. I turned their four problem statements into four responses: one match resolved at a time, explicit progress through the set, a review-all gate before exit, and guidance available in place. Validation runs on both sides, with every model-populated field marked, traceable to source, and blocking on submit until a person resolves it.
Early exploration of agents verifying fields
Requestor flow with NowAssist validation
I took the one real gap to product before offering it to the customer
One thing in that model had no native equivalent: no way to mark a value as verified, which is the primitive the whole validation layer depends on. A renewal year made a custom build the expensive answer for both sides. I refined it into a product opportunity and took it to product, design, and engineering for sign-off before presenting it as an option. Accepted, committed, shipped.
Agent DocIntel verification check modal
What I'd revisit
The model blocks on submit until every flagged field is resolved, which guarantees a human looks. It does not tell anyone whether the human looked carefully. I would now scope a signal for that, resolution time per field or override rate against source, so the validation layer can be tuned rather than assumed.
In their words
“Transformed a single Q&A session into multiple collaborative working sessions, navigating a critical renewal period with the client.” - Account Manager
"Even with a difficult and often uncooperative client, Nick handled the engagement with professionalism, patience, and humor. He proved to be both technically adept and highly skilled at managing difficult client maintaining Zen-like patience that helped diffuse tension and keep focus on results. This was a win." - Project Owner