The problem
The client is a 60-person professional services firm whose operations team came to us wanting "AI automation for the whole back office," with a list spanning contract review, expense processing, client onboarding paperwork, internal reporting, and more. It's a common starting point and, left unscoped, a recipe for a project that never ships, because automating everything at once means nothing gets the attention needed to work reliably.
We pushed to scope this down to one process first, chosen by actual time-cost data rather than whichever process felt most frustrating anecdotally, and prove it worked before expanding.
Picking the first process with data, not opinion
In the first week, the operations team logged the time spent on each candidate process over a representative week, instead of relying on gut sense about what was most painful. The data surprised the client: client-onboarding paperwork review, which staff called merely "annoying," consumed more hours per week than contract review, which everyone had assumed was the bigger problem because it felt more tedious.
Architecture
Process mapping before any automation
We mapped the onboarding review process in detail. A new client's intake forms and supporting documents (ID verification, engagement letter, conflict-of-interest disclosures) had to be checked for completeness and internal consistency before a case could formally open. A coordinator did this by manually cross-referencing four documents against a checklist, which took roughly 20 minutes per new client, with the firm onboarding 15 to 20 new clients a week.
AI-assisted document completeness and consistency checking
We built a tool that ingests a new client's four intake documents, checks each against the required-fields checklist, and flags specific inconsistencies (a name spelled differently in two documents, a date mismatch, a missing signature) with a clear description of what needs review, rather than a vague "something's wrong." It mirrors what the coordinator was doing, faster and more consistently; a person doing the same repetitive cross-check twenty times a week is prone to exactly the attention lapses the tool doesn't have.
Human-in-the-loop review workflow, not autonomous approval
The tool produces a review packet for the coordinator (which documents were checked, what was found, and a clear pass or needs-review status), but the coordinator makes the final call on whether a case can open, given the real legal and compliance stakes around conflict-of-interest checks. We scoped it explicitly to make the coordinator faster and more consistent, not to replace their judgment, which the firm's compliance lead said was a requirement, not a preference.
Integration with existing case management system
Rather than a standalone tool the coordinator had to check separately, the review results appear in the firm's existing case management system as a new case-status field and review checklist, so they fit the workflow the coordinator already uses daily.
Challenges and tradeoffs
- Document format variance across intake channels was the main engineering challenge. Some clients submitted intake documents as scanned PDFs, others as filled digital forms, and others as a mix, and getting consistent extraction across these formats took more iteration than the consistency-checking logic itself.
- Defining "consistent" required real input from the compliance team, not just engineering judgment. A minor name variation (a middle initial present on one document, absent on another) is a normal, low-risk inconsistency; a mismatched date of birth is a serious flag. We built this severity distinction with the compliance lead directly, because getting it wrong in either direction (too many false flags, or a missed real one) had real costs.
- Change management with the coordinator mattered as much as the tool's accuracy. The coordinator whose role this tool touched most was, understandably, wary at first that it was meant to replace her. We involved her directly in defining the review packet, which improved the tool's usability and made her an internal advocate once she saw it made her job faster without replacing her judgment.
Results
Onboarding review time per client dropped from roughly 20 minutes of manual cross-referencing to about 4 minutes of reviewing the tool's flagged packet, saving roughly four to five hours of coordinator time a week at the firm's onboarding volume. In an audit sample the compliance team ran after the first month, the tool's flags matched what a manual review would have caught 97% of the time; the remaining 3% were mostly genuinely ambiguous cases that went to review regardless.
The firm has since approved a second phase automating contract review, citing the first process's proven reliability and the coordinator's positive experience as the reasons they were comfortable expanding.
What we'd do differently
We'd bring the affected staff member (the onboarding coordinator, in this case) into the process-mapping conversation from day one rather than after the initial design was drafted. Her early input on what actually mattered in a "needs review" flag would have saved a design iteration and likely shortened the change-management period too.