Most consulting case studies describe one engagement, because one engagement is what there was. This one describes three for the same client inside a year — which is the part worth paying attention to, and the only part that cannot be manufactured.
How one engagement became three
Fix the finance queue
Expense validation built for the auditor: three-to-six months down to one.
Get it into production
Live in Mexico and extending to Asia — not a pilot.
Free the field force
Route optimisation: 13% less travel, an hour back per rep per day.
Point them at the right people
Smarter targeting for that freed-up hour: 1.5x ROI.
A validation queue measured in months
Around 2,000 invoices arrived every month as documents rather than as data. Each expense event had to be read by a person, matched against policy, and cleared — and the full validation and clearance cycle ran three to six months per event, with a meaningful share coming back as rework.
The obvious framing is document extraction: read the invoice, apply the rules, done. That framing fails, for a reason that has nothing to do with model accuracy. Expense validation is an audited process. A model that is right most of the time but cannot show its work isn't a partial solution — it's an unusable one, because the controls environment has no mechanism to accept it.
We stopped trying to remove the reviewer, and started making sure every minute they spent landed on something that deserved it.
So we built for the auditor first. Documents are digitised and extracted into structured fields; a validation layer combines deterministic policy rules with ML judgement; and crucially, each item is scored and routed by confidence — clear-cut cases resolve themselves, and only genuinely ambiguous or high-exposure items reach a human queue. Automated data-integrity checks run continuously and write their own evidence, which strengthened audit readiness rather than straining it.
An hour a day, back in every rep's pocket
With the first system live, the client brought us a different problem: field execution. Their sales force in Latin America was losing a significant share of each day to travel, and coverage suffered for it.
We architected and deployed an AI-driven route optimisation engine. The design decision that mattered was ownership: rather than building a tool per market, we built it as a centrally owned engine, so that every improvement compounds across territories instead of being rebuilt in each one.
Then: making that hour count
Freeing up a rep's time raises an immediate second question — where should they spend it? That became the third engagement: a unified engagement solution for the field force covering healthcare professionals, built on smarter targeting and adaptive promotional planning, with ML-based recommendation models integrated into the global marketing outreach pipeline.
Why they kept coming back
Three engagements in a year is not a procurement outcome. It is an operational one — it happens because the first thing you build keeps working after you leave.
Two things compounded. The first is that we got each system into production, which is where the difficulty actually sits. A model in a notebook is a hypothesis; a model in production has survived contact with real data, real users, and a real controls environment. The second is cheaper than the first: by the second and third engagement we already understood how the business ran, which meant less time spent learning the context and more spent on the problem.
The sequence itself tells the story. Validation freed the finance function. Routing freed the field force. Engagement told the field force where to go. Each one made the next one more valuable — and none of them would have happened if the first had stayed a pilot.
If you're facing something similar
Two things generalise. First: when a process resists automation, the blocker is usually accountability, not accuracy. Find out who has to sign off on the output and what evidence they need, design for that person, and the automation question becomes much easier to answer.
Second: be suspicious of any proposal whose success criterion is a demo. The expensive part of applied ML is not getting a model to work — it is getting it into a place where the business can rely on it, and keeping it there. That is the only part that earns a second engagement.