Running a collections floor from a single source of truth

Executive summary

A US tax-credit firm had done the hard part most companies never finish: its operations team had built genuinely full client context – every call, email, and document, per account, in one place. And it still couldn't run its collections floor well. That is the whole lesson of this engagement. Context was not the bottleneck; acting on it was. What we built sits entirely on top of context: summaries that say what a file's state is, prioritization that says who to work today, a dashboard that says whether the right work is actually happening, and a chat layer for the questions no fixed report anticipates.

The business

The founder came out of banking and finance, and made his money going into unglamorous places most people avoid – the equivalent of a plumbing business, run with computers and phones. One of those places is the tax-credit space: firms that help other companies claim tax breaks for innovation or for retaining employees through a crisis like COVID-19. The model is simple to state: full service, no upfront payment, and a cut of the proceeds taken only at the very end.

Why it's hard

That model has a specific shape. Acquisition is easy – it's free to start – so clients sign readily and relationships run long, with a lot happening in between: calls, emails, documents requested and chased. And a lot can go wrong in that gap, from a client going out of business to a client who gets their IRS check and then goes quiet on the fee they owe.

The economics are what make it brutal (and profitable): The firm effectively pays close to two years of salaries before it can expect a return. So you cannot give every client full attention at every moment – there is a long task list and a need for proactive check-ins, and the whole game is operational efficiency: work the right files, at the right time, without dropping the ones that go dark. Although the analogy is unconventional, this resembles an enterprise sales cycle with lots of interactions and work before monetization takes place.

The firm knew this and invested heavily in systems and process. Its operations team achieved what we'd call full client context – the complete picture, per account, in one place. And here is the twist the rest of this story turns on: full context was not the solution. Having everything in one place is table stakes. It tells you nothing about where things stand with a customer, how the team is doing, and what to do next.

Got a similar situation? Let's have a chat.

What we built

Everything we built sits one step past context – because that is where the value was:

Insight – summaries. A per-client summary that compresses the full context into the current state of a file: what's filed, what's owed, what's blocking, what happened last. It replaced the 15-minutes-of-prep-before-a-call ritual with a single trustworthy view.
Guidance – prioritization. A queue that orders the floor's day from real internal state – value at stake, stage, reachability, time zone – so a rep opens the account that actually deserves the next hour, not a static list.
Control – the dashboard. The management layer, in three questions: are the reps doing the required work, are the clients meeting their obligations, and where is effort quietly hiding risk (team, customers, bottlenecks). Scored pass or fail, not activity theatre.
Flexibility – chat over the platform. An MCP-based chat across the whole system, so the questions no fixed report anticipates still get answered without a new build each time.

The one rule everything hung on: the numbers the floor acts on were anchored to the firm's self-built platform, not their CRM; a setup all companies naturally converge towards. A collections rep who acts on a stale number once stops trusting the system forever – so the money-critical fields had to come from the system that was actually correct, with CRM fields and call notes secondary. That is the single source of truth in the title, and it is what made the summaries safe to act on.

And one thing we purposely did not build: an automatic generator for the legal narratives. It would have demoed beautifully and manufactured exactly the false confidence this engagement was trying to remove – fluent "legal-ready" output while the real evidence assembly stayed manual. Not building it was the right call, and the harder one to defend in a status meeting.

Fun fact, most companies do this in reverse: They bolt an AI chat onto their tools and databases first, then discover it mostly inflates the volume of information sloshing through their systems and their people's heads. The limiting factor was never the system side. That's why chat is the last layer here, not the first. And it is by far not the most important one.

Results

The metrics that matter to a floor like this are operational: connection rate, prep time, resolution on the call, overdue tickets and SLA breaches. Those are the levers on the cash-conversion cycle and OpEx – which is the actual business outcome. What we can put hard numbers on:

Caller prep: 15+ minutes to about 30 seconds. The pre-call scramble effectively disappeared.
~$2.7M in fees and refunds made visible – the firm's first single structured view of what was owed, to whom, and at what stage.
10,000+ calls graded against one rubric, turning "how is the floor doing" into a real baseline instead of anecdote.

An honest boundary: the dashboard scores behaviour and compliance, not collected dollars – deliberately, because the firm keeps revenue attribution on its own books. So the proof here is capability gained, not a collections lift I can claim.

The most interesting result wasn't a number, it was about adoption. A good tool does not get used just because it exists – it took an executive mandate plus measurement to move people off their routines. The first person to adopt it unprompted was a new joiner who, given the choice between this and the incumbent HubSpot workflow, simply picked the one that told them what to do. For the founder, what changed was control, visibility, and efficiency. For the operators, it was quieter: less dread before a call, more confidence they were working the right file.

How it actually went

Projects at this scale rarely go as streamlined as advertised, and this one didn't either. Both sides made calls that had to be corrected – which is the part most case studies leave out.

The invisible milestone. For months, leadership believed a piece of the legal-narrative work was automated. It wasn't – it was being assembled by hand, with one person as the throttle. It surfaced only when someone finally walked a single file end to end. The lesson the firm took – verify a milestone against the real workflow, not a status update – is the exact muscle the reporting was built to create.
The deadline that slid. A company-wide "escalate everything by the end of April" directive – dated, absolute, the right kind of target – was missed. What mattered was that the founder named the miss plainly and kept the remediation going, rather than quietly reframing it.
Ours. We scaled summaries from the four-rep pilot to more books before locking down who owned the loop between "a rep flags this is wrong" and "the fix lands." Cheap to fix at four reps, expensive at forty.

The founder put the whole thing better than I can:

"I was under the presumption that things were happening automatically. I got corrected – it's being done manually. So all the milestones I felt we achieved were actually not achieved."

Catching an untrue version of progress before it costs you – that is the product. Not the dashboard, and not the context underneath it. The floor being unable to lie to itself.