Back to Blogs
Revenue Earned, But Not Collected
The document problem within the cash conversion cycle – and why frontier models can’t fix it.
There is a category of revenue that companies have already earned and cannot yet collect. The work has been completed, the obligation is real and undisputed, yet the money is sitting on the wrong side of a piece of paper that no one has gotten to yet.
Finance teams rarely describe it that way. They describe it as Days Sales Outstanding (DSO), or as a billing backlog, or as a staffing problem in accounts receivable. But underneath those descriptions is usually the same mechanic: a transaction cannot be billed until a human reads a document, pulls specific values out of it, and confirms those values match what was supposed to happen. Multiply that by thousands of transactions a month and you get a structural delay between delivering value and being paid for it.
We ran into a clear version of this in liquid fuels logistics, where every physical movement of product generates a bill of lading, and no invoice can be issued until someone works out what that document says. But the pattern is not just specific to the energy sector. It shows up anywhere the paperwork trailing a transaction is inconsistent, high-volume, and legally load-bearing. Which is to say, nearly everywhere money changes hands between businesses.
Four other industries facing the same revenue bottleneck
Healthcare revenue cycle
Provider organizations do not usually lose money because a payer disagrees about coverage. They lose it because a claim went out with something missing. Experian Health’s 2025 State of Claims survey found the top three denial drivers were missing or inaccurate data, authorizations, and incomplete patient information – in other words, documentation, not medical disagreement.
The scale of the leak is significant: an analysis of more than 2,300 hospitals put net revenue leakage at $48.4 billion in 2025, up roughly 25% from the year before. Most of those denials are administrative, and most eventually get overturned. The revenue arrives late, or it arrives after rework costs have eaten the margin, or it doesn’t arrive at all because no one had the staff hours to appeal.
Trade finance
A letter of credit is, functionally, a promise to pay against paperwork. Banks examine documents on their face, and any mismatch – a consignee name spelled differently on the bill of lading than on the invoice, a shipment date one day past the stated window – makes the presentation discrepant and freezes payment.
International Chamber of Commerce (ICC) estimates place the share of presentations refused on first presentation between 65%-80%. That means that the majority of a payment mechanism designed specifically to give exporters certainty, routinely fails on the first attempt for reasons that have nothing to do with whether the goods shipped.
Construction
A subcontractor who has completed work waits a median of roughly 50 days after submitting a pay application, and a large share of that wait is not structural retainage – it’s simply package completeness. A pay application is a bundle: schedule of values, lien waivers from every tier with a claim, certificates of insurance, certified payroll, change order backup. One missing signature two tiers down stalls the whole billing. Rabbet’s 2025 report estimated slow and inconsistent payment cost U.S. construction around $299 billion, functioning as a tax on an industry that is not especially profitable to begin with.
Freight and logistics
The proof of delivery is the trigger for the entire payment cycle. Until the signed document comes back and gets reconciled against the bill of lading and the rate confirmation, the invoice cannot be matched by accounts payable, and it sits. When the two documents disagree – 40 pallets on the BOL, 38 with two damaged on the POD – the discrepancy has to be resolved by a person before any money moves.
These are all different industries with different regulators and different documents, but they’re all facing the same core problem. Revenue that exists, but is gated by a document-reading task.
So, why hasn’t AI already solved this?
On paper, this is one of the most obvious use cases in the enterprise. High volume, narrow scope, clear success criteria, direct line to cash. And yet the results have been dismal in aggregate. MIT’s NANDA research found that roughly 95% of enterprise generative AI pilots delivered no measurable impact on the P&L. RAND put the AI project failure rate above 80%, about twice the rate of conventional IT projects.
Those numbers get cited constantly as evidence that the technology is overhyped. That’s not entirely wrong, but there’s more to it than most headlines would lead one to believe.
- The tool is sized wrong for the job. The instinct over the past three years has been to route everything to whatever sits at the top of the leaderboard. For open-ended reasoning, that instinct is correct. For pulling six fields off a shipping document ten thousand times a month, it is expensive, slow, and frequently less accurate than a smaller model trained on that specific document type. NVIDIA researchers argued this directly in Small Language Models are the Future of Agentic AI, estimating that 40%-70% of enterprise AI tasks can run on sub-10-billion-parameter models. A frontier model knows a little about ten thousand things. A billing pipeline needs a great deal about one.
With respect to the economics, frontier capability is priced for the long tail of general reasoning that an invoice pipeline will never touch, and at production volume that premium is paid on every single transaction. Latency compounds it again – the difference between sub-100-millisecond and multi-second response times stops being an engineering detail when it sits inside a workflow that runs continuously.
- The output can’t be defended. A CFO cannot put a number in a filing, an auditor cannot sign off, and a bank cannot release funds against a figure whose origin is a black box. In regulated commodity flows – RINs, low-carbon fuel standard credits, customs declarations, clinical documentation – the requirement is not just the right answer, it is the ability to point at the specific document, the specific field, and the specific reason the system read it that way.
Automation without provenance produces a system that works beautifully in a demo and then stalls permanently at the pilot stage, because the people accountable for the numbers won’t stake their signature on output they can’t trace. No amount of frontier capability without governance and explainability at the architecture level can overcome this.
What right-sizing actually looks like
- Match the model to the task, not to the headline. Some steps in a document pipeline genuinely need strong reasoning – reconciling an ambiguous discrepancy against contract terms, for instance. Most steps don’t. A system that can route each step to an appropriate model, rather than sending everything to one endpoint, costs less and performs better on the narrow work that makes up the bulk of the volume.
- Don’t marry a single provider. The AI infrastructure market is volatile enough that a five-year commitment to one model family is a bet on which vendors are still standing in five years. Nobody has that information. Architecting for substitution is not hedging; it is the only defensible position when the underlying market is repricing itself quarterly and roadmaps are ephemeral, at best.
- Preserve the source, not just the extraction. The document is not disposable input to be discarded after the fields are pulled. Kept and made queryable, it becomes the thing that answers the counterparty dispute, supports the audit, and lets someone trace a number back to its origin months later without going to find a filing cabinet.
- Measure in days, not in accuracy percentages. Extraction accuracy is a proxy metric. The question a finance leader is actually asking is how many days sit between delivering the product and collecting the cash, and what one of those days is worth. That number is usually already calculated somewhere in the treasury function. Start there and work backward.
A short diagnostic
If you want to know whether this problem exists in your organization, three questions get you most of the way:
- Is there a step in your billing process where cash cannot move until a person reads a document?
- Do those documents arrive in inconsistent formats from parties you don’t control?
- Would a number extracted from one of those documents need to survive an auditor, a regulator, or a counterparty dispute?
Three yeses describe a document-gated revenue process, and it’s one of the places in the enterprise where AI has a short, legible path to the balance sheet.
Sources: MIT NANDA GenAI Divide; RAND Corporation AI project analysis; Experian Health State of Claims 2025; Kodiak Solutions revenue cycle analysis via Fierce Healthcare; ICC Banking Commission / UCP 600 discrepancy estimates; Rabbet 2025 Construction Payments Report; Belcak et al., NVIDIA, “Small Language Models are the Future of Agentic AI” (2025); The Hackett Group Working Capital Survey.
