Cleared invoices are the best training data you will ever own

Ask anyone who has built analytics inside a mid-market company what consumed the budget, and you will get the same answer: not the models, not the dashboards — the cleaning. Finding the data, deduplicating it, reconciling three systems that disagree about the same customer, filling the gaps, and then defending every number in the first management meeting.

Now consider what an e-invoicing mandate quietly hands you. Every invoice your company issues or receives under a clearance regime has been forced through a validation gauntlet before it counts: structured to a government schema, checked against code lists, totalled correctly to the decimal, stamped with an authoritative timestamp, and tied to a counterparty whose tax registration the system verified. The output is a dataset with properties that internal data teams spend years failing to manufacture.

It is schema-perfect, because malformed invoices do not clear. It is timestamped, because the clearance event is recorded by an authority with no incentive to be vague. It is counterparty-verified, because registration numbers are checked in the loop. And it is complete — the property that matters most — because under a mandate, an invoice that bypasses the pipeline legally does not exist. There is no shadow population of transactions sitting in someone's inbox.

The maths of "already validated"

The economics here deserve a moment, because they invert the usual analytics business case. Conventionally, data quality is something you pay for after the fact: extract the records, profile them, fix them, reconcile them, and accept that some fraction will never be trustworthy. The cost scales with the mess, the work recurs forever, and the result is always probabilistic — cleaned data is data you believe, not data you know.

Cleared data reverses the sequence. Validation happens before the record exists, at issuance, enforced by a counterparty you cannot negotiate with. The cost of that validation is already sunk — you paid it as compliance spend, because the law required it. Every analytical use you stack on top inherits the quality for free. A reconciliation exercise that consumes an analyst's quarter in a conventional stack simply has no equivalent here: there is nothing to reconcile, because there is one stream, and it has already survived the strictest review it will ever face.

This is why we tell clients, only half-joking, that the tax authority has accidentally become their data quality department. The rejection log alone is a diagnostic — we have written elsewhere about rejections functioning as a master data audit nobody ordered. But the cleared archive is the bigger prize: the records that passed.

One caveat belongs here, because honest framing matters. The cleared stream describes your invoiced commerce — it does not contain payroll, inventory movements, or the handshake deals that never became invoices. It is not the whole picture of the firm. It is, however, the verified spine of the picture, and a spine is exactly what every other dataset in the company lacks. Joining weaker data to a strong spine is a tractable problem; joining weak data to weak data is the swamp most analytics programmes drown in.

What to do with it, in ascending order of ambition

Start with dashboards that carry no reconciliation debt. Revenue and accounts-payable reporting built on the cleared stream needs no caveats, no "subject to reconciliation" footnotes, no quiet asterisk about the subsidiary whose extracts arrive late. The numbers are the same numbers the authority holds. For most finance teams this alone justifies the plumbing: a daily view of issued and received invoices that nobody can dispute, because the disputing already happened at clearance.

Then read your counterparties. The stream records, with verified identities and authoritative timestamps, how every supplier and customer actually behaves: who invoices promptly and who in end-of-quarter bursts, whose credit notes cluster suspiciously after price changes, where your revenue concentration genuinely sits once group entities are resolved. This is commercial intelligence your CRM approximates and your cleared archive states.

Then forecast on your own flow. Cash-flow forecasting models are only as good as their training history, and most mid-market firms train on exports that are partial, lagged, and inconsistently coded. A cleared archive is none of those things. A model trained on your own validated flow learns your seasonality, your payment-term reality, and your collection patterns — not a vendor's multi-tenant average of companies that merely resemble you. This is where the dataset stops improving reports and starts informing decisions.

Finally, watch for what should not be there. Duplicate invoices, broken sequences, amounts that drift from contract terms, counterparty behaviour that changes without commercial explanation — anomaly detection on the cleared stream catches what sampling-based controls structurally miss, because it reads everything. The authority's analysts run exactly this kind of screening on your data already. Running it yourself, first, is not paranoia; it is symmetry.

The inversion

Notice what happened across those four steps. The first is reporting, the second is analysis, the third is prediction, the fourth is control — and all four run on an asset your compliance project produces whether or not anyone uses it. The mandate forced you to build a pipeline that emits the cleanest dataset your company has ever generated. The only open question is whether the dataset lands somewhere you can use it, or evaporates into a service provider's archive and a folder of PDFs.

That question is settled by one design decision, made at implementation time: treat the cleared archive as a product. Give it an owner, a data store inside your own boundary, a retention design that serves analytics as well as audit, and a roadmap that climbs the four steps above. The day you make that decision, the accounting changes character — what the budget called compliance spend is now also data-asset spend, and it is producing the training data for every financial model you will want in the next five years.

Firms that skip the decision will meet it again later, expensively: buying analytics tools to approximate a dataset they already owned, or renting forecasting models trained on other people's flows. The asset was on the premises the whole time. It cleared customs years ago. Somebody just has to claim it.