Enterprises typically lose 2-5% of supplier spend to contracts that are mis-billed or under-leveraged. Volume tiers, indexation clauses, rebate thresholds, rate-change cutover dates — pricing rules that read as one paragraph in a contract but apply across tens of thousands of invoice lines — routinely go unchecked, because nobody re-derives that arithmetic by hand every month.
A spot check might sample a handful of high-value invoices once a year. The rest goes unexamined, month after month. Numbers are plausible but not reproducible, and there's no way to check the reasoning behind them.
Checks every invoice line against every applicable contract rule, every month. Every finding includes the contractual evidence and the calculation behind it — outputs ready for procurement and finance to action directly with suppliers.
Every supplier, every document, one view of what's actually in force.
Superseded amendments, replaced price lists, duplicates, things that should have died years ago — flagged, with the paper trail.
Every clause compared against your own standard: payment terms, auto-renewals, indexation, liability. A risk map, not a reading assignment.
Every invoice line recomputed from the contract.
Rebate tiers reached, compensation clauses triggered, credits earned — and never claimed.
Notice windows and auto-renewals, tracked before they close.
Commercial clauses are a compilation problem. We study how language models can translate pricing terms into formal, human-verifiable rules, and where that translation fails. Extraction is probabilistic, execution is deterministic, and the boundary between the two is the core design question of the system.
Invoice lines rarely reference the agreement that governs them, or name the product the way the contract does. Matching them is entity resolution under noisy data, at both the supplier and the product level. Our linkage models output calibrated match probabilities, and we study when a link is strong enough to justify a verification claim.
Most LLM systems cannot tell you how sure they are. We develop methods for calibrated confidence in zero-shot data extraction and reasoning, so every finding carries a defensible probability and weak evidence yields cannot verify rather than a guess.
ReliaParse is created by a team of academics and commercial experts based at the University of Helsinki.
15+ years of research in computational social science, including Oxford and University College London. Leading research work.
Widely published in statistical methods and LLM applications. Industry experience as a data scientist.
Runs pilot relationships, scopes engagements with partner organisations, and maps the route out of the research phase.
20+ years of entrepreneur in tech related startups & corporate venturing. Coordinating go-to-market & AI risk mitigation.
Background in basic research in signal processing, and expertise in commercialisation in the digital domain.
Not yet. We’re working with a small number of design partners to validate this on real contracts and invoice data before general availability. Early partners get hands-on access and a direct say in what we build next.
Your contracts (or the relevant pricing sections) and your invoice or billing history for the same supplier relationship. We don’t need your whole document archive — just the commercial terms and the bills.
No — it runs alongside your existing payments and operational data. Nothing to migrate, nothing to replace.
It can run inside your own environment, so your data doesn’t need to leave your systems. We work within whatever data handling arrangement you require, and we’re happy to sign a DPA before any documents are shared.
Because “probably right” isn’t good enough when the output is a dollar claim against a supplier. An LLM can misread a tier threshold or invent a missing rate and sound equally confident either way. We use deterministic, auditable arithmetic for every number we report — LLMs help us read contracts faster, but they never get the final say on a dollar figure.
If you manage supplier contracts with volume tiers, rebates, or indexation clauses — and suspect (or know) some of that isn't being billed correctly — we'd like to run your data through this and show you what we find, at no cost, in exchange for your feedback.
