SignalFinancial ServicesData story
A Receipt Is Not a Check
Unsiloed AI's headline number, 99.78%, shows what an index has to keep apart: where a figure came from, who else confirms it, and whether anyone has reproduced it.
Unsiloed AI's pitch is a receipt. The San Francisco startup's software pulls information out of tables, charts, and other parts of business documents, and TechCrunch reports that it "shows where the extracted information came from in the original document, so businesses can check the results rather than take the AI's output at face value."
It's also a good example of why the AI Value Index exists. AI is changing how businesses operate, and even making new kinds of businesses possible, faster than anyone is checking the results. Stories like this one carry real information, and Unsiloed is doing interesting work. But the numbers arrive with no common standard for what counts as evidence. So where is the rigor?
A receipt doesn't make a number right. It tells you where to go to argue with it. That gap is how AIVI reads any claim, and Unsiloed's own numbers show it clearly.
Three questions, not one. Where did the figure come from? Who else says so? Has anyone reproduced it? Provenance, corroboration and verification are separate, and a record needs all three kept apart. Unsiloed's receipt answers the first, value by value. TechCrunch's report gives no sign it answers the other two.
A receipt for a value isn't a receipt for a score. The headline figure is 99.78% accuracy across roughly 84,000 extracted data points from 542 financial documents. That is an aggregate. Checking it takes ground-truth labels and a scoring method, not the tool's own receipts. TechCrunch does not say how accuracy was scored.
A claim is not a deployment. AIVI's Index records named deployments with outcomes. Here, a bank and a law firm are unnamed, the 542-document customer is unnamed, and the one named customer, Bolna AI, comes with no accuracy figure. What exists is a claim, and its standing is the story.
The publisher is not the originator. The 99.78%, the roughly 84,000 and the 542 all originate with the company, shared in a technical brief. The benchmark scores also originate with the company, from tests it conducted. TechCrunch printed them, and noted that the 99.78% results "have, however, not been independently verified." AIVI's rule is that a press repeat is not corroboration.
One organization, several roles. Unsiloed is the subject here and the AI vendor in any deployment. Bolna AI is Unsiloed's customer and, as a voice AI startup, a builder of voice agents. Vendor and deployer are roles in a relationship, not types of company.
Two tests, two standings. On olmOCR-Bench, a public benchmark from Allen AI covering 1,403 PDF pages, the startup reports 88.0 against 66.0 for Reducto's agentic system and 39.9 for Unstructured. Someone else could, in principle, rerun that. The 542-document test used documents supplied by an unnamed customer, its comparison scores were not published, and no outsider can rerun it as reported. Same vendor, two self-reported scores, different evidence.
How AIVI records it
| Claim | Originated by | Reported by | Standing |
|---|---|---|---|
| 99.78% on ~84,000 data points, 542 documents | Unsiloed (technical brief) | TechCrunch | Not independently verified; scoring and comparison scores unpublished |
| 88.0 vs 66.0 vs 39.9, olmOCR-Bench | Unsiloed | TechCrunch | Company-run on a public benchmark; not independently reproduced |
| Bank, law firm, Bolna AI as customers | Unsiloed (Mishra) | TechCrunch | Two unnamed; no accuracy figure for any |
| $7.5 million seed, led by Felicis; seven people | Not stated in the report | TechCrunch | Reported; company site repeats the seed; YC page lists a team of 2 |
The receipt for 99.78% is the brief. AIVI did not find it published.
That is the standard AIVI is building: a data model that records, for every company it covers, where each claim came from, who else confirms it and whether anyone has reproduced it. The 99.78% may hold up. Right now, nobody outside the company has shown it.
How to read this: Every reported fact above, including the figures, benchmark scores, headquarters, headcount, customers and the $7.5 million seed round, comes from TechCrunch's November 18, 2026 report, except the company-site and YC entries in the table. The test figures and benchmark scores come from the company. The customer behind the 542 documents is not named, and the comparison scores were not published.
Get new AI Value Index records in your inbox.
Subscribe