Skip to content

Resources

We publish what we measure.

Including the parts that do not flatter us.

Most writing about enterprise AI is opinion with a statistic attached. We run AI systems in production for real companies, so we can publish what we observed instead. Every study states its method, its sample and what it cannot tell you, and every row can be downloaded under CC BY 4.0.

Published

Version 1.1Published 21 August 2026Updated 7 November 2026

The Agent Failure Index

Half of them reported success.

Every production failure Advizr has recorded, with its root cause, the signal it gave at the time, and whether it was ever closed. Published work on agent failure studies coding assistants in an IDE or frameworks in a lab. This is business automation running against real client data.

4,308 failures, 207 systems, 146 days

Read the study CSVJSON

Version 1.0Published 18 August 2026

The AI Evidence Index

43 of the 93 report no sample.

The enterprise AI debate runs on about a dozen numbers, and almost nobody quotes them correctly. What each study measured, its sample, and how it gets restated wrong.

93 studies, 77 publishers, 2024 to 2026

Check a statistic CSVJSON

Not published yet.

Listed before the result is known, on purpose. Saying what is being measured is part of what makes the result worth reading.

  • Later

    Model performance by task area

    Which models clear a written rubric on real business tasks, at what cost per passing case, ranked by the lower bound of the confidence interval rather than the headline rate.

  • Being counted

    The Agent Failure Index, version 1.1

    The same failure ledger, recounted on a much larger corpus, with closure measured: how many failures got a written fix and a prevention rule, and whether a rule ever stopped a repeat.

How we publish.

  1. 01

    Every number states its sample.

    Where a rate comes from a small sample we say so, and publish the count beside the rate. A rate from 30 cases and a rate from 3,000 are different claims that look the same once rounded.

  2. 02

    Client data is anonymised, never blended.

    A descriptor in our research is one real engagement, never a composite of several, and the anonymisation gate runs before any analysis and again on what we publish.

  3. 03

    Corrections are versioned, never silent.

    If a number or a sentence changes, the change is dated and listed on the study, with what it said before. A row that changes bumps the version.

  4. 04

    Our interest is printed.

    Advizr sells AI systems, which is a commercial interest in every subject here. The limits are written to be usable against us, and the rows are published so anyone can check.

Every version and correction, dated.

  1. The Agent Failure Index: version 1.1 published4,308 failures, 14 Jun 2026 to 7 Nov 2026, pinned to snapshot 333b898. The corpus is twelve times version 1.0's. Layer, signal, detection and prevention move to a stratified sample of 401 with published intervals, because coding 4,308 records to the same standard was not affordable and printing a bare share would have been the exact misuse this programme exists to catalogue. The per-row cause sentence is withdrawn: a name scan found 431 person-shaped strings in the corpus prose, and no denylist closes that. The crosswalks to MAST, the OWASP agentic list and EU AI Act Article 73 are dropped rather than carried forward uncounted.
  2. The Agent Failure Index: correction, prose onlyThe fifth finding said the published taxonomies cover four percent of this. Four percent was MAST alone. The sentence about the verification layer now compares it with the product code and the agent loop, which is what the counts support. No row and no figure changed.
  3. The Agent Failure Index: correction, prose onlyThe page said the closure endpoint was still failing. It was fixed on 24 August. The page said most notes came from an automatic capture path. That path has never filed a note. The page said the EU AI Act's high risk rules applied from 2 August 2026. Regulation (EU) 2026/1744 moved that date before it arrived.
  4. The Agent Failure Index: version 1.0 published359 failures, 14 Jul 2026 to 20 Aug 2026, pinned to snapshot d152f8a.
  5. The AI Evidence Index: version 1.0 published93 studies, 77 publishers, 2024 to 2026.

Run your whole company on AI.

Tell us how your business runs. If we're not the right fit, we'll say so on the call.

Your custom strategy and systems, delivered in under a month. Your named people using them every week by day 30. Or we keep working for free until both are true.

The page ends on a hand drawn chart of the sky that draws itself as it comes into view, with a small sail ship crossing its open sky from a far star to a beacon and a small planet.

© 2026 Advizr AI Inc. · Vancouver, BC, Canada

Delivered in under a month. Used by your named people every week by day 30. Or we work free until both are true.