Show an examiner how your AI decisions were made.
A self-serve evidence pack for insurers and lenders that use AI in decisions about people: underwriting, pricing, claims triage, fraud scoring. Fifteen minutes of answers, one decision-log export, and the pack is on your delivery page in minutes.
$49 · Readiness Check: your answers only, no log upload, five minutes
Start your evidence pack See the sample report (PDF)
What you get
- A readiness score for one AI workflow, and a pass / partial / gap matrix across 24 checks: 9 measured from your decision logs, 14 from your answers about governance practices, 1 from your retention policy.
- A framework map against the NAIC Model Bulletin on insurers' use of AI (adopted in Pennsylvania as Insurance Notice 2024-04), ISO/IEC 42001, the NIST AI Risk Management Framework and the EU AI Act.
- Risk-ranked findings with the evidence behind each one, and a 30-day remediation plan sized to the days per week your team can give it.
- A verifiable ledger: every decision record replayed into a SHA-256 hash chain, with a 40-line verifier you run yourself. Any edit, removal or reordering of a record breaks the chain. The ledger and the framework control mapping are open source: github.com/bickfordd-bit/decision-ledger.
- Delivered as
report.pdf,findings.json,ledger.jsonlandreport.html. The sample is the exact output for a synthetic insurer.
Why now
State insurance regulators expect insurers to keep a written program for AI systems and the evidence behind it, and to produce that documentation on request. If a vendor's tool makes or shapes the decision, you are still the one asked. The pack tells you where you stand before anyone asks, in writing you can hand over.
What a Pennsylvania examiner may ask about your AI: a one-page summary with sources.
More, each sourced and checked: the Notice 2024-04 documentation checklist · the NAIC AI Model Bulletin explained · what "Not evidenced" means · how to verify an AI decision log.
How it works
- Answer the intake: the system, the workflow steps, the review rule, the approved model versions, and fourteen yes / partial / no questions about your practices.
- Upload one or two decision-log exports (CSV or JSONL, decision records only, no personal data) and map the columns from your browser.
- Pay by card. The engine maps your logs, scores the checks, replays the ledger and renders the PDF. Your delivery page updates on its own, usually within a minute.
Questions
Is this an audit?
No. It is a self-assessment generated by software. It does not make you compliant with anything; it gives you an evidence-based picture of one workflow you can put in front of your risk committee, auditor or examiner, and a plan to close the gaps.
What does "Not evidenced" mean?
Your answers or logs did not show it. It is scored like a gap, because an examiner treats it the same way.
What do the logs need to contain?
One record per decision. The more of these it has, the more can be measured: decision ID, timestamp, model name and version, input reference, output, score, amount, policy checks, reviewer, override and reason, reason codes. Columns can be named anything; you map them on the form.
Do I have to send personal data?
No, and please don't: export references or hashes, not names, addresses, dates of birth or health details. Files are stored in a private bucket in the United States and deleted after 30 days. Privacy and retention.
What is the $49 Readiness Check?
The same intake without the log export. You get the 15 answer-based checks scored in the same report format, and the 9 log-measured checks marked "not measured" with what the full pack would add. It is the honest cheap version: it tells you exactly where you stand on governance practices and exactly what a log export would prove. Most buyers start there.
What if it is not useful?
Reply to your Stripe receipt or email the address below with the order ID and you get a refund, no questions.
About
bickford is the AI-governance software of Bickford Technologies, Philadelphia. It turns AI workflow records into audit-ready evidence. Built by Derek Bickford (Drexel LeBow '06), 15 years in enterprise software, who also built the engine, its control mapping and its tests; every framework reference in the report carries the date it was checked against public sources.