Company profile · 2 min read
Bead AI: Evidence-linked SOX testing
Bead AI is selling an evidence trail for internal audit work.
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
Published · Updated
Bead AI is selling an evidence trail for internal audit work.
What it does
The homepage says Bead can follow an existing SOX testing plan, pull evidence, test controls, identify exceptions and produce working papers with a decision log. That sequence matters. The product is not only generating an audit summary; it is claiming to preserve the evidence and the reasoning behind each judgment in the format an audit team already uses.
The page says the system can automate 80% of manual work and test 100% of controls while reducing testing time by 70%. Those are Bead’s claims. They are useful hypotheses for a pilot, not independent audit results.
The security page adds a second decision. Bead says it has SOC 2 Type II, US data residency, AES-256 encryption at rest, tenant isolation, zero-retention inference, and that evidence is not sent to model providers or used for model training. Those are vendor commitments to verify with procurement and the actual contract.
Buyer and task
The buyer is an internal-audit team that spends too much time copying evidence, marking up documents and rebuilding the same workpaper structure. The first task should be one control with a messy exception, not a clean sample. Upload the existing testing plan, trace each requested artifact, and inspect how the reviewer sees the decision log and linked evidence.
The important boundary is human judgment. An agent can gather, test and draft. The auditor still needs to decide whether the evidence is sufficient, whether the exception is real and how the work should be signed off.
The first test
Take one completed control test and one exception-heavy test. Compare the evidence trail, the working paper output, the missing-data prompts and the correction path. Then ask how a changed testing plan is versioned. If a reviewer cannot reconstruct why the system reached a conclusion, the automation has moved the copying burden without solving the audit problem.
Sources checked
official Speedrun profile · company homepage · security
Sources checked — September 20, 2026.
About the author
Cofounder of Lazyweb (a16z Speedrun). Previously grew subscriptions and built native ads at Duolingo. Writes First1000 (90,000+ subscribers).
I cofound Lazyweb and publish Mudpie. This is an owner-written publication, not an independent testing organization. Research notes distinguish observations, sourced reporting and editorial judgment.
First1000 ↗ · X ↗