Compliance
Regulatory research assistant
Can the organization show which corpus, source version, prompt, answer, and analyst edits supported a filing response?
AI systems and auditability engineering for regulated workflows
Can you reconstruct an AI decision six months later?
Audit Grill-Me helps regulated organizations find AI workflow auditability gaps before regulators, auditors, customers, or litigators do.
$9,500 fixed scope10 business daysBuilt for audit, risk & compliance
decision_record · loan_adjudication · 2024-11-086-month replay
model.versioncould not determine
prompt.templatenot retained
retrieved.sourcesno manifest
reviewerj.okafor (inferred)
approval.trailoutcome_2024-11.csv
evidence.retainednot retained
5 of 6 fields cannot be reconstructed. This is the auditability gap.
The auditability gap
Most AI risk is not model risk. It is workflow risk.
Your organization is deploying AI. But when a decision is questioned months later, the evidence needed to defend it lives in the gaps between systems, people, and tools.
Step 01
Input
Step 02
AI system
Step 03
Human reviewer
Step 04
Decision
input.datarequest.context
model.versionprompt.templateretrieved.sources
reviewer.identityoverride.history
approval.trailevidence.retained
7 of 9 evidence points are never captured. The failure is in the workflow, not the model.
The six-month reconstruction test
For one real AI decision, can you produce all eight?
Pick a workflow where AI touches a decision. Count what you could produce today, with evidence, not from memory. The honest count is your starting position.
0 / 8
Reconstructable evidence
If your team cannot produce the evidence trail, the decision exists but the record does not.
Run the full risk check01Model versionProvider, model, and configuration behind the output
02Prompt / system instructionThe exact system and user prompt, fully rendered
03Input dataThe input as it was at decision time - a snapshot
04Retrieved sourcesDocuments, chunks, and ranking the model retrieved
05Original outputThe output as first generated, before any edit
06Human reviewerWho reviewed it - and what they actually saw
07Approval / override recordWho approved or changed it, when, on what basis
08Retained evidenceStored, immutable, and time-bounded
Audit Grill-Me evidence readiness sprint
A fixed-scope assessment, not an open-ended engagement.
Ten business days. One workflow. A defensible answer to a question your auditors will eventually ask, delivered as evidence, not opinion.
Audit Grill-Me sprint
$9,500fixed
10
business days
1
workflow
5
deliverables
0
code changes required
Accepting applications; normally one active Audit Grill-Me Sprint at a time.
Artifact 01
AI Workflow Evidence Map
A plain-English map of where AI is used, what inputs it consumes, what outputs it creates, who reviews it, what systems touch it, and where evidence is or is not retained.
Artifact 02
Six-Month Reconstruction Test
A pass/fail stress test of whether the company could reconstruct a specific AI-assisted decision six months later.
Artifact 03
Auditability Gap Register
A severity-ranked list of missing evidence, weak controls, unclear ownership, retention failures, logging gaps, source-linkage failures, and human-review weaknesses.
Artifact 04
Audit Committee Briefing Memo
A concise, board-ready memo explaining the workflow, auditability risks, evidence gaps, and recommended remediation sequence.
Artifact 05
30/60/90-Day Remediation Backlog
A practical action plan showing what to fix first, what can wait, and what should be implemented before the workflow scales.
Built by Invariant Engineering
A narrow test for a load-bearing risk.
Audit Grill-Me combines software engineering, auditability, and regulated-workflow judgment. It is built for the teams who have to answer when an AI-assisted decision is challenged.
Ian Hafkenschiel
Founder, Invariant Engineering Group
01
US-trained lawyer and senior software engineer
02
Built from production AI workflow constraints
03
Designed for internal audit, risk, compliance, and governance teams
The thesis behind the sprint
The Auditability Gap
The book names the problem: non-reproducible intelligence in regulated systems. Audit Grill-Me turns that argument into a test for one real workflow.
Where the test applies
One method, several regulated workflows.
The public simulator uses fictional examples. The private sprint applies the method to one approved workflow under agreed handling constraints.
Insurance
Claims triage assistant
Can a disputed routing decision be tied back to the exact model, inputs, override, and retained evidence?
Legal
Contract review assistant
Can the original clause finding, suggested redline, lawyer change, and source library version be reconstructed?
Audit Grill-Me
Test whether you can defend it before someone asks.
One workflow, ten business days, a fixed price, and an evidence packet your audit committee can read.