Actor
Who or what acted?
Assure Evidence and control for AI agents
Know what your AI did. Prove it.
Assure shows what AI agents did, whether they were authorized to do it, and what evidence supports that conclusion.
Keep your logs. Assure connects them into evidence you can investigate, test and independently verify.
Evidence chainVersion binding
Approval mismatchThe approval existed.
The wrong version was sent.
The wrong version was sent.
01 / The problem
AI agents don't just answer.
They act.
Edit filesCall toolsGenerate documentsChange recordsSend messagesMake recommendationsTrigger workflows
When something goes wrong, a log line may not answer the question that matters:
“What actually happened?”02 / The shift
Logs show activity.
Assure shows evidence.
Keep your logs. Assure does not replace them. It connects events to actors, exact versions, authority and outcomes.
Raw logs
10:21 document.updated
10:24 approval.completed
10:26 document.sentLogs are events.
Evidence is context.
Assure
- 10:21 Purchase Agreement v3 created
- 10:24 Human approved Purchase Agreement v3
- 10:25 AI created Purchase Agreement v4
- 10:26 Purchase Agreement v4 sent
FindingSent artifact did not match approved artifact.
There was an approval. There was a send. But the approved document was not the document that left the system.
03 / Evidence model
Six questions.
One connected record.
Mandate
What was it allowed to do?
Artifact
What exact thing changed?
Approval
Who approved which version?
Action
What actually happened?
Outcome
What happened next?
04 / How it works
From event trail
to reviewable evidence.
Assure Core remains deterministic and evidence-first. It warns and evaluates; it does not autonomously block customer systems.
- 01Record
Collect the minimum required evidence.
- 02Connect
Link related actions, artifacts and approvals.
- 03Test
Run deterministic business-control checks.
- 04Warn
Surface evidence-backed exceptions.
- 05Reconstruct
Replay what happened later.
- 06Export
Produce a scoped evidence package.
05 / Flagship workflow
Demo / SyntheticAn AI agent
prepares an offer.
No real customer data. No protected form content. A synthetic real-estate transaction workflow makes the control visible in under five minutes.
01Transaction data
02Forms
03Conditions
04Document
05Human approval
06External delivery
AI creates APS v7Version commitment recorded
Broker approves APS v7Approval bound to the exact hash
AI changes closing dateSystem creates APS v8
APS v8 is sentExternal delivery observed
Approval mismatchExternal artifact did not match the approved version.
Assure does not decide whether the contract is legally correct. It answers a narrower question: Was the document sent the document an authorized person approved?
Enforce is disabled.Assure warns and evaluates. The application itself may require human approval.
06 / Authority
AI agents need
a mandate.
Assure records which mandate applied when the action happened.
It records and evaluates authority evidence. It does not claim to enforce the boundary unless a separate application gate actually does.
Transaction agent
Read
Draft
Request approval
Send Outside recorded boundary
07 / Controls
Test the business control.
Not just the software.
Results are evidence states, not optimism. Missing evidence displays as UNKNOWN—never as a pass.
Every externally sent agreement has an approval for the exact sent version.
EXCEPTIONNo material edit after approval without new approval.
EXCEPTIONEvery agent action maps to a valid principal and mandate.
PASSEvery external delivery can be reconstructed.
PASSCustomer A evidence never appears in Customer B's tenant.
PASSUnderlying private content was independently verified.
UNKNOWN08 / Replay
When someone asks “Why did this happen?”
start with evidence.
01Source event
02Document version
03Agent action
04Approval
05External action
06Outcome
Assure reconstructs observable evidence. It does not claim access to hidden model chain-of-thought.
09 / Optional intelligence
Evidence first.
Intelligence second.
Assure Core works without these systems. Their outputs are not automatically fact.
Assure Core · deterministic evidence
Brain
AION
AOK
Sentinel
Institutional memory.
Have we seen this before? What changed? Was it fixed?
Investigation.
What hypotheses explain this finding? What test could falsify them?
Prioritization.
Which investigation is worth running first?
Warnings.
Is behaviour changing in a way that deserves attention?
Current boundaryBrain, AION and AOK remain governed optional layers. Sentinel is experimental, uncalibrated and Observe/Warn only. No block or intervention claim.
10 / Open Agent Observatory
Tested against real
open-source agent systems.
We study public metadata and controlled, synthetic runtime behaviour. We do not observe project users, and project maintainers do not endorse Assure.
Registered projects and harnesses at exact public pins.
Recurring, synthetic, network-bounded runtime evidence.
No real-model claim. Deterministic local adapters only.
11 / Regulatory direction
The direction
is clear.
Across jurisdictions, AI governance is moving toward stronger accountability, documentation, testing, monitoring and human oversight. Applicability depends on role, system and context.
Accountability under existing privacy law.
Canadian privacy regulators emphasize appropriate purposes, safeguards, transparency, accountability, traceability and human review. Canada does not have one universal private-sector AI-agent statute.
Privacy commissioners · 2023 ↗Risk-based rules, phased over time.
Enforcement and selected transparency and GPAI duties apply in 2026; specified high-risk obligations apply later. Not every AI agent is a high-risk system.
European Commission · updated 2026 ↗A targeted and sectoral landscape.
No single general federal equivalent to the EU AI Act. Federal activity, sector obligations, state laws and voluntary frameworks—including NIST AI RMF—coexist.
Congressional Research Service · 2025 ↗Lifecycle model-risk expectations.
OSFI E-23 applies to federally regulated financial institutions, not automatically to every software vendor. Vendors may still face downstream evidence and diligence demands.
OSFI · effective 2027 ↗Educational information only. Assure does not provide legal advice or guarantee compliance.
12 / Assurance market
AI assurance is becoming
an enterprise category.
The world's largest assurance firms publicly describe work in AI governance, risk, controls, testing, documentation and assurance.
PwCPublic AI-assurance services covering governance, controls, monitoring, testing and evidence-based review.PwC US · 2025-06-03 ↗DeloittePublic AI controls and assurance services spanning governance, risk, controls, testing and reporting.Deloitte US · 2026-08-12 ↗EYEY describes AI assurance work around governance, model objectives, training/validation procedures, data reliability and management controls.EY Canada · 2026-01-01 ↗KPMGPublic AI-assurance services covering risk assessment, control testing, model validation, evidence and reporting.KPMG US · 2025-09-25 ↗
Assure is infrastructure for assurance.It is not the independent auditor. No partnership or use of Assure by these firms is claimed.
13 / Frameworks
Map evidence to the frameworks
your organization uses.
NIST AI RMFMapped
Descriptive risk and governance themes; no certification claim.
ISO/IEC 42001Partial
Evidence can support selected management-system activities; no certification coverage.
SOC-type assurance conceptsPartial
Control, sampling, workpaper and retest concepts only; no SOC report.
OSCAL exportNot yet mapped
Capability matrix records the tested OSCAL adapter as missing.
Primary references: NIST AI RMF · ISO/IEC 42001 · NIST OSCAL
14 / Privacy and deployment
You shouldn't have to centralize
every prompt and document.
Metadata-first evidence
Record the minimum evidence needed for a control or reconstruction.
Hash and reference binding
Bind versions without claiming a hash proves content is correct.
Redaction and tokenization
Reduce exposure while recognizing tokenized values may still be personal data.
Customer-controlled artifacts
Keep content under customer control where the deployment supports it.
Deployment boundaryPrivate-edge and BYOC references are implemented and CI-evidenced, not claimed as active customer deployments. Some controls require authorized access to underlying content.
15 / Review
Give reviewers evidence.
Not screenshots.
Assure can prepare structured evidence. Independent assurance remains independent.
Scoped evidence packageReviewable · Exportable · Verifiable
01actors02agents03versions04mandates05artifacts06approvals07actions08exceptions09control results10remediation11retest
Source manifest Reviewed 2026-09-16
Principles for responsible, trustworthy and privacy-protective generative AI technologies
Accountability, appropriate purposes, safeguards, transparency, traceability and human review under applicable Canadian privacy law.
- Publisher
- Office of the Privacy Commissioner of Canada and provincial/territorial privacy authorities
- Published
- 2023-12-07
- Last reviewed
- 2026-09-16
The enforcement framework of the AI Act
The Act applies progressively; enforcement and selected transparency/GPAI duties apply in 2026, with specified high-risk-system dates later.
- Publisher
- European Commission, Directorate-General CONNECT
- Published
- 2026-08-24
- Last reviewed
- 2026-09-16
Regulating Artificial Intelligence: U.S. and International Approaches and Considerations for Congress
No broad federal AI regulatory authority had been enacted; U.S. activity remained a mix of targeted federal provisions, agency authority, voluntary measures and state laws.
- Publisher
- Congressional Research Service
- Published
- 2025-06-04
- Last reviewed
- 2026-09-16
Artificial Intelligence Risk Management Framework 1.0
The AI RMF is a voluntary framework for incorporating trustworthiness considerations into AI design, development, use and evaluation.
- Publisher
- U.S. National Institute of Standards and Technology
- Published
- 2023-01-26
- Last reviewed
- 2026-09-16
Guideline E-23 — Model Risk Management (2027)
Federally regulated financial institutions are expected to apply risk-based, lifecycle-wide model governance, inventory, review, approval and monitoring; effective 2027-05-01.
- Publisher
- Office of the Superintendent of Financial Institutions
- Published
- 2025-09-11
- Last reviewed
- 2026-09-16
ISO/IEC 42001:2023 — Artificial intelligence management system
Requirements for establishing, implementing, maintaining and continually improving an AI management system.
- Publisher
- International Organization for Standardization
- Published
- 2023-12-18
- Last reviewed
- 2026-09-16
OSCAL releases and compatibility commitment
OSCAL is a machine-readable control and assessment model. Assure's adapter remains not yet mapped and makes no compatibility claim.
- Publisher
- U.S. National Institute of Standards and Technology
- Published
- 2023-11-08
- Last reviewed
- 2026-09-16
PwC Launches First-to-Market Solution to Provide Assurance for AI
Public AI-assurance services covering governance, controls, monitoring, testing and evidence-based review.
- Publisher
- PwC US
- Published
- 2025-06-03
- Last reviewed
- 2026-09-16
Deloitte Expands End-to-End AI Controls and Assurance Capabilities
Public AI controls and assurance services spanning governance, risk, controls, testing and reporting.
- Publisher
- Deloitte US
- Published
- 2026-08-12
- Last reviewed
- 2026-09-16
Our commitment to audit quality — EY Canada 2026
EY describes AI assurance work around governance, model objectives, training/validation procedures, data reliability and management controls.
- Publisher
- EY Canada
- Published
- 2026-01-01
- Last reviewed
- 2026-09-16
KPMG expands AI Trust services with new AI Assurance capabilities
Public AI-assurance services covering risk assessment, control testing, model validation, evidence and reporting.
- Publisher
- KPMG US
- Published
- 2025-09-25
- Last reviewed
- 2026-09-16
Start with one workflow
Bring us one
AI workflow.
We'll show you what can be recorded, reconstructed and tested.
Assure Core · Evidence and controlEnforce disabledNo compliance claim© 2026 One Lab