01
Audit trail
Every recorded action appended to a hash chain the database protects from edits — verifiable on demand.
Regulated governance
AI governance that leaves evidence.
Policies say what should happen. The platform records what did — and lets an auditor check it.
Botminds builds AI governance into the platform the agents run on, for regulated, mission-critical work at scale. Every recorded change by every person and every agent is sealed into a verifiable audit trail; signatures are bound to exactly what was signed; read rules hold on every read path; agents are scored on every publish, and workflows and meshes on a schedule; and configuration is versioned and can be rolled back.
BeforeWho may?
Roles at three layersplatform, workspace, project
Access policiescompany sign-in, two-step, who may invite
Read rulesper persona and per row, on every API read path
DuringWho decided?
Approvalsa person approves, rejects or edits
E-signaturesbound to exactly what was signed
Separation of dutiesmaker and checker, two different people
AfterCan we prove it?
Evaluation gatesa worse change is flagged, the verdict kept
Versionssnapshot, diff, roll back, freeze
Observabilitytraces, cost, failures, owners
One audit chainappend-only · each entry sealed with the hash of the one before
verify → intactWhat is AI governance?
AI governance is the set of policies, controls and evidence that keep an organization’s AI accountable: who may build and use it, what it may touch, how it is tested before release, how its decisions are recorded and reviewed, and how a change is undone. A policy document says what should happen. An AI governance platform enforces it and keeps the proof.
Scales out on KubernetesServices and workers add capacity as work arrives, and give it back when it is done.
Built for millionsMillions of documents, records and agent runs — the indexes and workers grow with them.
Any cloud, any regionAzure, AWS, Google Cloud or on-premises, in the region your data must stay in.
Always onRedundant services, rolling releases and continuous backups, for mission-critical work.
Secure and governedSingle sign-on, encryption, least privilege and a sealed audit trail.
The framework
An AI governance framework, built into the platform.
Most AI governance frameworks describe the same duties: know who acts, control what they reach, test before release, record decisions, keep people in charge, and be able to undo. On Botminds each duty is a control that runs.
02
E-signatures
A meaning — Reviewed, Approved, Released — bound to a fingerprint of exactly what was signed.
03
Roles and policies
Roles at three layers, workspace sign-in policies, and tools granted to each agent by name.
04
Read rules
Who may read a collection, and which rows — on every API read path, the assistant included.
05
Versions
Snapshots, a diff before you restore, rollback, and a freeze with its reason recorded.
06
Evaluation gates
Each agent publish scored on golden data against a pinned baseline, the verdict kept beside the change; workflows and meshes scored on a schedule.
07
Calibrated confidence
Scores fitted per field from reviewers’ decisions, so straight-through rests on measured evidence.
08
Observability
Traces, failures and model cost — and an action center that gives every fault an owner.
AI audit trail
An AI audit trail you can verify, not just read.
The Botminds audit trail is append-only at the database, sealed entry by entry with a hash chain, and checked on demand by a verify call that names the first entry that does not match.
entry 1041Priya Ramanapproved the payment run
prev
9f3c…e1seal
4b7a…02entry 1042Invoice agentwrote 14 values to INV-2207
prev
4b7a…02seal
c2d9…7fentry 1043Tom Beckersigned “Released” on INV-2207
prev
c2d9…7fseal
81e0…5achain head
81e0…5akept apart, so a cut-off end is caughtAt the databaseChanging or deleting an entry is refused. Retention follows the rule you set, oldest entries first.
VerifyIntactentries checked, seals recomputed, head matches
VerifyBroken at entry 1042what this entry records was changed after it was written
integrity
Sealed on write
Each entry stores the previous entry’s hash and a SHA-256 seal over its own content, set by the database itself.
proof
Your own seal
Keep the chain’s latest seal in an inspection file or a ticket; a later verify proves every entry before it.
review
Three feeds
Project, workspace and platform feeds, filtered by who, what and when, and exported as CSV or JSON.
timeline
A document’s whole story
One timeline joins stage moves, value edits, agent runs, re-extraction and indexing.
E-signatures
Electronic signatures with separation of duties.
Some decisions need a name on them: a batch released, a credit decision approved. A signature on Botminds is bound to a fingerprint of the record, checked against the rules of its meaning, and visibly superseded the moment the record changes.
- 1FingerprintThe server computes a SHA-256 fingerprint over the record’s declared content — for a document, the file itself.
- 2Prove it is youThe signer re-enters their password, and their sign-in code where two-step is on.
- 3Rules of the meaning✓Not an author of the record✓Did not sign it “Reviewed”else refused — separation of duties, on the audit trail
- 4Signed: ApprovedTom Becker · 14 Oct, 16:02valid
Later
The record is editeda content field changes after signing
The same signature now readssuperseded
Nothing is overwritten: the signature store is insert-only. People sign; agents read signatures and ask the right person to sign.
Meanings you declare
Each collection declares its meanings and who may sign with each — Reviewed by one team, Approved by another.
Authors kept apart
Whoever created the record, or changed a signed field, can be kept out of signing it.
Distinct people
Whoever signed one meaning can be kept from signing another on the same record.
Maker-checker on stages
A stage can require a different person to move a record out — checked for clicks, bulk moves, agents, timers and the API.
AI agent governance
AI agent governance starts with who may do what.
People get roles at three layers and workspace policies. Agents get the tools they are granted, and the changes they propose can wait for a person.
people
Three layers of roles
Platform roles for the installation, workspace roles per business unit, and project roles that list exactly which actions and screens each role may use.
sign-in
Workspace policies
Company sign-in, required two-step sign-in, who may bring people in, the role newcomers join with, and who may suspend an account.
agents
Agent permissions
Each agent reaches the tools granted to it by name. Choose which tools wait for a person; the run pauses in the conversation until someone decides.
review
Access Explorer
Who can reach what, at platform, workspace and project scope, in one view — and for one person, every role and rule that applies, and why.
review
View as
See the platform as a role or persona sees it, to check what they can reach before they do.
agents
A record of every run
Each step of each run is streamed, kept and replayable. Written guardrails and spend quotas sit beside it. See agent harness.
Data access
Read rules on every API read path — including the assistant.
A rule that hides rows on a grid but not in search, or in search but not in what the assistant cites, is not a rule. Collection read rules are enforced for the person asking on every API path that reads collection data, and they fail closed.
The ruleClaims collection
Who may readAdjusters
Which rowsregion = my region
Lockedchanged by workspace admins
✓
Record gridsrows and columns on screen
✓
Counts and statisticstotals never leak hidden rows
✓
Document lists and pagesviewer and downloads
✓
Searchkeyword and meaning
✓
The assistant’s knowledge searchwhat an agent may cite
✓
Piper’s toolswhat the builder can read
✓
Cards on pagesevery app built on the data
✓
Datasheets and record exportsthe same rule on the way out
Fails closed
Unknown settingtreated as restricted
Rules cannot be readthe request is refused
Knowledge search failssays it failed — never “nothing found”
Per persona, per row
A collection is readable by everyone or by named personas; a row filter gives each reader their own slice — their region, their book, their cases.
Counted by a test
A test walks every API read path. Each passes through the read gate or sits on a reviewed list with its reason; a new path joins one or the other before it ships.
Designed for shared agents
Read rules follow the person asking. Automations and API keys read what you attach to them — so attach to a widely reachable agent what everyone who reaches it may read.
Versions and change control
Change control: snapshot, compare, roll back, freeze.
Model risk management asks what changed, when and why — and whether you can go back. Configuration is versioned module by module, so an agent’s instructions or a collection’s schema can be compared and restored on their own.
- Snapshot
A named snapshot of a module — automatically on change, or by hand before a release.
- Compare
See exactly which settings differ from any snapshot.
- Roll back
Restore configuration after an explicit confirmation; your data stays as it is.
- Freeze
Hold a project or workspace steady; the freeze and its reason are recorded.
- Promote
Move a solution between environments as a signed package, secrets left behind.
Versions per agent
Agents, skills and prompts keep their own history, so one instruction change is visible on its own.
Preservation holds
Hold a document, a collection or a project and every delete path refuses until release — which is recorded.
Signed supply chain
Images signed and verified at release, Hub packages signature-checked, and a package carrying a secret refused at publish.
Testing before release
Evaluation gates: every agent publish scored, every verdict on record.
Generative AI changes behaviour when its instructions, model or tools change. Model risk management for generative AI needs tests that run on every change and a verdict recorded where reviewers will see it. On Botminds, agents are scored on every publish, and workflows and meshes nightly, weekly or on demand.
CandidateThe saved changenew instructions, model or tools
BaselineA pinned baselinethe run you chose to compare against
Same golden documentsCompared pair by pairPass-or-fail judges on a model you choose; repeated runs check the results are consistent
- flaggedA real dropthe whole 95% interval of the difference is below zero
- flaggedUnder the floorthe score is below the pass mark set for this automation
- floor decidesToo close to callthe floor alone decides — no false flag on noise
- flaggedNever measureda floor is set but nothing was measured — unmeasured is not passing
- passesHolds or improvespasses, with the comparison on record
Every agent publish you gate is scored, and workflows and meshes are scored nightly, weekly or on demand. The verdict is recorded beside the change, and a worse result is flagged for the people who approve it — in advisory or enforce mode.
inputs
Golden datasets
Inputs paired with the right answers, kept as collections with owners and read rules.
scoring
Pass-or-fail judges
Model judges answer yes or no. You choose the judge model, from the same family as the agent or another.
monitoring
Drift
Score distributions compared over time; a sustained rise in audited errors switches straight-through off for that field until a person restores it.
confidence
Calibration
Confidence curves fitted per field and promoted when they lower calibration error. See human in the loop.
Governance at scale
AI governance at scale: thousands of users, one chain per project.
Governance has to hold at the same volume as the work. Every person, agent, automation and integration passes the same controls on every request, and every recorded change lands in its project’s audit chain — in every workspace and every region.
Every actor
Thousands of peoplecompany sign-in, two-step
Thousands of agentstools granted by name
Automationsworkflows, meshes, schedules
IntegrationsAPI keys and connectors
The same controls, every request
- 1Sign-in policyper workspace: SSO, two-step, who may invite
- 2Roleplatform, workspace and project layers
- 3Read ruleper persona and per row, at every API read path
- 4Approval or signaturewherever the decision needs a person
- 5Sealedappended to the project’s audit chain
Every workspace, every region
Retail lendingEU region
Mortgage intakeintact
Credit decisionsintact
ClaimsUS region
Motor claimsintact
Fraud reviewintact
TreasuryAsia-Pacific region
Paymentsintact
Sanctions screeningintact
one audit chain per projectevaluations on every agent publish, and on a schedule for workflows and meshessnapshots of every moduleAzure · AWS · Google Cloud · on-premises
Every recorded change, sealed
Each recorded change by each person and agent is appended to its project’s chain, across thousands of users and agents, and any chain verifies on demand.
Read rules at every read path
The same rule answers on grids, counts, search, exports, cards and the assistant — enforced inside each service as it scales out.
Evaluations on every agent publish
Each gated agent publish is scored against its baseline as it ships, and workflows and meshes on a schedule, so the evidence grows with the estate, not with the review team.
SR 11-7 · EU AI Act · NIST AI RMF
What regulators ask for — and how the platform supports it.
AI governance for regulated industries comes down to a short list of questions that recur across supervisory guidance and law. The platform supplies the controls and the evidence your answers rest on; your policies, risk assessments and people own the answers.
| What is asked | Where it is asked | How the platform supports it |
|---|---|---|
| Keep records of what the system did, automatically | EU AI Act, record-keeping (Art. 12) · SR 11-7, documentation | Append-only, hash-chained audit trail with a verify call; a kept record of every agent run |
| Let people oversee, override and stop | EU AI Act, human oversight (Art. 14) · NIST AI RMF, Manage | Approve, reject or edit on the tools you choose; review stages; cancel any run; straight-through above a threshold you set or one derived from audited reviews |
| Help users interpret outputs | EU AI Act, transparency (Art. 13) | Each extracted value carries its source page, words and confidence; a preview shows exactly what an agent was told |
| Validate before use; analyse outcomes | SR 11-7, model validation · NIST AI RMF, Measure | Golden datasets, pass-or-fail judges, paired comparison with a pinned baseline, verdicts recorded on every agent publish |
| Monitor in use | SR 11-7, ongoing monitoring · NIST AI RMF, Measure and Manage | Drift statistics, a drift alarm per field, calibration error per field, an action center with owners |
| Declare and keep accuracy | EU AI Act, accuracy and robustness (Art. 15) | Calibrated confidence per field; thresholds certified from audited reviews at a stated confidence |
| Clear roles, policies and controls | SR 11-7, governance · NIST AI RMF, Govern | Roles at three layers, workspace policies, read rules, separation of duties on signatures and stages |
| Control and document change | SR 11-7, implementation and use | Versioned configuration with diff, rollback and freeze; signed packages between environments |
| Manage risk across the lifecycle | EU AI Act, risk management (Art. 9) · NIST AI RMF, Map | Evidence for each step, alongside your own risk assessment and model inventory |
How to use this table. Obligations under the EU AI Act follow from what your system does and your role in it; SR 11-7 is applied by your model-risk function; the NIST AI RMF is a voluntary framework. The table shows where the evidence for each comes from. See security for the platform’s certifications.
Side by side
AI governance platform vs policy tools vs assembled controls.
Buyers usually weigh three routes: a governance tool that catalogues AI use, point controls added to an existing AI stack, or a platform where the controls are part of how the AI runs.
| Catalogue and policy tools | Controls you assemble | Botminds | |
|---|---|---|---|
| What it is | Inventory, questionnaires, policy workflows | Logging, access and testing tools wired around your AI | The platform the agents run on, controls inside it |
| Audit trail | Of the governance process | Logs, if each team adds them | Append-only, hash-chained, verifiable |
| Access to data | Not its concern | Per tool, often inconsistent | One read rule on every API read path |
| Human sign-off | Attestations about the system | Built per workflow | Signatures and stage rules on each decision |
| Testing before release | Records that it happened | Separate pipelines | Evaluations on every agent publish and on a schedule, verdict kept |
| Rollback | Not its concern | Depends on each tool | Snapshots, diff, rollback and freeze |
| Scale | An inventory of systems | As far as each tool goes | Every user, agent, workspace and region |
They combine. A catalogue tool can list the Botminds agents you run; the evidence behind each entry comes from the platform. See AI agent security for how the platform is protected.
Questions people ask
AI governance, answered.
More from the platform: the full architecture, and how security works at Botminds.
What is AI governance?
AI governance is the policies, controls and evidence that keep AI accountable: who may build and use it, what it may reach, how it is tested, how decisions are recorded and how changes are undone. An AI governance platform enforces those controls and keeps the proof.
What is an AI governance platform?
Software that turns governance policy into running controls and evidence: access and approval rules, audit trails, testing gates, version control and monitoring. On Botminds these controls are part of the platform the agents run on.
What is AI agent governance?
Governance for AI that acts, not only answers: which tools each agent may use, which actions wait for a person, what data it may read, and a record of every step. See agent harness and human in the loop.
What is the difference between AI governance and data governance?
Data governance controls the data — quality, ownership, access, retention. AI governance adds the system that uses it: how models and agents are tested, approved, monitored and changed. Read rules on Botminds serve both.
How does SR 11-7 apply to generative AI?
SR 11-7 is US supervisory guidance on model risk management, which many banks apply to generative AI and agents: sound development, independent validation, ongoing monitoring and governance. Botminds supplies evaluation gates, drift monitoring, versioned change and an audit trail for your model-risk function to rely on.
How does Botminds support the EU AI Act?
With controls and evidence for duties such as record-keeping, human oversight, transparency and accuracy — an audit trail, approvals and review stages, sourced values and calibrated confidence. Which obligations apply depends on your use case and your role.
What should an AI audit trail record?
Who did what, when, to which record and with what result — for people and agents — in a form where a quiet edit would show. Botminds seals each entry with the hash of the one before and verifies the chain against a seal you can keep outside the platform.
How do you enforce separation of duties for AI decisions?
With rules that run. A signature meaning can keep a record’s authors and the signer of another meaning apart; a review stage can require a different person to move a record out. Both are checked at every door and refused attempts are recorded.
Does AI governance scale to thousands of users?
Yes. Every person, agent and integration passes the same controls on every request, each project keeps its own audit chain in every workspace and region, and evaluations run on every gated agent publish, and on a schedule for workflows and meshes.
Can we roll back an AI change?
Yes. Configuration is versioned in twelve modules. Compare any snapshot with the current state, roll back after confirmation, and freeze changes with a recorded reason. Your data stays as it is.
Governance an auditor can check.
Bring your control framework. We’ll walk it row by row against the platform and show the evidence each control leaves.