A buyer’s guide By Nick Jang, founder · updated 26 October 2026
Choosing AI governance software, for a Canadian bank.
The kinds of tool on the market, twelve questions to ask any vendor before OSFI’s E‑23 takes effect on 1 May 2027, and a four-week pilot that shows the answers.
01 The kinds of tool
- GRC platforms
- Built for policies, controls, risks and audits across the enterprise. Strong on workflow and reporting; models and agents are usually one more register among many.
- Model risk and validation tools
- Built for the model lifecycle and its tests: inventory, tiering, validation, monitoring. Strong on the statistics; agents and vendors may sit outside.
- AI governance platforms
- Built for AI policies, risk assessments and regulation mapping across many jurisdictions. Strong on breadth; the depth of any one regulator varies.
- Agent gateways and policy engines
- Built to sit in the path of an agent's tool calls and allow or refuse them. Strong on enforcement; the inventory, the approvals and the examiner's record usually live elsewhere.
02 Twelve questions to ask any vendor
Ask them of every tool on your list, ours included.
| The question | Why it matters |
|---|---|
| Does it hold models, AI systems and AI agents in one inventory? | E-23 asks for one enterprise view; agents and vendor models that sit in other tools are the ones that get missed. |
| Does it use E-23's own fields and risk rating, or a generic template you must adapt? | An examiner reads against the guideline; a mapping you maintain yourself drifts. |
| Does it enforce separation of duties, or only record who did what? | A tool that lets the developer approve their own model only documents the gap. |
| Can a high-risk move be blocked until the evidence is on file? | A door that opens with fields missing turns the record into a to-do list. |
| How are third-party models and AI vendors covered under B-10? | Vendor models are inside E-23's scope, and the supplier arrangement is inside B-10's. |
| For AI agents, is the policy enforced before each tool call, and by what? | A policy that is written down but not checked at the call is guidance, not a control. |
| Is a person required to release high-risk agent actions, and is the developer kept out? | Human approval only counts when the right person, and not the builder, gives it. |
| Can someone named stop an agent at once, and is that recorded? | The kill switch is the control examiners ask to see working. |
| Can an examiner verify the record without trusting the vendor? | Hash chains, time-stamps and an offline verifier let evidence stand on its own. |
| Where does it run: vendor cloud, your cloud, your data centre? | Data residency and outsourcing rules may decide the deployment before features do. |
| Does it work in French as well as English? | Federal institutions and Quebec teams need both. |
| What does a pilot prove, and on whose data? | A pilot on your own inventory shows the gaps that matter to you; a demo shows the vendor's. |
03 A four-week pilot that shows the answers
- Week one
- Import your existing model inventory spreadsheet and see which E-23 fields are owed, model by model.
- Week two
- Walk one high-risk model through review and approval, and try to approve it with the developer as the approver.
- Week three
- Put one AI agent under a written policy and try a call it should refuse and one it should hold for a person.
- Week four
- Export the evidence and have someone outside the project verify it without the vendor in the room.
04 Where MINERVA fits
MINERVA Studio is the record and the controls for Canadian federally regulated institutions: one inventory of models, AI systems, agents and vendors in E‑23 and B‑10 terms; approvals that refuse the developer and wait for evidence; agent policies as code with held calls and a kill switch; and a hash-chained, time-stamped record an examiner verifies offline, in English and French, on our cloud or in your data centre.
What it does not do: it does not host model validation compute (tests run in your own notebook through our Python SDK and are recorded against the model), it does not run your agents, and it does not make an institution compliant.
05 Questions people ask
What software helps Canadian banks with OSFI E-23?
Model risk management and AI governance software that keeps one inventory of models, AI systems and agents in E-23's own fields, rates them, enforces review, approval and separation of duties, monitors them, covers third-party models under B-10, and keeps a record an examiner can verify. GRC platforms, model validation tools, AI governance platforms and agent gateways each cover part of this; ask the twelve questions on this page of any of them.
What should AI governance software for a bank do?
Keep one inventory of models, AI systems and agents; rate them; enforce review, approval and separation of duties; hold moves until the evidence is on file; cover vendors; enforce agent policies before each tool call with a person for high-risk actions and a kill switch; and keep evidence an examiner can check without trusting the vendor.
Is a GRC platform enough for model risk management?
It can hold the register and the workflow. Check whether it rates models in E-23's terms, blocks approvals until evidence is on file, keeps the developer away from approval, and covers AI agents and vendor models; many institutions pair a GRC platform with a model risk or AI governance tool.
How do you run a pilot of AI governance software?
On your own data, for about four weeks: import your inventory, walk one high-risk model through approval, put one agent under a policy, and have someone outside the project verify the exported evidence.
Where does MINERVA Studio fit?
MINERVA Studio is the record and the controls for Canadian federally regulated institutions: one inventory of models, AI systems, agents and vendors in E-23 and B-10 terms, approvals that refuse the developer and wait for evidence, agent policies as code with held calls and a kill switch, and a hash-chained, time-stamped record an examiner verifies offline. It does not host model validation compute; validation tests run in the bank's own notebook through its Python SDK and are recorded against the model.
Compared: OSFI E‑23 software, side by side. Platform: OSFI software · OSFI E‑23 platform · Agentic AI governance · AI vendor risk. Guides: OSFI E‑23 · Agentic AI governance · AI vendor risk · Policy as code · Know Your Agent. A plain-language guide, not legal advice.