How do you choose AI governance software?
Choose AI governance software by deciding first where it must run, then writing requirements for inventory, assessment, approval, enforcement and evidence. Ask each vendor to show every requirement working, not described. Run a short proof of concept on your own traffic and a real approval, and score the candidates against weighted criteria with the evidence you collected.
The market uses the same words for different things. "Governance" can mean a register with questionnaires, or a control point that refuses a request. Requirements written before the demos keep the comparison honest.
Start with the deployment model
Where the software runs decides what data it may see, which in turn decides what it can do. Settle this before comparing features.
| Model | What it means | What it implies |
|---|---|---|
| Multi-tenant cloud | The vendor runs one service for many customers | Quick to start. Prompts and logs inspected by the product usually leave your network, so check contracts, regions and retention |
| Dedicated tenant | The vendor runs an instance for you alone, in its cloud or yours | Better isolation and region choice. The vendor usually still operates it and has some access |
| Self-hosted | You install and operate the software on your own infrastructure | Prompts and evidence stay with you. You own upgrades, backups, availability and capacity |
If prompt inspection must happen without data leaving your network, that single constraint may remove most of the shortlist. Find out early.
Requirements and questions to ask
| Area | What you need | Questions to ask |
|---|---|---|
| Inventory | A register of AI systems with owner, purpose, vendor, models and lifecycle status | Which fields are required? How is a record kept current when a model or vendor changes? |
| Discovery | Finding AI use nobody declared | Which sources does it read: proxy logs, DNS, endpoints, traffic? Does a finding become a record automatically or after a person confirms it? |
| Assessment | Questionnaires for systems, vendors and agents | Are questionnaires versioned? What happens to past answers when a question changes? |
| Risk classification | A tier for each system with reasons | Can it show which factor produced the tier? Is an unknown answer treated as safe? Who can override, and is the reason recorded? |
| Approvals and separation of duties | A recorded decision with conditions and an end date | Can a requester approve their own request? Do approvals expire? What happens when facts change after approval? |
| Enforcement at the point of use | Decisions applied where requests pass | Which traffic paths are covered: API, browser, agents? Does a refused request reach the model provider? |
| Evidence and audit | Records an auditor can check | Is the log append-only and tamper-evident? Can evidence be exported in full, in an open format? |
| Framework mappings | Links from controls to the frameworks you report against | Is status computed from live records or entered by hand? Which controls are not computed, and are they labelled? |
| Integrations | Identity, logging, ticketing, GRC | Which integrations ship today and which are on the roadmap? What does it send out, and to where? |
| Deployment and data residency | Running where your constraints require | Where are prompts, logs and evidence stored? Does it work without internet access? How are updates delivered? |
| Cost visibility | Usage and spend by team and model | Is cost recorded per request or imported in bulk? Are estimated savings kept apart from measured spend? |
The categories of vendors in this space
Several kinds of product are sold as AI governance. Each starts from a different place, and many buyers end up combining two.
- AI governance platforms. Built around the register, assessments, approvals and reporting against AI frameworks. Check how far they reach into enforcement.
- AI security platforms. Built around runtime protection of AI applications and models, such as prompt inspection and threat detection. Check how decisions and ownership are recorded.
- GRC suites with AI modules. Extend an existing risk and compliance platform with AI registers and questionnaires. A natural fit if the suite is already your system of record.
- SSE and CASB vendors. Extend web and cloud access controls to AI services, often with data loss rules. Strong on access, usually lighter on per-system governance records.
- LLM gateways. Route and control API traffic to model providers, often with cost tracking. Strong at the point of use, usually lighter on assessment and approval.
Red flags
- Risk scores nobody can explain. A number with no factor-level reason cannot be defended to an auditor or challenged by an owner.
- No record of who approved. If the product cannot show the approver, the basis and the date, it records status, not decisions.
- "Compliant" badges. Software can hold evidence and map it to requirements. Whether you comply is a determination for your organisation and its counsel.
- Evidence that cannot be exported. Records you cannot take out in full are a lock-in risk and a problem for external auditors.
- Mandatory vendor cloud for prompt inspection. If sending prompts to a third party is a constraint for you, this is a blocker, not a detail.
A proof of concept in five steps
- Agree scope and success criteria. Pick three to five real AI systems, the traffic paths to cover and the outcomes that count as a pass.
- Install in your environment. Use the deployment model you would use in production, with your identity provider and network rules.
- Build the inventory. Load declared systems and run discovery against your own logs. Compare what it found with what you expected.
- Run a full decision. Take one system through assessment, classification, review and approval by someone other than the requester, then change a fact and see what happens.
- Test enforcement and evidence. Send allowed and disallowed requests, then export the evidence and check it with your audit team.
A scoring sheet you can copy
Set weights that add up to 100 before the demos start. Score each criterion from 0 to 5 only on evidence seen, and multiply by the weight.
| Criterion | Example weight | Evidence to request |
|---|---|---|
| Deployment fit and data residency | 15 | Architecture diagram, data flow, where each data type is stored |
| Inventory and discovery | 15 | Discovery run on your logs during the proof of concept |
| Assessment and classification | 15 | A classified system with the reason for each factor shown |
| Approvals and separation of duties | 10 | An approval record with approver, basis, conditions and end date |
| Enforcement at the point of use | 15 | Logs of allowed and refused requests from your test |
| Evidence and audit export | 10 | A full export reviewed by your audit team |
| Framework mappings | 5 | Mapping for your frameworks, with computed and manual controls labelled |
| Integrations | 5 | Working connections to your identity and logging systems |
| Cost visibility | 5 | Spend by team and model from test traffic |
| Vendor maturity and support | 5 | Roadmap, support terms, security documentation |
How MERIDVAR fits
MERIDVAR combines a governance record with enforcement at the point of use. It is self-hosted: a single-tenant appliance on a Docker host on your private network, with no MERIDVAR cloud and no telemetry. It runs without internet access once installed, including signed licences and signed updates.
- An AI inventory from traffic and declarations, where an observation becomes a record only when a person confirms it
- Three versioned questionnaires and rules-based risk classification, with a reason for every factor and no numerical score
- Approval by an administrator who did not make the request, with conditions and an end date. An approval reads "review required" when facts change
- An admission gateway that applies policy such as approved model lists, prompt DLP and team budgets, where a refused request never reaches the model provider
- An append-only, hash-chained evidence log, audit packages and control mappings with computed status for the EU AI Act, NIST AI RMF and ISO/IEC 42001
Weigh it honestly: the product is in beta, controls apply only to traffic routed through the gateway, you operate the appliance yourself, and MERIDVAR does not certify compliance. See deployment and the product overview.