Search for AI governance tooling and the results are confident, well-designed, and aimed at someone else. The platforms on page one are built around a model registry: a system of record for the AI you develop, with risk classification, bias testing, and conformity documentation attached to each entry. A serious product, for a company that ships AI. If your whole AI footprint is ChatGPT, Claude, Gemini, and Microsoft Copilot subscriptions, the registry is empty on the day you buy it and stays that way.
Most of these pages are published by vendors ranking their own product — not a scandal, just a reason to read them as catalogs rather than advice. The useful question is not which platform scores highest. It is what posture your company holds, because posture decides whether you need a platform at all.
One phrase, three products
Three distinct categories share the label and the listicles mix them freely. They answer different questions and are priced accordingly.
| Category | What it does | Built for | Market range |
|---|---|---|---|
| Lifecycle governance platform | Registry of models and agents, risk tiering, control crosswalks, evidence and audit artifacts | Organizations that develop or deploy AI systems of their own | Rarely published; quoted per system and per framework |
| Compliance automation | Maps one control set to ISO/IEC 42001, SOC 2, ISO 27001; collects evidence; readies you for a certificate | Companies already running a security-compliance program | Add-on to an existing subscription |
| Data security and discovery | Finds unsanctioned AI use, classifies sensitive data in prompts, blocks or redacts at the boundary | Any company whose staff use assistants, which is all of them | Often a tier upgrade on tools you own |
Credo AI, IBM watsonx.governance, Holistic AI, ModelOp, Monitaur, and Trustible sit in the first column. Gartner published its first Magic Quadrant for the category in June 2026 — a signal that it is real, and a misleading one about who it is real for, since analyst coverage follows enterprise budgets. Verify features and pricing on each vendor’s own page; comparison posts age badly here.
The third column transfers down-market cleanly. A 200-person company has the same exposure as a 20,000-person one to an employee pasting a client contract into a personal account. It just has fewer models to govern — none.
The posture test that settles the shortlist
Every framework here draws a line the vendor pages blur. The EU AI Act separates the provider who develops an AI system from the deployer who uses one under its own authority. The NIST AI Risk Management Framework splits its four functions the same way: GOVERN operates across the whole organization, while MAP, MEASURE, and MANAGE apply to each AI system individually. That structure is why platform value and platform pricing both scale with the number of systems you own.
So count them. Three questions:
Do you develop or fine-tune models, or ship AI features in a product customers use? If yes, you are a provider, you have systems to register, and a lifecycle platform earns its price. Most companies in the 50-to-500 band answer no.
Do you use AI to make or materially inform decisions about people — hiring, promotion, lending, insurance, benefits, tenancy? If yes, you have a legal exposure that is a counsel question before a tooling question, and the one case where a small company genuinely needs per-system assessment.
Is your AI footprint licensed assistants plus whatever is embedded in the SaaS you already buy? If yes, you are a deployer with a wide tool surface and a short inventory. Your problem is visibility and behavior, not model risk.
Most of this segment lands on the third answer and buys for the first. That costs twice: the license fee, and the months of program-building a platform demands before it produces anything. The underlying argument is the one we make in the case against treating licenses as capability — a purchase is not a control.
What already ships in the console you pay for
Before evaluating a vendor, open the admin surface of the subscription you hold. A meaningful share of what the platforms sell is already there, gated behind a license tier rather than a new contract.
Microsoft documents this plainly for Copilot. Data Security Posture Management for AI in Microsoft Purview discovers AI activity, surfaces sensitive data appearing in prompts and responses, and runs data-risk assessments for oversharing. Prompts and responses land in the unified audit log, and eDiscovery can reach assistant interactions for a legal matter. Foundational controls sit at the A3, E3, and G3 tiers and optimized controls at A5, E5, and G5 — so the gap between you and a governed Copilot deployment is sometimes a license upgrade and an afternoon of configuration.
The business and enterprise tiers of the major assistants carry a comparable set: single sign-on, directory provisioning, workspace retention, admin visibility into membership, a contractual commitment that business data is not used to train models, and in higher tiers an API for retrieving conversation records. Read each vendor’s current trust page rather than any comparison table, including ours.
Two things follow. Consumer logins are the real hole — none of this reaches an employee on a personal account, which is why the sanctioned-account roster is the first artifact to reconcile. And if your controls are unconfigured, buying a platform to watch them is premature. The ordered version of this work is in the policy and training guide this sits under.
Which framework actually reaches you
Three regimes drive this market and reach a US mid-market company unevenly. The urgency in vendor copy rarely survives the test.
NIST AI RMF 1.0 is voluntary US guidance, published January 2023, organized as Govern, Map, Measure, Manage. It is not a certification and nobody can audit you against it. Its value to a deployer is vocabulary and a checklist, both free. Its generative-AI profile, NIST AI 600-1, is the more relevant document if assistants are your whole footprint.
ISO/IEC 42001:2023 is the certifiable one — published December 2023 as the first AI management system standard, awarded by independent accredited certification bodies. Pursue it when a customer, an insurer, or a procurement questionnaire asks for it. Not before.
The EU AI Act applies on a market-and-effect test rather than headquarters location, so a US company can be in scope by deploying AI systems in the EU or affecting people there. Two dates matter. Article 4’s AI literacy duty for providers and deployers has applied since 2 February 2025. And the high-risk timeline moved: under the Digital Omnibus package, the Annex III deadline shifted from 2 August 2026 to 2 December 2027, with Annex I deferred to 2 August 2028, while Article 50 transparency obligations stayed at 2 August 2026. Pages written before the change still quote the old date to create urgency. Test scope first, then take the specific question to counsel.
All three ask first for an inventory and a policy, and neither artifact needs a platform. The artifact-by-artifact version is in our guide to what each compliance regime asks you to produce.
What it costs, and which numbers to distrust
Pricing here is mostly unpublished, and the certification figures in circulation contradict each other by more than tenfold. The defensible breakdown is component-level. Vanta publishes ISO 42001 costs as readiness at $3,000 to $10,000 and up, implementation at $10,000 to $40,000 and up, the initial two-stage audit at $5,000 to $20,000, surveillance audits at $3,500 to $9,000, and ongoing monitoring at $3,000 to $10,000 a year. Auditor fees for a 21-to-50-person company land near $14,000 for full certification.
Elsewhere on the same results page you will find claims of $85,000 to $150,000 for a 50-to-200-person company and $180,000 to $320,000 for mid-market, with no source attached. Treat unsourced ranges as positioning. When a vendor will not publish, ask for the three inputs the quote is built from: AI systems governed, frameworks mapped, connected evidence sources.
For a deployer the honest budget is smaller and differently shaped: a license-tier upgrade for discovery and audit logging, a policy drafted or reviewed properly, and a facilitated session to move behavior, which sits in a $2,000 to $15,000 market range. A custom automation build, if warranted later, runs $25,000 to $150,000. A lifecycle platform belongs on that list only once you have systems to put in it.
Buy in this order
Sequence matters more than selection. Each step makes the next one cheaper, and skipping forward wastes the spend.
- Inventory. Every AI tool in use, including the ones embedded in SaaS you bought for other reasons and the personal accounts finance is reimbursing. Reconcile your identity provider, your expense data, and what people say when asked without consequence. McKinsey found C-suite leaders estimated 4% of staff use generative AI for at least 30% of daily work while 13% of employees reported doing so — three times the assumed penetration. Assume your estimate is low.
- Configure what you own. Sanctioned accounts for everyone, single sign-on, directory provisioning, retention set deliberately, audit logging on. Largest risk reduction per dollar, no new vendor needed.
- Write the policy. One page: sanctioned tools, prohibited inputs, where a human must review output, how to report a mistake. A platform cannot produce this; the judgments in it are yours.
- Train the people the controls will not cover. Covered below.
- Add discovery if your surface is wide. Dozens of SaaS products with AI features, or contractors outside your directory, justify it. A single assistant on managed devices does not.
- Buy a lifecycle platform when you have systems to register, a certificate to earn, or a regulator asking per-system questions. Reassess every two quarters; this market moves fast enough that waiting is often the better trade.
Whoever runs this is usually an operations leader or chief of staff holding it as a second job. Whether they and the board above them can read what they oversee is a separate purchase, covered in the guide for leaders and directors.
What no platform can do
Every control here enforces at a boundary: block an upload, mask a string, refuse an unsanctioned account, write a log line. Genuine value, and also the limit.
What it cannot see is the quality of a judgment inside a permitted action. An employee uses the sanctioned tool, on an approved document, within policy, and accepts an output containing a confident error that reaches a client. Nothing flagged, because nothing was violated. Meanwhile the clearest sign that governance is working — people reporting their own near-misses — appears on no dashboard, because reporting is a behavior rather than an event a system can observe.
That residual is why the practiced half matters. People who have used an assistant on their own real documents, with a facilitator pointing out where it was wrong, develop a reflex to check. People who have read a policy develop a reflex to comply with it, which is not the same thing. The behavioral layer underneath the controls is the subject of our guide to security and risk awareness training that fits AI use.
If you want an outside read on your posture before signing anything, book a free AI-readiness call. We run hands-on fluency sessions with teams on their own documents and daily work, and we will say plainly when a license upgrade and a rewritten policy would serve you better than anything we sell.
Frequently asked questions
Do we need an AI governance platform at 50 to 500 people?
Usually not yet. The capability these platforms are priced around — a registry of AI systems with per-system risk assessment and conformity documentation — has nothing to hold if your footprint is licensed assistants. Configure the controls you already pay for, write a policy, and buy a platform when you develop AI systems of your own, pursue a certificate, or face a regulator asking per-system questions.
What is the difference between an AI governance platform and our admin console?
The console governs the product it belongs to: accounts, retention, audit logs, data handling. A platform sits above every AI system and holds the program — registry, control crosswalks, evidence trail. With one or two assistant products the console covers most of the ground. With twenty AI systems across four frameworks it does not.
How much does AI governance tooling cost?
Most vendors do not publish. Quotes are built from three inputs: AI systems governed, frameworks mapped, connected evidence sources. For certification, Vanta’s breakdown runs $3,000 to $10,000 for readiness, $10,000 to $40,000 for implementation, $5,000 to $20,000 for the initial audit. Distrust the unsourced six-figure ranges in comparison posts.
Does the EU AI Act apply to a US company of our size?
It can. The test is market and effect rather than headquarters, so deploying AI systems in the EU or affecting people there brings you in scope. Article 4’s literacy duty has applied since 2 February 2025; the Annex III deadline moved to 2 December 2027, while Article 50 transparency stayed at 2 August 2026. Confirm scope with counsel before buying on the strength of a deadline.
Do we need ISO 42001 certification?
Only if someone is asking for it. It is a voluntary management-system standard awarded by accredited bodies, and its commercial value is answering a customer, insurer, or procurement questionnaire. Adopting the structure without certifying delivers much of the internal benefit for far less.
What is shadow AI discovery and is it worth paying for?
It finds AI use you did not sanction — personal accounts, AI features inside SaaS bought for other reasons, data leaving through tools nobody inventoried. Worth paying for when your surface is wide: many SaaS products, contractors outside your directory, unmanaged devices. Otherwise your identity provider and expense data will tell you most of it.
Who should own AI governance when there is no compliance function?
Whoever owns the acceptable-use policy — commonly an operations leader, chief of staff, or finance — with IT supplying the inventory and account configuration. Handing it entirely to IT is the usual mistake: what may be entered about a client is a business judgment, not a setting.
Do we need agent governance yet?
Only if agents are taking actions in your systems — writing records, sending messages, moving money. Agent registries and runtime permission controls are the fastest-moving part of this category, which argues for waiting. If staff are drafting and analyzing with assistants, this is not your problem this quarter.
What should we buy first?
Nothing. Inventory your AI use across your identity provider, your expense data, and a no-consequence conversation with staff, then configure the controls in the subscriptions you hold. Both steps are free, remove the largest share of risk, and tell you which category, if any, you need.
Key takeaways
- The phrase covers three products: lifecycle platforms built around a model registry, compliance automation aimed at a certificate, and data-security discovery for assistant use. Only the third transfers cleanly to a company that bought licenses.
- Posture decides the shortlist. Providers who develop AI systems need a registry; deployers using licensed assistants have an empty one.
- Check your admin console first. Microsoft documents AI-activity discovery, sensitive-data reporting, and prompt-level audit logging inside Purview, tier-gated at A3/E3/G3 and A5/E5/G5 rather than requiring a new contract.
- The frameworks reach you unevenly: NIST AI RMF is voluntary and uncertifiable, ISO/IEC 42001 is worth pursuing when a customer asks, and the EU AI Act’s Annex III deadline moved to 2 December 2027 while Article 4 literacy duties applied from 2 February 2025.
- Distrust unsourced pricing. Certification costs break into readiness, implementation, audit, and surveillance components; the six-figure mid-market figures carry no source.
- Buy in order: inventory, configure what you own, write the policy, train the people the controls cannot cover, add discovery if your surface is wide, and consider a platform only once you have systems to register.
- Controls enforce at the boundary and log the attempt. They cannot judge whether a permitted output was any good, and never see the near-miss someone chose to report. That gap closes with practice on real work.
Dirk Jan van Veen, PhD
