Enterprise AI — AI Glossary

Reference details

Related terms

• Model Risk Management (MRM)
• AI Audit
• Shadow AI

How it works

1. Identity is delegated to the organisation's existing provider over SSO or SAML, so model access is governed by the same group membership that governs everything else.
2. Requests are tagged with the authenticated principal and written to an append-only audit log, which is what makes a later question about who asked what answerable.
3. Data residency is enforced by routing inference to region-pinned endpoints, and the contractual commitment that training does not occur on submitted data sits alongside it.
4. Role-based access control is applied to the retrieval layer rather than the model, so a user's query searches only the documents their existing permissions already allow.
5. Private deployment options — VPC endpoints, on-premise, or a dedicated tenancy — remove the shared-infrastructure question that security review otherwise cannot close.
6. The whole arrangement is priced and contracted per seat with an SLA, because what is being bought is the governance envelope rather than the inference itself.

Example: What the governance envelope actually costs

Take 5,000 seats. An enterprise platform at $30 per seat per month is 5,000 x $30 = $150,000 per month. Now price the same usage metered on a raw API. If each user makes 15 requests on each of 21 working days, that is 5,000 x 15 x 21 = 1,575,000 requests per month, and at 2,000 input and 500 output tokens per request the volume is 3,150,000,000 input tokens and 787,500,000 output tokens. At $3 per million input and $15 per million output that is 3,150 x $3 = $9,450 plus 787.5 x $15 = $11,812.50, so $21,262.50 per month. The two figures are $150,000 against $21,262.50, a factor of about 7.1, and the $128,737.50 difference is the honest price of the governance layer rather than a markup to negotiate away. Whether that is worth paying is a calculable question rather than a matter of taste. If the alternative is building SSO integration, audit logging, RBAC-aware retrieval, and a region-pinned deployment in-house, then $128,737.50 per month is $1,544,850 per year against the loaded cost of the engineers who would build and then maintain all four. For 5,000 seats that trade usually favours buying; for 50 seats the same arithmetic gives $1,500 against roughly $212.63 of metered usage, and the governance envelope is then costing more than seven times the inference while covering far less organisational risk.

Commonly confused with

Enterprise AI is not a capability tier, and the name misleads people into expecting a better model. The weights behind an enterprise offering are usually the same weights behind the consumer product; what differs is identity, logging, residency, contract, and support. The second confusion is with on-premise deployment, which is one option inside the category rather than a synonym for it — most enterprise deployments are hosted, distinguished by a VPC boundary and a data-processing agreement rather than by hardware location. It is also not the same as fine-tuning on proprietary data, which is a capability choice available at every tier; an organisation can fine-tune without any of the governance apparatus, and can have the full apparatus around an entirely off-the-shelf model.

When to use it

Reach for it when:

• When a compliance regime makes the audit log a requirement rather than a convenience, which is most of regulated finance and healthcare.
• Where data residency is a legal constraint, since region-pinned inference is difficult to retrofit onto a deployment that did not start with it.
• At seat counts where the per-seat premium is small against the engineering cost of building identity, logging, and permission-aware retrieval in-house.
• When procurement requires a vendor security review and an SLA, which metered API access generally does not provide.

Reach for something else when:

• For small teams, where the per-seat premium can exceed the metered cost of the same usage by several times over.
• Expecting better model quality, since the capability is usually identical and the difference is entirely in the envelope.
• As a substitute for permission-aware retrieval design, which the platform enables rather than solves.
• For prototyping, where the procurement cycle costs more calendar time than the experiment it would govern.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Enterprise AI — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path