Skip to main content

How do companies decide which AI models to make available to employees?

Summary

  • Companies that do this well run a managed "model menu": a short list of approved models, chosen on evidence and enforced in one place, then revisited as new models ship.
  • The core tests are quality on the company's own work, cost per task, legal and regulatory limits, provider terms, data handling, capacity, and which groups actually need each model.
  • Evaluations often say no: a newer model version can raise cost without a meaningful quality gain, and declining to offer it is a valid outcome.
  • Enforce the decision with access grants, not a policy memo: a standard tier for everyone, wider tiers for specific groups, plus rate limits, budgets and usage data to review choices.
  • On Databricks, Unity Gateway governs models as Unity Catalog securables, so approving or excluding a model is an EXECUTE grant or a tag-based GRANT policy you can audit.

How do companies decide which AI models to make available to employees?

Companies that do this well treat it as a managed "model menu": a short list of approved models, chosen on evidence and enforced in one place. Each candidate is judged on how it performs on the company's own work, what it costs per task, whether its terms and data handling fit legal and security policy, and whether there's capacity to serve it reliably. The decision is enforced through access grants and revisited as new models ship, because the model that was the best choice for a job a few months ago often has a better or cheaper option today.
Both extremes fail. Standardize on one provider and you risk missing better models or lower costs elsewhere. Support many with no control point and access, budget controls, security policies and usage logs scatter across separate systems. The workable middle is broad choice that's governed centrally.

What goes into a model approval decision?

Decision factor What to check
Quality on your work Results on an internal evaluation set built from real tasks. Public benchmarks do a poor job of predicting real-world performance on work like coding.
Cost per task The price of the quality you actually need. Much routine work, such as everyday coding, doesn't need the most expensive model.
Legal and regulatory limits Export-controlled model families, vendor- or region-restricted models, and corporate policies that prohibit specific models.
Provider terms The applicable terms for each model. Databricks' release notes say customers are responsible for ensuring their compliance with applicable terms.
Data and security Where requests go, what gets logged, and which guardrails apply to inputs and outputs.
Capacity and availability Rate limits, available capacity and fallback options for the workloads that will depend on the model.
Audience A standard set for everyone, and wider sets for power users or an AI platform team.

Evaluations often return negative results. In Databricks' post on managing AI coding costs, one company found that a newer model version didn't meaningfully improve quality while increasing cost, so it declined to make that version available internally. Saying no is part of the process.

How Databricks approaches model access

On Databricks, the model menu lives in Unity Gateway, which is built on Unity Catalog. Model services are Unity Catalog securables, so an approval decision becomes a set of grants you can audit.

  • Start from the default and narrow it. By default, all account users have EXECUTE on system.ai and can query its system-provided model services. To restrict that, remove the account users group's default schema-level EXECUTE grant and grant EXECUTE on approved model services only, as the model service governance docs describe. To keep all current and future system.ai services to an AI platform team, manage privileges on the schema itself.
  • Exclude by attribute. Models and model services in system.ai carry system tags such as ai.model_family, so GRANT policies can grant access to everything except a model family you've decided against. The GRANT policy reference lists models and model services as generally available securable types.
  • Offer tiers. Create a separate model service for each audience, such as a standard service for all account users and a power-user service for a smaller group, each with its own rate limits.
  • Bring approved external providers. Register a provider account once as a model provider service. Databricks encrypts and stores the credentials, and callers never see them.
  • Cap and watch spend. Budgets support shared and per-user thresholds, per-user overrides for groups that need more, and alert or block actions. Enforcement uses near real-time estimates, so treat budgets as cost control, not an absolute cap on billed amounts.
  • Review with data. The system.ai_gateway.usage table and the built-in usage dashboard show which models teams use, with token counts, latency and errors. Lineage tracks which consumers depend on a model service, so you can scope the impact before changing or retiring one. See Unity Gateway observability.

FAQs

Should every employee get access to every model?

Rarely. A standard set for everyone plus wider sets for specific groups gives choice without losing control of cost, data handling and compliance. Separate model services with their own grants and rate limits make that tiering explicit.

How often should the approved list change?

Whenever the evidence changes. New models arrive often, and a model that was the best option a few months ago may have a cheaper or better successor. Review usage, cost and evaluation results on a regular cadence.

Can we block one model family but keep the rest?

Yes. In Unity Gateway, a GRANT policy can match the ai.model_family system tag and grant access only to models and model services outside the excluded family.

Is Unity Gateway generally available?

Yes. The Unity Gateway release notes announced general availability on August 4, 2026. Some capabilities, including service policies, remain in Beta.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.