Ship AI without the risk

Know what leaves your walls before you turn a tool on.

• Confirm whether a vendor trains on your data by default, and turn that off where it's optional — several vendors in this registry (e.g. Business/Enterprise tiers) default to training opt-out; free and Pro tiers often don't.
• Route customer, financial, and legal data only through tools whose contract explicitly covers that data class — a generic "AI assistant" plan is rarely enough on its own.
• Prefer vendors offering zero data retention or configurable retention windows for regulated data.

Give people the tool, not the keys to everything it can touch.

• Provision AI tool access through the same SSO/SCIM your org already uses — most enterprise-ready tools in this registry support SAML SSO on their top tier.
• Scope agentic tools (coding agents, support agents, contract-drafting agents) to the specific systems they need, not broad standing credentials.
• Review and revoke access on offboarding with the same urgency as any other SaaS credential — AI tools are often missed in offboarding checklists.

Some decisions stay human, permanently — not just during a pilot.

• Anything customer-facing (a support reply, a marketing send, a sales email) gets human review before it goes out, until measured accuracy earns a narrower exception.
• Anything financial (an invoice approval, a forecast submitted upward, a contract term) requires named human sign-off — AI can draft it, but a person owns the decision.
• Keep an audit trail of what was AI-generated vs. human-authored, especially for regulated functions like Finance and Legal.

AI tool spend sprawls fast — treat it like the recurring cost it is.

• Track spend per seat, not just per tool — unused seats on usage-based AI tools (billed per resolution, per credit, per conversation) are easy to miss.
• Re-check pricing on a schedule — this registry flags any tool whose pricing was last verified more than 120 days ago; do the same for your own procurement record.
• Consolidate before renewing: the Stack Builder's Optimize stage exists specifically to catch overlapping tools before an auto-renewal locks them in.

The failure modes are different from normal software — plan for them specifically.

• Prompt injection and data exfiltration: a support or coding agent that reads untrusted content (a web page, an email, a ticket) can be steered into leaking secrets or taking unintended actions — treat any agent with browsing or file access as operating in an adversarial environment.
• Shadow AI: employees adopting consumer AI tools outside procurement is the most common real-world data-leak path — a sanctioned, easy-to-use alternative in this registry beats a policy nobody reads.
• Supply-chain risk: a vendor's underlying model, fine-tune, or retrieval source can change without your sign-off — pin versions where the vendor allows it, and re-test after any vendor-announced model upgrade.

Evaluation (evals) is how you get evidence a tool works — don't take a vendor's benchmark as your own.

• Build a small, representative evaluation set from your own real tasks and data before rollout — vendor benchmarks measure the vendor's chosen tasks, not yours.
• Re-run the evaluation set after every model or vendor version upgrade, not just at initial rollout — a silent model swap is the most common cause of a tool's accuracy quietly changing underneath you.
• For regulated functions (Legal, Finance, HR), keep a human-graded evaluation alongside any automated one — automated grading alone under-catches the failure modes that matter most in those functions.

Guardrails are what stop a bad output from becoming a bad outcome.

• Filter both directions: redact PII and secrets before they reach a third-party model, and screen outputs for policy violations before they reach a customer or a downstream system.
• Scope agentic tools to an explicit allow-list of actions and systems — design against an agent doing something correctly that it should never have been allowed to attempt in the first place.
• Set a hard spend and rate cap per tool per period as a technical control, not just a budget line — usage-based AI tools can run away on a bug or a bad prompt loop in a way flat-fee software can't.

Governance doesn't end at rollout — it's what tells you whether it's still working.

• Log prompts, outputs, and the human decision that followed (approved, edited, or rejected) for every AI-assisted action in a regulated function — this is both your audit trail and your next eval set.
• Watch for drift: track accuracy and approval rate per tool over time, not just at launch, so a quiet vendor-side model change shows up as a metric moving, not a customer complaint.
• Alert on anomalies — a spend spike, an unusual access pattern, or an agent taking actions outside its normal scope — the same way you'd alert on any other production system.