By the Shadow AI Policy team
Most AI tool approval processes fail because they're built for risk prevention, not for getting an answer back to an employee in a reasonable amount of time — so employees stop asking and just use the tools anyway.
This post gives you a working template: a request form with the right fields, a tiered review system based on what data the tool touches, shortcuts for vendor security checks, hard deadlines for decisions, and a communication workflow so requesters don't feel like their request disappeared into a black hole.
The goal of an approval process isn't zero AI risk — it's knowing what tools are in use and why. A process that takes six weeks to return an answer guarantees employees will route around it. Build for speed first, then layer in scrutiny where the data sensitivity actually justifies it.
The typical AI tool approval request lands in IT's inbox, gets forwarded to Legal, sits in a queue, and resurfaces three weeks later with a list of follow-up questions. By then, the employee has either abandoned the idea or started using the tool on their personal device. Neither outcome is good.
The fix isn't more process — it's tiered process. A Canva AI request to help marketing design social graphics is not the same as a request to run ChatGPT on customer support transcripts containing personal data. Treating them identically is where approval workflows collapse under their own weight. If you want to understand the broader pattern of employees using unapproved tools, our overview of shadow AI covers why it happens and what's at stake.
The template below is designed for teams of 50–500 people. It assumes you have IT and at least one person who can ask basic security questions — but not a dedicated GRC team or an enterprise procurement department.
Keep the intake form short enough that a department manager can fill it out in under five minutes. Every field that isn't strictly necessary for routing the request to the right reviewer is a field that will be skipped or filled in badly. Here are the fields that actually matter:
That last field matters more than it looks. Discovering that someone already signed up for a paid plan and handed over company data before IT knew about it changes the urgency of the review. The data checklist is the routing mechanism for the tiered system — it's the single most important field on the form.
Deliver the form as a Google Form, a Typeform, a Jira ticket type, or a simple email alias with a template — whatever your team will actually use. The format is less important than the habit of using it consistently.
Once the form comes in, route it immediately based on what data the tool will touch. Three tiers covers almost every case:
| Tier | Data Types | Reviewers Required | Decision Deadline |
|---|---|---|---|
| Tier 1 — Low | No company data processed; tool operates on publicly available content or the employee's own drafts only | IT lead only | 3 business days |
| Tier 2 — Medium | Confidential business data, internal documents, non-PII customer data | IT lead + department head sign-off | 7 business days |
| Tier 3 — High | Customer PII, employee data, financial records, health information, or data subject to HIPAA, GDPR, CCPA, or FERPA | IT lead + Legal/compliance + department head | 14 business days |
Tier 3 tools require a vendor security assessment and, in regulated industries, a check on whether the vendor will sign a Business Associate Agreement (for HIPAA-covered data under 45 CFR § 164.308(b)) or a Data Processing Agreement (for GDPR personal data under Article 28). If the vendor won't sign either, the answer for regulated data is no — not "let's discuss." You can find a broader breakdown of what to look for in a policy framework in our AI acceptable use policy template guide.
Most tools that employees request fall into Tier 1 or Tier 2. Reserve your thoroughness for Tier 3 — that's where the actual regulatory and liability exposure lives.
A full vendor security questionnaire can run 200+ questions and take weeks. For a 100-person company evaluating whether to approve a writing assistant, that's overkill. Instead, run a focused 10-point check that you can complete in under two hours using publicly available information:
For common tools your team is already asking about, check our AI tool risk directory — it summarizes data handling, training opt-outs, and BAA/DPA availability for tools like Microsoft 365 Copilot and Google Gemini based on their published terms, so you don't have to read through vendor documentation from scratch.
A vendor that won't answer the 10 questions above — or whose answers require a sales call to get — is telling you something about how seriously they take security. That's a usable data point even before you see the SOC 2.
Deadlines only work if someone owns them. Assign a named IT or ops contact as the approval coordinator for every request. That person's job is to make sure the review doesn't stall — not to make the decision unilaterally, but to chase the right people when a deadline is approaching.
Set a rule: if a decision isn't made by the deadline, the coordinator escalates to the department head and sends a provisional response to the requester. "We haven't finished the review — here's what we still need and when you can expect an answer" is always better than silence. Silence is what creates shadow AI.
For Tier 1 and Tier 2 requests, if the only holdup is a reviewer who hasn't looked at it, consider whether the tool can be approved with conditions (e.g., "approved for use with non-confidential drafts; not approved for customer data") while the remaining questions are resolved. Partial approvals are underused — they let the business move forward while you finish due diligence on the parts that need it.
Keep a log of every decision — approved, denied, approved with conditions, and pending — in a shared location. A simple spreadsheet works. This becomes your approved tool registry, which is the foundation of any real AI governance program. Without it, you'll be re-reviewing the same tools every time someone new requests them.
Every decision needs a response to the person who made the request, using plain language. Four outcomes, four templates:
The denied response is the one most teams get wrong. Saying "no" without an alternative or an explanation pushes the employee toward using the tool anyway — just without telling anyone. Give them a reason and, where possible, a path forward. If the problem is that the tool doesn't have enterprise-grade data controls, tell them: "The free version of this tool trains on your inputs. The enterprise tier doesn't — here's what that costs, and here's how to get budget approval."
If your team doesn't yet have a policy that covers how approved tools should be used once they're cleared, generate a tailored policy kit that includes usage guidelines alongside the approval framework.
About Shadow AI Policy: We build AI acceptable use policy tools for HR and operations teams at 50–500 person companies. We publish guides on shadow AI, acceptable use policies, and AI governance, updated as regulations and AI tools change.
Treat it as a Tier 2 or Tier 3 review depending on what data has already been processed — and move faster, not slower. The priority shifts from "should we allow this" to "is there an active exposure we need to contain." If the tool has been processing sensitive data without a signed DPA or BAA, loop in Legal immediately to assess whether a breach notification obligation has been triggered under GDPR Article 33 or your applicable state privacy law. Don't punish the employee for using the process wrong if the process wasn't clearly communicated — fix the communication and complete the review.
Free tier status doesn't reduce the review requirement — it often increases it, because free tiers typically have weaker data protection terms than paid or enterprise plans. Check the vendor's terms of service specifically for clauses about whether user inputs are used for model training. Many free tiers include this by default, and the opt-out is either buried in settings or unavailable. If the tool is used team-wide, treat it as a Tier 2 or Tier 3 review regardless of cost, and check whether an enterprise plan with a DPA is available before approving broad rollout.
Yes, but it can be lighter. If a vendor you already have a contract with — say, your CRM or your productivity suite — adds an AI feature, your existing contract may or may not cover how that feature handles data. Check the vendor's updated terms and your existing DPA to see whether the AI feature is explicitly included. Microsoft 365 Copilot, for example, is governed by Microsoft's data protection addendum, but it must be licensed separately and has its own deployment prerequisites. Don't assume existing vendor relationships mean new AI features are automatically approved.
Start with a single-field intake: require employees to email a designated address with the tool name, their use case, and what data it will touch — nothing else. Have one person review that email and respond within five business days with approved, denied, or more questions needed. Maintain a running list of decisions in a shared doc. That's a functioning approval process. You can add tiers, forms, and security checklists once the habit of asking before using is established. Getting to 80% compliance with a simple process beats 0% compliance with a sophisticated one nobody uses.
Free tool: Check what an AI tool does with your data → Training on your content, retention, admin controls and BAA availability for 26 major tools.
Tailored to your industry and the AI tools your team uses. Free preview, then $149/mo to keep it current as the rules and vendor terms change — or $79 for a one-time snapshot.
Generate my policy kit →Writing policies for several clients? MSPs, IT consultancies and fractional CISOs keep a roster of client kits that refresh monthly, under their own branding. See partner plans →