An AI intake process has one job that most of them forget: it has to be faster and easier than not using it. Every request that goes through the front door is a request that does not go round the side, and an inventory entry you did not have to discover.
The BlackFog and Sapio survey of 2,000 UK and US employees found 63% considered it acceptable to use AI tools without IT oversight when no approved option exists. That is not a compliance failure to punish. It is a service level to beat.
This playbook builds an intake that triages by risk, decides most requests in days, and spends review effort only where data sensitivity or autonomy justify it.
Preconditions
- A named executive owner for AI risk. Someone who can approve the tiering rules and break ties. Without one, every hard case escalates to nobody.
- A short AI use policy stating prohibited uses and the data classes that cannot go to an external model without review. One page. The policy is covered separately on this desk.
- Your existing data classification. The intake reuses it. Do not invent an AI-specific one.
- A list of already-approved tools. Even if it is short. The intake's fastest path is "use the approved one".
Phase table
| Phase | Duration | Output |
|---|---|---|
| Design the form and tiers | Week 1 | One form, three tiers, written criteria |
| Seed the inventory | Week 2 | Every known AI tool and feature entered as a record |
| Pilot | Weeks 3 to 4 | Two teams use it; measure time to decision |
| Launch | Week 5 | Announced with the approved-tools list |
| Review | Quarterly | Tier criteria adjusted from real decisions |
The form
Ten fields. Every additional field costs you requests.
| Field | Why it is there |
|---|---|
| What you want to do, in one sentence | Purpose drives everything else |
| Tool or model, and vendor | Needed for the inventory and vendor review |
| Is it a new tool, or a new AI feature in a tool we already use? | Embedded features are the most missed category |
| Highest data class it will touch | Main tiering input |
| Does output go to customers or the public? | Second tiering input |
| Can it take actions, or only produce text? | Third tiering input: autonomy |
| Will it make or shape decisions about people? | Employment, credit, access: the high-risk flag |
| Who owns it day to day | The inventory needs a name |
| Number of users | Scale changes exposure |
| When you need it | So the queue can be honest about dates |
These fields line up with what the NIST Generative AI Profile suggests an inventory entry carry: underlying models and versions, access modes, data provenance, and human oversight roles (GV-1.6-003). The intake form is the inventory's data entry screen. Do not run them as separate processes.
The three tiers
Tier 1: Approve on submission
Public or internal-only data, output reviewed by a person before use, no actions, no decisions about people, and an approved tool or an enterprise-licensed version of one. Decision: automatic, logged, inside one business day. Most requests should land here. If they do not, the approved-tools list is too short.
Tier 2: Light review
Confidential data, or customer-facing output, or a new vendor. Review by security and privacy against a fixed checklist: data retention and training terms, where data is processed, access controls, logging, and an exit path. Decision inside five business days. The reviewer can approve, approve with conditions, or escalate.
Tier 3: Full assessment
Regulated or restricted data, actions taken without a human in the loop, decisions about people, or anything that might fall into a high-risk category under the EU AI Act. These go through a written use-case assessment with a named approver above the requester's line. Covered in assessing a generative AI use case on this desk.
Write the tier criteria as yes/no questions on the form, so the tier is computed, not argued.
Who approves what
| Tier | Approver | Record |
|---|---|---|
| 1 | Automatic, against the criteria | Inventory entry created |
| 2 | Security and privacy reviewer | Checklist plus conditions |
| 3 | Business owner plus the named AI risk executive | Written assessment and decision |
The business owner signs tier 3 because the business owns the benefit and the risk. Security advises and can escalate; it should not be the only signature, or every refusal becomes a security veto and every approval becomes security's liability.
Embedded AI in tools you already own
The category most intakes miss. A collaboration suite, CRM, or support platform ships an AI feature in a release note, and it is on by default. Nobody filed a request because nobody bought anything new.
Treat a vendor enabling AI on data you already send them as a new intake request, raised by the tool's internal owner. Ask your top twenty SaaS vendors directly whether AI features are on by default and whether customer data trains anything. File the answers in the inventory.
Connecting tools to systems
The BlackFog survey found 51% of respondents had connected AI tools to other work systems or apps without IT approval. Connection is where a text tool becomes an actor: a chat assistant with access to a mailbox or a ticketing system can read and send.
Make "can it take actions" and "what does it connect to" mandatory form fields, and route any request that grants an AI tool write access to another system to tier 2 at least. Agent identity questions, such as who owns the credential and how it is revoked, belong to the identity program; see the orphaned agents piece on the identity desk.
Measuring the intake
Track four numbers monthly:
- Median time to decision per tier
- Share of requests by tier
- Requests withdrawn or abandoned
- Tools found by discovery that never came through intake
The last one is the honest score. If discovery keeps finding tools the intake never saw, the intake is too slow, too hard, or unknown.
Failure modes
Everything is tier 3. Usually because the criteria were written by people who only see the hard cases. Pilot with real requests and look at where they land.
The reviewer becomes a bottleneck. One person reviewing every tier 2 request is a queue. Train a second reviewer before launch.
Approval without conditions tracking. "Approved, provided logging is enabled" is worthless if nobody checks. Conditions get an owner and a date in the inventory record.
No expiry. Models and vendor terms change. Tier 2 and 3 approvals should be reviewed at least yearly, and on any material change to the model, vendor terms, or data scope.
How this maps to the frameworks
Running this intake produces evidence that several public references expect, without adopting any of them wholesale.
- NIST AI RMF is voluntary and organized around Govern, Map, Measure, and Manage. The Generative AI Profile, NIST AI 600-1, asks organizations to enumerate generative AI systems for the inventory (GV-1.6-001). The intake form is that enumeration.
- ISO/IEC 42001:2023 specifies requirements for an AI management system for any organization providing or using AI. A documented intake with named approvers is the kind of operational process a management system audits.
- The EU AI Act obligations differ by risk classification and role. Tier 3's "decisions about people" flag is where an Annex III high-risk use would surface; those obligations apply from 2 December 2027 under the Omnibus amendment. Classification is a legal question, so route a tier 3 flag to counsel as well as to the assessment.
For a full walkthrough of the AI RMF functions and a minimum set of artefacts, see NIST AI RMF in practice.
Skip this if
Your organization has fewer than about fifty people and uses AI only through two or three enterprise-licensed tools. A one-page policy and a shared list of approved tools does the same job. Build the intake when requests start arriving faster than one person can answer them in a chat thread.
What to cut if you only have two weeks
Ship the form, tier 1 auto-approval, and the approved-tools list. Route everything else to one reviewer with a five-day target. You will learn the tier 2 and tier 3 criteria from the first month of real requests, which is better than guessing them.
Next steps
- Name the AI risk executive and get the tiering criteria approved.
- Publish the approved-tools list before the form, so tier 1 has somewhere to point.
- Seed the inventory from what you already know, then compare it with discovery data.
- Pair this with a shadow AI policy people follow.
- For tier 3 requests, start from the generative AI risk assessment template. AI governance platforms are compared in GRC Compass.