An AI use case template captures a proposed AI idea in a standard shape: who it serves, what the AI does, what measurable result it should change, which data and systems it needs, how much autonomy it gets, and what risks it raises. The template below gives a ten-section intake form, a one-page canvas, a prioritisation scorecard and a worked example.
What intake decides
Most organisations have more AI ideas than capacity. Intake is the step where an idea becomes either a funded piece of work with an owner, or a decision not to proceed that people can see the reasons for. It has to answer four questions quickly: is the problem real and measured, is the data available, how risky is it, and how much autonomy does it need. Everything else, including model choice, comes later.
Keep the form short enough that business teams fill it in themselves. The AI governance lead reads it, assigns a provisional risk tier under the organisation’s AI governance policy, and routes it. NIST AI RMF MAP 1.1 asks for the intended purpose, context and users to be understood and documented, and MAP 3.1 and 3.2 ask for expected benefits and costs to be examined; an intake form is where that starts. The OECD Framework for the Classification of AI Systems, published in February 2022, sorts the same questions into five dimensions: people and planet, economic context, data and input, AI model, and task and output.
The AI use case intake form template
Copy the sections below into your form tool or a document. Brackets mark the fields to complete.
A. Requester and sponsor
A. REQUESTER AND SPONSOR
Requested by: [name, team, date]
Business sponsor: [name, role; commits budget and people]
Process owner: [name; owns the workflow today]
Working title: [short name]B. Use case statement
B. USE CASE STATEMENT (one sentence)
For [user or role], use [AI capability] to [task] so that
[KPI] moves from [baseline] to [target] within [timeline],
while respecting [constraints].
Example: For claims handlers, use document extraction to
pre-fill first-notice-of-loss forms so that average
handling time moves from 22 to 12 minutes within two
quarters, while keeping all claims data on our servers.The sentence follows the same parts as the Use Case Sentence Formula in our AI use case framework tool, which scores a single idea in a few minutes.
C. Current process and baseline
C. CURRENT PROCESS AND BASELINE
How the work is done today: [steps, roles, systems]
Volume: [cases per day / month]
Baseline measure and source: [value, from where, period]
Pain points: [delay, errors, cost, backlog]D. Data and systems
D. DATA AND SYSTEMS
Data the use case needs: [documents, records, fields]
Where it lives: [systems, owners]
Classification: [Public / Internal /
Confidential / Restricted]
Personal data: [no / yes: categories]
Special-category data: [no / yes]
Access today: [who may read it]
Systems it must read from: [ ]
Systems it must write to: [ ]
Data quality known issues: [ ]E. Autonomy and human role
E. AUTONOMY AND HUMAN ROLE
[ ] Assist system drafts, a person does the work
[ ] Augment system recommends, a person decides
[ ] Automate system acts within approved limits
[ ] Autonomize agents coordinate multi-step work
Who reviews output, and at which step: [ ]
Actions that must wait for approval: [ ]F. Risk screen
F. RISK SCREEN (answer yes / no / unsure)
F1 Could it involve a practice prohibited by EU AI Act
Article 5 (e.g. social scoring, emotion recognition at
work, manipulative techniques)?
F2 Does it fall in an Annex III area: biometrics, critical
infrastructure, education, employment, essential
services such as credit or life and health insurance,
law enforcement, migration, justice or elections?
F3 Is it a safety component of a regulated product?
F4 Will people interact with it, or will it generate
content others see (Article 50)?
F5 Does it make or shape decisions about individuals?
F6 Will it use personal or special-category data?
F7 Will it act in other systems (send, pay, change)?
F8 Is any answer "unsure"?
Any yes in F1: stop. Any yes in F2, F3, F5 or F7, or any
unsure: route to a full risk assessment.G. Options
G. OPTIONS
[ ] Configure an approved platform or existing tool
[ ] Buy a product: [vendor, product]
[ ] Build: [team]
Why this option: [ ]H. Effort estimate
H. EFFORT ESTIMATE (rough)
Integration work: [systems, connectors needed]
Data preparation: [cleaning, labelling, permissions]
People: [roles, weeks]
Running cost: [compute, licences, per month]I. Value estimate
I. VALUE ESTIMATE
Hours saved or revenue protected per month: [ ]
Quality or compliance benefit: [ ]
How it will be measured after launch: [ ]J. Intake decision
J. INTAKE DECISION (completed by the AI governance lead)
[ ] Proceed to discovery [ ] Full risk assessment first
[ ] Park until [date / condition] [ ] Reject
Provisional risk tier: [0 / 1 / 2 / 3]
Reasons and conditions: [ ]
Decided by / date: [ ]
Inventory ID assigned: [ ]AI use case canvas
The canvas is the same information on one page, for steering groups that compare several ideas at once. Fill it from the intake form rather than separately, so the two never drift apart.
| Problem and users | AI capability and autonomy | Outcome and KPI |
|---|---|---|
| Who struggles with what, how often (C) | What the AI does; Assist, Augment, Automate or Autonomize (B, E) | Baseline, target, timeline (B, I) |
| Data and systems | Risks and controls | Effort and cost |
| Sources, classification, owners (D) | Risk screen result, tier, required approvals (F, J) | Integration, data work, people, running cost (H) |
| Human role | Option | Decision and next step |
| Reviewer, approval points (E) | Configure, buy or build (G) | Discovery, assessment, park or reject (J) |
AI use case prioritisation template
Score each sub-criterion from 1 to 5 and average within each dimension. Averaging keeps one enthusiastic rating from carrying an idea. The sub-criteria below match those in the framework tool, so a portfolio scored here and single ideas scored there can be compared.
| Dimension | Sub-criteria (1 = low, 5 = high) |
|---|---|
| Business value | Revenue or margin lift; direct cost and labour savings; process efficiency and cycle time |
| Feasibility | Data readiness and availability; ease of integration with systems; team skills and support |
| Risk exposure | Compliance and regulatory exposure; security and data sensitivity; change and adoption resistance |
Plot value against feasibility. High on both: do first. High value but hard: break into stages. Easy but low value: do only if it unblocks a team. Hard and low value: decline. Then read the risk score as a list of controls the idea will need before go-live, which is the argument our guide to ranking use cases by value, risk and effort makes in detail. The exception is the risk screen: a prohibited practice is rejected whatever the score, and a likely high-risk system goes through a full AI risk assessment before anyone builds it.
How to fill it in, and who owns it
- Requester completes sections A to I, with the process owner’s help on C and D. Thirty minutes is a fair target.
- AI governance lead checks the risk screen, completes J, assigns a provisional tier and an inventory ID, and replies within [ten] working days.
- Data owner confirms section D before discovery starts, because data access is where most first projects stall.
- Security and privacy see every form with a yes in F6 or F7.
- The AI governance committee sees every idea routed to a full assessment, and the ranked portfolio each quarter.
For a quick legal check on F1 to F4, the EU AI Act compliance checker asks up to twelve questions and returns the likely class, and the AI risk classification agent records the reasoning once the idea moves forward. The editable DOCX of the intake form and canvas, plus the governance policy, risk register, incident plan, model card and readiness checklist, is in the AI Governance Template Pack.
Worked example: four ideas at a manufacturer
The company and scores are fictional. Value and feasibility are averages of the three sub-criteria above.
| Use case | Autonomy | Value | Feasibility | Risk | Screen result | Decision |
|---|---|---|---|---|---|---|
| Technicians ask questions of maintenance manuals | Assist | 3.7 | 4.3 | 1.7 | No flags | Proceed; Tier 3 |
| Agent matches supplier invoices to orders and proposes postings | Automate, with approval above a set amount | 4.3 | 3.3 | 3.0 | F7 yes | Full assessment; Tier 2; approval gate on postings |
| Ranking job applicants for production roles | Augment | 3.3 | 3.0 | 4.7 | F2 and F5 yes: employment is an Annex III area | Full assessment; Tier 1; committee decision |
| Website chatbot that quotes prices to customers | Automate | 3.0 | 2.3 | 3.7 | F4 yes; prices are binding offers | Park until the price data is clean |
The manuals assistant goes first: high feasibility, low risk, and a measurable baseline in technician search time. Its completed form fits in a few lines:
INTAKE FORM (excerpt, fictional): maintenance manual assistant
B For maintenance technicians, use question answering over
equipment manuals so that the average time to find a
procedure moves from 14 to 5 minutes within one quarter,
while the manuals stay on our own network.
C About 300 look-ups a day across three plants; times taken
from a two-week shadowing study in August 2026.
D 12,400 PDF manuals in the document management system;
Internal; no personal data; every technician may read them.
E Assist. Technicians check the cited page before acting.
F All no.
J Proceed to discovery. Tier 3. Condition: every answer
cites the manual and page. Inventory ID AI-0042.The invoice agent is worth more but needs an approval step on postings and a test set of past invoices. The applicant ranking idea is not rejected, but it now carries the deployer duties of the EU AI Act, which apply to Annex III systems from 2 December 2027 under the Digital Omnibus on AI, Regulation (EU) 2026/1744.
Before running intake at scale, the AI readiness assessment shows whether data, infrastructure or governance gaps will hold the first ideas back.
How VDF AI fits
Intake goes faster when the default answer to “where would this run?” is already settled. VDF AI installs on your own infrastructure, on-premises, in a private cloud or air-gapped, so ideas that touch Confidential data, like the invoice agent above, can be built without a new external data flow to approve.
The autonomy levels in section E map onto platform settings: assistants that only draft, agents whose tools are granted per role, and approval gates that hold actions such as postings until a person confirms them. Every execution is logged with the user, model version, sources and tool calls, which gives section I its post-launch measure and the audit trail a Tier 1 or Tier 2 use case needs. Choosing and ranking the use cases stays with your teams.