Sample | AI Use Case Triage | Decision brief
Sample: AI Use Case Triage decision brief
What you keep after a 90-minute AI Use Case Triage: a two-page decision brief plus appendix.
An illustrative sample in the real format: Halden Building Supply does not exist, and every name and number is invented. Pages one and two are the decision; the appendix holds the working.
The bottom line (first pass)
Halden Building Supply (fictional company): a building-products wholesaler with 800 employees and 22 branches in Ontario and Quebec (from the intake). Eight AI ideas, one 90-minute session.
Put first for Discovery (recommendation): quote drafting; the Discovery Sprint confirms it. It outranks invoice matching because it touches revenue your team reports losing and a rep approves every quote, while invoice matching cannot post to your read-only finance system until you decide suggested matches suffice.
Do not pursue as proposed (recommendation): automated credit limits and résumé screening. Each hands an automated system a decision about a person.
The other five are CAVEAT: unresolved until each owner answers the question below.
Every fact here is sourced in the appendix: reported (and by whom), document reviewed, assumption, or not checked.
Next step: a two-week AI Discovery Sprint on quote drafting, fixed fee from $15,000 USD, $1,500 triage fee credited in full: a map of how quotes get made and an indicative business case. Decision call: 30 minutes with the VP Operations within ten business days.
Eight ideas, graded
| Case | Data it would touch | Reason | First-pass grade |
|---|---|---|---|
| Quote drafting for inside sales | Confidential: price books, account pricing. Emailed requests: potentially personal | Runs on data you hold, a rep approves every quote, and it targets lost accounts. | PROCEED: put first for Discovery |
| Product descriptions for about 14,000 web products | Public: product catalog | Public data with review before publishing, but supplier content reuse rights are unconfirmed. | CAVEAT |
| Supplier invoice matching (accounts payable exceptions) | Confidential financial data in the enterprise resource planning (ERP) system | A recurring backlog, but the ERP is read-only, so it could only suggest matches. | CAVEAT |
| Branch staff Q&A on returns and safety policy | Internal: policy documents | Some policy pages are outdated, so answers could repeat superseded returns policy. | CAVEAT |
| Sales call notes summarized into the CRM | Personal: names, phone numbers, site addresses | Writes notes about identifiable people, and recording-consent practice varies by branch. | CAVEAT |
| Website chat that answers order status | Confidential and personal: ERP account and order data | Customer-facing and reads account data; general FAQ answers without account data could start first. | CAVEAT |
| Automated credit limits for new trade accounts | Personal and financial: credit applications, owners' guarantees | Automates a financial decision about people, with no written basis, replacing a same-day process. | REFUSE |
| Résumé screening for branch hires | Personal: applicant records | Automated interview decisions carry fairness and legal exposure; managers already read applications. | REFUSE |
Listed in recommended order; the five CAVEATs are unranked. Grading key. PROCEED - Our call: this one first. Discovery confirms it. CAVEAT - Unresolved. Answer the stated question before deciding. REFUSE - Do not pursue as proposed. The reason is stated, and the appendix says what would change the answer. Every grade is a first-pass recommendation on the evidence in the intake and the session, and weighs both data exposure (who could see the data) and harmful outputs (what a wrong answer could do).
The questions to answer first
| Case | Question to answer first | Who owns the answer |
|---|---|---|
| Quote drafting | How often is custom or project pricing needed, and is account pricing in the price book or held by reps? | Director of Sales |
| Product descriptions | Do your supplier agreements let you rewrite manufacturer product content? | Marketing (owner to confirm) |
| Supplier invoice matching | With the ERP read-only, is suggesting matches for a clerk to post enough, or must the system post them? | Controller |
| Branch staff Q&A | Who owns the policy pages, and which version is current? | Operations (owner to confirm) |
| Sales call notes | Has the privacy owner confirmed this use of the recordings, beyond recording consent, and will a person review each note? | Privacy owner (not yet named; owner to confirm) |
| Website chat | Can the chat confirm a logged-in customer and answer only about that customer's account (account-level authorization)? | IT Manager |
| Automated credit limits | Is there a written limit-setting policy that decisions could be tested against? | Controller |
| Résumé screening | What is HR's view, and what are the written, job-related criteria for branch hires? | HR (owner to confirm) |
Each answer can move a grade and has one owner ("owner to confirm" if absent from the session). End of the decision brief; the appendix holds the working and evidence labels.
Appendix: the three cases in detail
For each case examined in the session: the workflow as reported, the material facts, what the system would read and do, the unknowns, and what would change the answer.
Quote drafting for inside sales: PROCEED
- •Reported workflow (Director of Sales): contractors email requests, often with attached material lists or drawings. A rep looks up each line in the catalog and the current price book, applies account pricing, and emails a quote back. Turnaround is about a day and a half (reported by the Director of Sales).
- •Material facts: quote speed was the stated reason for two lost accounts this year (reported by the Director of Sales; not checked). The catalog and price books sit in two systems Halden controls (reported by the IT Manager). In the ten requests reviewed, most lines matched catalog items and several needed a rep's judgment on substitutes or project pricing (document reviewed).
- •Data and action boundary: reads the request, its attachments, the catalog and the price book; writes a draft quote to a review queue; never sends. A rep checks and sends every quote.
- •Reason (recommendation): it runs on data you already hold, a person approves every output before a customer sees it, and it targets the delay your sales team reports losing accounts to. Our call is this one first; Discovery confirms it.
- •Unknowns: how often requests need custom or project pricing across all quotes (not checked); whether account pricing is recorded in the price book or held by reps (not checked); what personal information arrives in requests and attachments (not checked).
- •Next check: sit with inside-sales reps, collect a larger sample of real requests with the quotes sent, and map where the time goes. If most of it goes to judgment calls that are not written down anywhere, this case drops to CAVEAT.
Automated credit limits for new trade accounts: REFUSE
- •Reported workflow (Controller): applicants submit a credit application, usually with an owner's personal guarantee. An analyst reviews it and sets a limit, typically within a day. About 30 new accounts a month (reported by the Controller).
- •Material facts: the application collects owners' personal and financial details (document reviewed). There is no written record of how limits are set today (reported by the Controller).
- •Data and action boundary: would read credit applications and guarantees (personal and financial) and act: set a limit with no person deciding.
- •Reason (recommendation): do not pursue as proposed. The proposal gives an automated system a financial decision about identifiable people, with no documented basis to test its decisions against, to replace a process reported to work within a day.
- •Unknowns: how limits are set in practice; whether credit bureau data is used and on what terms; your obligations to explain a refusal (not checked; this brief is not legal advice).
- •What would change the answer: a written limit-setting policy that decisions can be tested against, a backlog the current team cannot clear, and a privacy and legal review. A safer version exists now: AI drafts the credit memo the analyst already writes, and the analyst decides. That version would be screened on its own, starting from CAVEAT.
Résumé screening for branch hires: REFUSE
- •Reported workflow (VP Operations): branch managers read applications in the applicant tracking system and choose whom to interview. About 60 branch hires a year (reported by the VP Operations).
- •Material facts: applications carry personal information, and shortlisting decides who gets an interview. HR was not in the session, so the current screening criteria were not seen (not checked).
- •Data and action boundary: would read applicant records (personal) and act: rank or exclude candidates.
- •Reason (recommendation): do not pursue as proposed. An automated call on who gets an interview carries fairness and legal exposure, and the managers who make the hire already read the applications today.
- •Unknowns: HR's view and any written criteria; employment and human-rights obligations in Ontario and Quebec (not checked; not legal advice).
- •What would change the answer: the HR owner in the room, written job-related criteria, a legal review, and a design where a person makes every call. Even then, the likely version assists (summarizing applications against stated criteria) rather than screens.
Appendix: your data, a provisional map
| Source | Classification (basis) | Cases | Boundary | Unresolved | Owner to confirm |
|---|---|---|---|---|---|
| Product catalog | Public (reported by the Director of Sales) | Quotes, descriptions | Our proposed precaution: a person reviews before publishing | Reuse rights for supplier content | Marketing |
| Price books and account pricing | Confidential (reported by the Director of Sales) | Quotes | Stated by you: stays inside your cloud tenant (your own account with your cloud provider) | Where account pricing is recorded | Sales operations |
| Emailed quote requests and attachments | Potentially personal, pending inspection (assumption) | Quotes | Our proposed precaution: processed inside your tenant until inspected | What personal information they carry | IT Manager |
| ERP: orders, invoices, delivery addresses | Confidential and personal (our classification) | Invoice matching, order-status chat | Stated by you: read-only, so invoice matching can suggest but not post. Our proposed precaution: account-level authorization before any order-status answer | Which fields each case needs | Controller |
| CRM contacts and call recordings | Personal (reported by the Director of Sales) | Call notes | Our proposed precaution: inside your tenant, with human review before any write to the CRM | Purpose confirmed by the privacy owner; consent practice by branch | Privacy owner (not yet named) |
| Policy documents (shared drive) | Internal (reported by the IT Manager) | Staff Q&A | Our proposed precaution: inside your tenant | Which version is current | Operations |
| Credit applications and guarantees | Personal and financial (document reviewed) | Credit limits | Not approved for this proposed automated use | Credit bureau terms, if used | Controller |
| Applicant tracking system | Personal (reported by the VP Operations) | Résumé screening | Not approved for this proposed automated use | Criteria and HR's view | HR |
Provisional: classifications are as reported in the session or proposed by us, not verified against your systems. "Stated by you" is a restriction your team set; "our proposed precaution" is a recommendation you can accept or change.
Appendix: scope and evidence
In the room: VP Operations (budget owner), Director of Sales, IT Manager, Controller. HR and Marketing were not present.
Intake received: a completed intake listing eight AI ideas, with company size and branch count, 48 hours before the session. Documents reviewed: the intake; ten recent quote requests and the quotes sent in reply (supplied by the Director of Sales); a price-book extract for one product category; the blank credit application form. Excluded: no system access, no interviews with front-line staff, no vendor, security-architecture, cost or legal review.
Facts behind the screened cases: the web SKU (stock-keeping unit) count and the catalog classification (reported by the Director of Sales); the recurring accounts payable backlog and the ERP's financial data (reported by the Controller); policy pages that predate last year's returns change (reported by the IT Manager; not checked); recording-consent practice that varies by branch (reported by the Director of Sales).
Evidence labels: reported (and by whom), from the intake, document reviewed, assumption, not checked. Recommendations (labeled recommendation, and every grade) are Magma's judgment on the facts supplied, not facts in themselves. This brief is a bounded synthesis of the evidence supplied, written within 48 hours of the session, not two days of investigation.
Left for the Discovery Sprint: how the work really runs day to day and what it is worth. Both need observation of the people doing it, which a 90-minute session cannot give. For quote drafting, the Sprint starts from the three unknowns in the appendix, with the inside-sales reps and a larger sample of real requests, and confirms or overturns our call.
Not in this brief: vendor choice, security architecture, running cost, legal advice. Those belong to later work, not to a 90-minute session.
Prepared by Magma Inc. A two-page decision brief plus appendix, graded PROCEED / CAVEAT / REFUSE as a first pass: we name what not to build as clearly as what to build.
Your list, graded
Bring your list of AI ideas.
In 90 minutes, find out which one to put first, which ones to stop, and what your data allows. $1,500 USD, a two-page decision brief plus appendix like this one within 48 hours, and the fee credits in full toward the Discovery Sprint.
See the AI Use Case Triage