How to Prioritize Your First AI Use Case
A scoring framework for AI use case prioritization at mid-market manufacturers — rank candidates by value, feasibility, and data so your first agent ships.
Get AI use case prioritization wrong and you'll spend nine months and a six-figure budget on a project that demos well and dies in committee. Get it right and your first agent pays for itself before the next budget cycle. I've watched both happen at a $250M manufacturer. The difference wasn't the technology or the vendor. It was a disciplined way to rank the candidates and the discipline to say no to the shiny ones. This is the framework I'd hand a VP of Ops or Head of IT who's about to pick their first AI use case.
The trap is that everyone has a favorite idea. The plant manager wants predictive maintenance. The CFO wants AP automation. Sales wants a quoting bot. Without a scoring method, the loudest voice wins, and the loudest voice is usually wrong about what's actually buildable this year.
The four-factor score
Rank every candidate on four factors, 1-5 each. Multiply, don't average — a zero on any factor should sink the idea, and multiplication punishes weakness harder than an average does.
| Factor | What it measures | The killer question |
|---|---|---|
| Value | Annual dollars at stake | How much labor cost or margin leak does this process represent? |
| Feasibility | How buildable with today's tools | Can an agent actually do this, or does it need judgment a model can't fake? |
| Data readiness | Is the input clean and accessible | Is the data in a system we can read, or trapped in someone's head? |
| Owner pull | Does ops want it | Will a named person run this Monday morning, or is IT pushing it? |
The score = Value × Feasibility × Data × Owner Pull. A candidate scoring 5×5×1×4 = 100 loses to one scoring 4×4×4×3 = 192, because the first one has no usable data. That's the whole point. The sexiest idea with no data feed is worth nothing.
Score Value in dollars, not adjectives
Value is annual dollars, full stop. Estimate it the boring way: volume × time-per-unit × loaded labor rate × the fraction an agent can absorb.
Worked example — order entry:
- 1,800 orders/month × 4.2 minutes each = 126 hours/month
- At $45/hour loaded = $5,670/month = $68,000/year in touch time
- An agent absorbs ~65% of it = ~$44,000/year in recoverable cost
That's a 4 on Value. It's not glamorous. It's also a real number you can defend to finance, which beats "transformational impact" every time.
Feasibility: what AI is actually good at right now
Be honest about what today's agents do well versus where they fall over. Score Feasibility high for tasks that are:
- Text-in, text-out — reading an email, drafting a response, extracting fields from a PDF
- High-volume and repetitive — the same decision made hundreds of times a day
- Tolerant of a human check — the agent proposes, a person approves the edge cases
Score it low for tasks that need:
- Physical-world sensing — anything tied to a PLC, sensor, or vision system (real, but a capital project, not a first use case)
- Rare, high-stakes judgment — a $2M contract negotiation, a safety call
- Reasoning over data that doesn't exist in any system
First use cases should be boring office work: order entry, invoice matching, quote prep, customer email triage, warranty claim first-pass. That's where feasibility is highest and integration risk is lowest.
Data readiness is the silent killer
More first projects die on data than on anything else. Before scoring, answer three questions:
- Where does the input live? ERP, MES, email, a shared drive, or a person's memory? The further right on that list, the lower the score.
- Can we read it programmatically? An API or database connection scores 5. "We export a spreadsheet weekly" scores 2. "It's in PDFs in someone's inbox" scores 1 — though modern document extraction can rescue this one.
- Is it clean enough to trust? If 30% of your part numbers are typo'd free text, the agent inherits the mess.
If data readiness is a 1, don't kill the idea — but fix the data first and re-score. The data work often delivers value on its own.
Owner pull: the factor everyone skips
The best-scoring use case on paper still fails if no one in operations wants it. Owner pull asks: is there a named manager who will report this agent's number in the monthly review and fight for it when it stumbles? IT-pushed projects with no ops sponsor have a brutal failure rate. A genuinely enthusiastic owner is worth a point or two of feasibility you can engineer around.
A worked ranking
Here's how a real first-pass shortlist might score:
| Use case | Value | Feasibility | Data | Owner | Score |
|---|---|---|---|---|---|
| Order-entry agent | 4 | 5 | 4 | 4 | 320 |
| AP invoice matching | 5 | 4 | 4 | 3 | 240 |
| Quote turnaround agent | 4 | 4 | 3 | 5 | 240 |
| Predictive maintenance | 5 | 2 | 2 | 4 | 80 |
| Warranty claim triage | 3 | 4 | 3 | 2 | 72 |
Predictive maintenance has the highest value and the lowest score. That's not a bug. It's the framework keeping you out of an 18-month capital project when you should be shipping an order-entry agent in 90 days.
The one-page rule
If you can't fit your top candidate's case on one page — the dollar baseline, the data source, the named owner, the target metric, and what happens to the cases the agent can't handle — you don't understand it well enough to build it. Force the one-pager. It exposes the weak ideas before they cost you a quarter.
Your next step
AI use case prioritization isn't a workshop exercise. It's a ranking, a number, and a single ship decision. If you want a head start, our free First 5 Agents teardown scores the five highest-ROI agents for a manufacturer your size against exactly these four factors, with dollar estimates already filled in. Grab it, then book a 30-minute call and we'll run your candidate list through the framework live and tell you which one to ship first.
Let's see what's worth building first.
A 15-minute call: tell me where your AI or planning is stuck, and I'll tell you the one thing worth building first — and whether it's worth doing at all.