Every AI vendor promises ROI. Few of them show you the math.
That's a problem — because if you can't run the numbers yourself, you're flying blind on a six-figure decision. Here's how we actually calculate ROI for a mid-market manufacturer, using a scenario that's close to real engagements we've run.
The Scenario: Precision Parts Co.
Precision Parts Co. is a 200-person job shop in eastern Nebraska. They run Epicor Kinetic, have about $42M in annual revenue, and a 14-person ops team that spends a significant chunk of each day answering the same question in different forms: Where's my order, and why is it late?
Before we touch any code, we do an AI Discovery Sprint — a structured process audit. Here's what we found:
- Order status inquiries: Planners fielded ~35 calls/emails per day. Average handle time: 8 minutes each.
- Manual ERP report pulls: 3 ops staff spent roughly 45 minutes per day pulling, formatting, and distributing status reports that could be automated.
- Expedite decisions: Supervisors made 10–15 reactive schedule changes per week based on stale data — often creating downstream problems that took another 2–3 hours to unwind.
None of this was visible in any dashboard. It only surfaced during structured interviews.
Running the Numbers
| Cost Bucket | Calculation | Annual Cost |
|---|---|---|
| Order status inquiries | 35/day × 8 min × $38/hr × 250 days | $65,000 |
| Manual reporting | 3 staff × 45 min/day × $38/hr × 250 days | $21,400 |
| Reactive expediting | 12 changes/wk × 2.5 hrs × $52/hr × 50 wks | $78,000 |
| Total recoverable cost | ~$164,000/yr |
We built an AI copilot layer on top of their existing Epicor instance — no data migration, no ERP replacement. The copilot handles natural-language order status queries, auto-generates and distributes the daily ops report, and surfaces early schedule conflict alerts so supervisors can act before the expedite fire starts.
What It Actually Cost
| Item | Cost |
|---|---|
| Implementation (one-time) | $48,000 |
| Annual support & model tuning | $9,600/year |
| Year 1 net savings | ~$106,000 |
| Payback period | Under 6 months |
Those numbers are conservative. They don't include margin improvement from fewer late shipments, or the capacity freed up when planners shift from firefighting to forward-looking work.
Why Most ROI Calculations Are Wrong
Vendors will show you a slide with a big number — "15% productivity improvement!" — with no basis in your specific workflow. That number is meaningless without knowing:
- What processes are actually costing you time — which requires structured discovery, not a demo
- Whether your ERP data is clean enough to power the copilot — dirty data kills AI projects faster than anything
- What the realistic adoption curve looks like — week 1 is not week 12
A real ROI calculation starts with your ops reality, not a vendor's benchmark slide.
Where to Start
If you're a midwestern business in the 10–300 employee range, the first step isn't buying software. It's understanding exactly where your team is bleeding time — and whether AI can actually stop it. That's what the AI Discovery Sprint is for.