Most AI business cases start with the technology and work backwards to the benefit. A vendor demonstrates a capability. An internal champion gets excited. Someone writes a business case that justifies the deployment after the decision to deploy has already been made.
The problem isn't the technology. It's the sequencing. When you build the case backwards, you're optimising for approval rather than for value. You pick the metrics that support the conclusion you already have. And when the deployment lands, there's no clear baseline to measure against — so you can't tell whether it worked.
Here's the framework we use before any AI agent deployment begins.
Why most AI business cases are built backwards
The honest reason is that forward-looking ROI analysis is harder to do, and the people doing it are often closer to the technology than to the workflows it would change. It's easier to describe what the AI does than to map exactly which human activities it eliminates, how many times per day those activities happen, and what they cost.
The result is business cases full of phrases like "estimated efficiency gains of 20–30%" — numbers that feel reasonable but aren't traceable to anything specific. When the deployment happens and someone asks "did it work?", there's nothing to measure against.
The fix is simple in principle: start with workflows, not capabilities. Identify the specific human activity being replaced or reduced. Count how often it happens. Price the time. Then calculate whether the deployment cost is justified by that number.
The four variables that determine ROI
For any AI agent deployment in a manufacturing operation, ROI depends on four variables:
- Query volume — how many times per day does the team perform the activity the agent would handle? This is the multiplier. Everything else scales with it.
- Time per query — how long does the manual activity currently take, from start to finish? Include navigation time, context-switching, and any follow-up steps. Most people underestimate this.
- Blended hourly rate — what does the time cost, at a blended rate across the people doing it? This doesn't need to be precise; a rough figure is enough to make the case directionally.
- Reduction factor — what fraction of the manual time does the agent eliminate? For well-defined retrieval tasks (order status, payment confirmation), this is typically 85–95%. For tasks with complex exception handling, it's lower — but still substantial.
The calculation is straightforward once you have these numbers. What most teams discover is that the monthly cost of the manual workflow is considerably higher than they expected — because the cost was distributed across hundreds of small actions that no one was tracking individually.
Run the numbers
Use the calculator below to map your own team's current cost and projected recovery. The defaults reflect the profile from our order status deployment — a team of 5 doing 20 queries per day at 3 minutes each.
Two real examples
Example 1: Order status queries. In our garment manufacturer deployment, a team of around 15 people across customer service, sales, and operations was making ERP queries 20–40 times per day. Each query took 2–3 minutes. At a blended rate of $25/hour, that's $13,750–$27,500 per month in navigation overhead — before you count the decision latency cost of answers arriving later than they should. The deployment cost was recovered in under 60 days.
Example 2: Invoice reconciliation. In our ongoing invoice reconciliation deployment, the manual process was consuming roughly 20 hours per week across two finance team members — matching invoices against POs, investigating discrepancies, and routing decisions. At $35/hour blended, that's $3,000+ per month just for the matching phase. The agent handles clean matches automatically; humans only touch exceptions. The time recovered compounds each month as invoice volume grows.
What the numbers don't capture
The time-and-cost calculation is the floor, not the ceiling. There are two value components that don't appear in an ROI calculator but compound over time.
The first is decision quality. When a sales rep can check order status mid-call instead of putting the customer on hold, the quality of that interaction is different. The answer arrives in context. The conversation doesn't interrupt. The customer's experience improves. This doesn't show up in a cost calculation, but it affects retention — and retention compounds.
The second is infrastructure reuse. The first AI agent deployment builds the infrastructure that makes the second one cheaper. Integration is done. Trust is established. The team knows how to use the interface. Each additional workflow adds a fraction of the original cost while delivering roughly the same category of value. This is why the most useful number to track isn't the ROI of any single deployment — it's the cumulative return on the infrastructure investment over 12–24 months. For the full architecture of how this infrastructure layer — decision governance, approval workflows, and audit trail — differs from the intelligence layer that surfaces signals, see Decision Infrastructure vs. Decision Intelligence.
AskOps: what a scoped deployment looks like
When we scope a deployment via AskOps, we start with exactly this framework. We map the highest-frequency ERP queries your team makes today, count them, time them, and price the overhead. Then we show you what the recovery looks like and what a production deployment costs.
The scoping call is 60 minutes. At the end of it, you have a number — not an estimate, but a calculation grounded in your actual query volume and your actual blended rate. Most teams find it considerably higher than they expected. That's usually when the conversation shifts from "is this worth doing?" to "when do we start?"
For Dynamics 365 Business Central teams, AskOps is bundled into OpsGrid — in live beta — our full decision infrastructure layer for BC. If your team is on Business Central, OpsGrid is the fastest path to a production deployment with a clear ROI baseline from day one.