This post is for senior operators at $50M-$500M companies who've tried AI and found that pilots stall, ROI evaporates, or vendors over-promise. We make the case that the model is rarely the bottleneck. The bottleneck is workflow design, governance, and operating-model fit.
We'll show how Joust runs the Operating Review to find the real money inside operations, and the patterns we use to ship work that holds up after we leave.
Why most AI pilots stall
The model isn't the bottleneck.
Companies that have tried AI usually share a pattern: a high-energy pilot, real engineering effort, and then nothing in production six months later. The post-mortems blame data quality, model accuracy, or change management. In our experience the cause is upstream of all three. AI work fails when it's designed around the model instead of around the workflow it's supposed to fit into.
Where pilots actually break
Across the AI engagements we've reviewed, the failure point is rarely the model itself. It's the workflow the model was supposed to live inside. The model produces output the workflow doesn't consume, or it produces good output but no one owns the next step. The handoff is undefined, or the governance is missing, or the team doing the work was never included in the design.
Specifically:
- No clear owner for the AI output once it leaves the model.
- No defined exception path when the AI is uncertain or wrong.
- No measurement loop, so no one knows whether to invest more or shut it down.
When AI work fails, it's almost always the workflow that broke, not the model.
Where stalls show up
| Stall pattern | Where it shows up | How often we see it |
|---|---|---|
| No production owner | AI output produces results no one is measured on or asked about. Pilot becomes a side project. | Most engagements we review |
| Undefined exception path | Model is uncertain on 5 to 15 percent of cases. No one knows what happens to those cases. | Most engagements we review |
| Workflow not redesigned | AI is bolted onto the existing process. The 80 percent improvement opportunity is left on the table. | Frequently |
The workflow-first methodology
Start with the work, not the model.
Workflow-first means we start by mapping the work, then ask where AI fits, then design the operating model around it. The order matters. Companies that lead with the model usually end up with a tool no one uses; companies that lead with the workflow end up with a tool the team can't imagine working without.
A short example: order-to-cash exception handling
A typical B2B distributor handles 14 percent of orders through manual exception handling. The model question is straightforward: can a classifier sort these into seven recurring exception types with high confidence? In our experience, yes.
The workflow questions are harder:
- Who owns the dispute drafts the AI generates?
- Who reviews exceptions the AI is uncertain about?
- What does the AR team do with the four hours per day that opens up?
- How does the operating model change when 38 percent of exceptions stop touching a human at all?
The model question takes a week. The workflow questions take three.
How we run the Operating Review
Three weeks, in three phases.
Each phase produces an artifact you can read and act on, even if you stop the engagement after that phase. We've never had a client stop early, but the option matters; it's how we keep the work honest week to week.
-
Map the operations. Week one is interview, walk-the-floor, and inventory work across the systems and teams that hold the volume. Output: an Operations Map naming the workflows worth examining.
-
Quantify the tax. Week two pulls eighteen months of transaction data and sizes each workflow in dollars, hours, and risk. Output: an Operations Tax Model your CFO can read.
-
Rank the work and roadmap it. Week three turns the map into ranked recommendations (replace, augment, redesign, or retire) and a 12-month implementation roadmap. Output: a Workflow Recommendations Report, the Implementation Roadmap, the governance artifacts, and a fixed-price implementation SOW for whatever you choose to build first.
What happens next
If your company has tried AI and felt the stall, the Joust Operating Review is the cheapest way to find out whether you're looking at a model problem, a workflow problem, or both. Three weeks, fixed scope, five engagement deliverables, two governance artifacts, and a 12-month implementation roadmap.
To start a conversation, book a 30-minute conversation or email Ron Davis at ron@joustagency.com.
Ron Davis
Founder
Three decades building enterprise platforms. Started Joust to close the gap between strategy decks and the work they're supposed to change.