Grant Thornton released its 2026 AI Impact Survey in April with a number that should stop every manufacturing CFO mid-deck: of 100 manufacturing respondents, exactly zero reported significant revenue uplift from their AI investments. Zero reported significant cost savings either. Across every other industry the firm tracked, 12% of executives said AI was already moving the top or bottom line. Manufacturing — the sector that talks the most about AI — is alone at the back of the pack on results.
This is a strange place for the industry to land. Manufacturing has been investing in machine learning since long before "agentic AI" became a deck slide. According to a separate Redwood Software outlook , 98% of manufacturers are exploring AI-driven automation. Only 20% say they feel fully prepared to use it at scale. The exploration-to-execution gap is real, and the cost of staying inside it is starting to show up in margin pressure.
What's actually broken — and what isn't The survey points the finger at three things, and none of them are the model.
The first is governance. Only 7% of manufacturers have a tested incident response playbook for AI failures. When a vision system mislabels a defect, a forecasting agent under-orders a critical component, or a scheduling engine commits to a date the shop can't hit, most manufacturers have no defined path to catch it, escalate it, or prevent the next one. That isn't an AI problem. It's an operations problem that AI keeps hitting.
The second is data. Grant Thornton's follow-up analysis notes that "clean" data isn't the same as "AI-capable" data. A manufacturer can have well-formed ERP, MES, and quality records and still not have the cross-system context an agent needs to make a useful recommendation. The plumbing between systems — and the workflows that depend on it — was never designed for autonomous decisions. It was designed for someone with a forklift to act on.
The third is strategy. The survey describes most manufacturing AI initiatives as "vendor-defined competitive reactions" rather than deliberate margin plays. That tracks with what we see in the field. A control engineer adopts one platform's vision tool, a quality director pilots another's defect classifier, finance buys a forecasting add-on, and the COO inherits a portfolio of disconnected pilots that don't share data, don't share governance, and don't compound.
Why "more pilots" won't fix it There's a tempting reading of the survey that says manufacturers just need more time and more pilots. PwC's 2026 AI Performance Study argues against that view directly: 75% of AI's measurable economic gains are being captured by the top 20% of companies, and those leaders are not running more pilots. They're running fewer, more deeply integrated ones — wired into the operational workflows that touch revenue.
This is the part that doesn't make a great press release: AI ROI in manufacturing comes from the unsexy work that surrounds the model. Intake. Routing. Approvals. Exception handling. Status updates back to ERP. Whoever owns that orchestration layer captures the value. Whoever ships a model and walks away does not.
A practical example: predictive maintenance. The model can flag a likely bearing failure on Line 3 with high confidence two weeks out. That's worthless on its own. The value shows up only when the prediction triggers a parts-availability check, opens a maintenance work order with the right parts kit, sequences it against the production schedule, alerts the technician with the procedure, captures the actual replacement, and feeds the outcome back to retrain the model. Most manufacturers have the prediction. Almost none have the rest of that loop wired together.
The orchestration layer is the missing product This is exactly the gap a unified operations platform is built to close, and it's why Symphona has spent more time on the workflow side of AI than on the model side. Symphona Flow gives manufacturers a no-code process layer that connects ERP, MES, CMMS, quality systems, and supplier portals, so an AI prediction or recommendation can trigger a real chain of work — not just a notification. The loop closes inside one platform, with full audit trails the survey explicitly calls out as missing.
Symphona Serve handles the human side of that loop — the work that gets dispatched to maintenance technicians, quality engineers, supplier-facing buyers, and shop floor leads. Tasks intake from a Flow process, an AI Agent, or a form, get assigned to the right person with the right context, and roll up to dashboards leadership can actually use to track whether AI investments are converting to throughput.
When the AI is wrong — and it will be — Symphona Resolve is the structured fallout layer. Predictions that miss, agent recommendations that conflict, and integration errors all land in a single queue with SLAs, root-cause tracking, and the option to auto-resolve common failures. That's the "tested incident response playbook" Grant Thornton found 93% of manufacturers don't have.
The point isn't that any one product is magic. The point is that AI without orchestration, task management, and exception handling is a science project. With those three layers in place, the same model that produced zero ROI in the survey starts producing the cost savings the rest of the economy is already booking.
What manufacturers should actually do this quarter Three concrete moves, none of which require ripping anything out.
First, pick one operational metric the C-suite already cares about — first-pass yield, planned-to-actual variance, warranty claim cycle time, supplier on-time delivery — and reverse-engineer the workflow that produces it. Map who touches the work, what systems they jump between, and where the handoffs break. That map is the AI roadmap. Anything not on it is a science project.
Second, separate the model from the workflow. Most manufacturers have decent AI tools sitting unused because nothing routes work to them and nothing acts on their outputs. A no-code orchestration layer can wire them in without a six-month systems integration project, and it gives ops leaders something they don't currently have: a single place to govern every agent and process touching the floor.
Third, instrument the loop. The survey's most damning finding is that manufacturers can't tell whether AI is working. Without exception tracking, SLA monitoring, and outcome capture flowing back to the model, there's no way to prove ROI to the board — or to retrain the model on what's actually happening in production.
If you're a manufacturing operator looking at the AI proof gap and trying to figure out which of your pilots can actually graduate to production, explore how Symphona works for manufacturing or book a consultation . We can map the workflows around your existing AI investments and show you which of them have a real shot at the ROI the rest of the economy is already collecting.
.svg)
.svg)
.avif)
.avif)
.avif)
.avif)
.svg){kind=icon}
.svg){kind=icon}
.svg)
.svg)
.svg)
.svg)
.svg)
