Every wave of enterprise technology has started in the industries with the most digital infrastructure and worked its way down to the ones with the least. Cloud started in SaaS, then went to media, then to retail, then to manufacturing. Data warehouses started in finance, then went to e-commerce, then to logistics. Agentic AI is following the same path, and the people building the tools have mostly built them for the first two stops.
That’s a problem, because the biggest prize is the third stop.
The shape of the opportunity
Industrial operators — we use the term loosely to mean anyone running a physical supply chain: precision manufacturers, aerospace and defense primes and their tier-1/2 suppliers, industrial services, logistics, specialty chemicals — share a specific profile that makes them unusually well-suited to agentic AI.
- High coordination load, low data volume. A mid-market manufacturer processes maybe a few thousand invoices a month. Not the billion-a-day volume you need to justify a data lake. But every one of those invoices touches 4–6 people: buyer, AP clerk, plant manager, controller, sometimes the CFO. That’s a lot of coordination per transaction.
- Heterogeneous systems that don’t talk to each other. The ERP is from 2011. The CRM is something someone set up in a weekend. The warehouse management system predates the ERP. Procurement runs half-on paper. Every integration is a quarter-long project.
- Exception-heavy workflows. Nothing in a precision manufacturing environment is routine. Every order has a spec sheet, every shipment has a non-conformance report waiting to happen, every invoice has a discrepancy.
- Skilled labor shortage. Finding and keeping a senior controller, a quality engineer, or a production planner is harder than it’s ever been. The work that agents can do is exactly the work those people are currently overloaded with.
Those four conditions make industrial operators a textbook fit for agentic AI. High judgment work, high coordination overhead, heterogeneous systems, and a human capacity constraint that’s getting worse, not better.
Why the first wave of AI tools didn’t fit
Most of the enterprise AI tools built in the last two years were designed for SaaS companies and financial services firms. That meant they were built around three assumptions that don’t hold in industrial environments:
Assumption 1: The data is clean and in one place.
Not even close. A typical manufacturer has data spread across ERP, MES, WMS, CRM, email, Slack, shared drives, scanned PDFs, handwritten receiving notes, and at least one Excel file labeled FINAL_v7_real.xlsx. Any tool that needs a well-structured data warehouse as a prerequisite is dead on arrival.
Assumption 2: The users are technical.
The people who will actually use the agent in a precision manufacturing environment are the AP clerk, the buyer, the production planner, and the plant controller. They are not prompt engineers. They do not want to learn a new interface. They want the work to show up in their existing tools — the ERP, email, a Teams chat — already done.
Assumption 3: You can ship a general-purpose tool.
Every finance, procurement, and quality workflow in a mid-market manufacturer is a local dialect. The policies, the approval chains, the exceptions, the document templates: all of them are idiosyncratic. A general-purpose AI tool can’t learn those local dialects without months of configuration by someone who deeply understands both the tool and the business. There is almost no one in that intersection.
The vendors who built tools for the first two waves kept running into these three walls in industrial accounts and mostly retreated to easier markets. That’s the gap.
What fills the gap
The tools that will actually work for industrial operators share a few properties:
- They install inside the existing stack. No rip-and-replace. No new interface for the users. Agents run inside the ERP, email, and chat tools people already use.
- They’re deployed, not configured. Someone who understands both the technology and the industry delivers the first working agent. The client doesn’t have to figure it out from a documentation site.
- They handle messy data natively. PDFs with stamps, scanned documents, emails with attachments, chat threads. Messiness is a feature of the environment, not an edge case.
- They prove ROI on a short clock. Industrial CFOs are skeptical by nature and don’t sign six-figure deals on vibes. The tool has to earn its keep in weeks, not quarters.
- They respect the regulatory reality. If you’re serving defense supply chains, you need to think about ITAR, CMMC, data sovereignty, and US-owned hosting from day one. Most SaaS-native AI tools never had to.
That’s what we built Forge to be. Not because we wanted to build another AI platform — there are plenty of those — but because we kept meeting industrial operators who had a clear problem, a budget, and a willingness to move, and no vendor who could actually deliver.
The eighteen-month compression
Here’s the part that surprises people: we think industrial AI adoption isn’t going to happen gradually. It’s going to happen in a compressed window, roughly between now and late 2027, and the operators who wait are going to look up and find their peers already have 20–30% capacity advantages they can’t easily close.
Three things are driving the compression:
- The underlying models got good enough around late 2024. The multi-step reasoning that agents need actually works now. It didn’t in 2023.
- The skilled labor shortage is getting worse. There’s no scenario where a mid-market manufacturer hires its way out of the current controller / buyer / planner deficit. Agents are the only option that scales.
- The cost of falling behind is visible in the P&L. Operators who deployed early in 2025 are already showing capacity and cycle-time advantages that their peers can see in quarterly calls. Once a competitor gets a 60% cycle-time improvement on invoice reconciliation, you don’t have two years to think about it.
This is why we’re taking ten founding clients in 2026 and no more. Not because we don’t want the business — we do — but because the delivery pod can only do about ten deployments properly in a year if they’re going to land the way we need them to. The operators who get in early are the ones who set the local benchmark. The ones who wait will be playing catch-up against teams that have already compounded their advantage.
The first wave of AI built tools for people who already had data and engineers. The next wave has to build tools for people who have neither — and show up at their door ready to deploy.
That’s the bet we’re making. If you run a precision manufacturer, a defense supplier, a logistics operation, or any industrial services business, and the coordination work is eating your people, we’d love to show you what 90 days actually looks like.