Industrial AI runs on an asset inventory most operators don't have

On June 18, Accenture said it would spend about $4.2 billion to buy visibility into industrial networks. The deal is framed as cybersecurity — a majority stake in Dragos, plus all of runZero and NetRise — but strip the security language off it and you’re left with the part of industrial AI almost nobody budgets for: knowing what’s actually on the network. That asset inventory, the unglamorous map of every connected device on a plant floor or a grid, is the substrate industrial AI runs on. Most operators don’t have one. Accenture just paid four billion dollars betting they’ll pay to get it.

The headline is a security story. The operator read is a data story.

TL;DR: Accenture’s ~$4.2 billion acquisition of Dragos, runZero, and NetRise is, underneath the threat-detection framing, a bet on asset visibility — the ability to see and catalogue everything connected to an operational-technology network. Dragos estimates only about 10% of OT networks have any visibility at all, and operators rank asset inventory their number-one investment priority (SANS, 2025). That same missing map is what kills industrial AI projects: a predictive-maintenance model or an autonomous control loop can only reason over equipment it can see. The model was never the bottleneck. The inventory of what’s physically there is. Build the asset map first; the AI is the last step on top of it.

What Accenture actually bought

The three companies do different jobs, but the through-line is the same: turning an opaque industrial environment into something legible.

Company	What it does	The capability that matters
Dragos	Vendor-neutral OT threat detection	Sees activity across industrial control systems regardless of equipment brand
runZero	Asset discovery, exposure assessment, attack-surface intelligence	Finds and enumerates what’s actually connected — including the devices nobody knew were there
NetRise	Firmware analysis, software supply-chain visibility	Knows what’s running inside the device, down to the firmware

Dragos CEO Robert Lee described the goal plainly: give industrial operators “one place to see everything on their OT network, understand what’s running on it.” That sentence is the whole deal. The combined platform carries about $208 million in annual recurring revenue as of June 2026, up 53% year over year, and Accenture CEO Julie Sweet called it “expanding our addressable market, creating a new platform-led growth opportunity.” Accenture has grown its security business from $700 million in 2016 to $10 billion in fiscal 2025; this is the next leg of that, and the leg is visibility.

Notice what’s not in the pitch: a better model. Nobody is paying $4.2 billion for smarter detection algorithms. They’re paying for the map.

The visibility gap is real, and operators know it

Here’s the number that reframes the deal. Dragos estimates only about 10% of OT networks worldwide have any kind of monitoring or visibility in place. Ninety percent of the industrial world is running equipment it can’t fully see.

This isn’t a secret inside those plants. In the SANS 2025 State of ICS/OT Security survey, drawn from more than 330 industrial practitioners, asset inventory and visibility was the number-one technology investment area in 2025 — about half of respondents — and it stays the top priority heading into 2026 and 2027, at roughly 54%. The people running these environments already know the gap is foundational. The market just put a price on it.

And visibility pays. The same body of OT research found that organizations with comprehensive visibility contained ransomware incidents in about five days, against an industry average of 42. The lesson generalizes past security: you respond faster, decide better, and automate more safely when you can actually see the thing you’re acting on.

Why this is an AI story, not just a security one

An AI model acts on the data it’s given. An agent acts on the systems it can reach. In an OT environment, both run into the same wall: a large share of the equipment isn’t in any system of record, so it isn’t in the data the AI sees either.

Think about what gets sold as “industrial AI” — predictive maintenance, anomaly detection, autonomous scheduling of assets, energy-optimization loops. Every one of those is a function of a complete, current picture of the equipment. A predictive-maintenance model that doesn’t know a pump exists will never predict its failure. An anomaly detector blind to half the controllers on a segment will call a quiet network healthy. An autonomous loop optimizing throughput across machines it can’t fully enumerate is optimizing a fiction. The model can be state of the art. If the asset map has holes, the output has holes, and the holes don’t announce themselves.

This is the same failure mode as autonomous field-service dispatch running on field data nobody maintains: the algorithm is solved, the substrate underneath it isn’t, and the automation faithfully amplifies whatever the substrate already is. On a plant floor the substrate is the asset inventory. Get it wrong and you don’t get a visible error — you get a confident model quietly reasoning over equipment that isn’t in its world.

Securing it and automating it are the same prerequisite

Accenture is buying this to defend critical infrastructure. But defense and automation start at the identical place. You can’t protect an asset you don’t know exists. You also can’t apply AI to a process running on equipment you can’t see. The first step for both is the same boring artifact: a complete, current inventory of the operational environment — what’s connected, what firmware it runs, what it communicates with, and what its normal looks like.

That overlap is why this deal reads as more than a cybersecurity acquisition. The visibility layer Accenture paid up for is dual-use. The map that lets you catch an intrusion is the map that lets you trust a predictive model. Operators tend to fund the security version because there’s a compliance gun to their head, and skip the automation version because it feels optional — then they’re surprised when the AI pilot underperforms. It’s the same missing map both times.

There’s a close cousin to this on the software side. Shadow AI inside companies is an inventory problem, not a policy problem — you can’t govern the tools you can’t see. OT asset visibility is the same idea pointed at the physical world: you can’t automate the machines you can’t see. The pattern repeats because the bottleneck repeats. It’s legibility, every time.

The working version: build the map, then point AI at it

The order that works is the unglamorous one, and it doesn’t require a $4.2 billion acquisition to start.

Discover and catalogue what’s actually connected. Not the spreadsheet from the last audit — the live picture. Asset discovery on an OT network routinely turns up devices nobody remembered installing: a contractor’s laptop still bridged in, a legacy HMI on an IP no one documented, sensors added during a retrofit that never made it into the CMMS. That gap between the drawing and the reality is exactly the gap that breaks an AI model trained on the drawing.

Write down what the signals mean. A live device list isn’t enough; you need agreed definitions of the states your AI will act on. When one system reports a machine “idle” and another reports it “down” for the same physical condition, an autonomous loop built on top will make decisions on a contradiction. That’s a data contract — the written, owned statement of what each signal means — and it’s the part no vendor platform writes for you, because it’s your operation’s reality, not their product. This discovery-and-contract work is what I do before anything gets automated.

Then add the AI. With a complete asset map and agreed definitions behind it, predictive maintenance, anomaly detection, and scheduling have something solid to stand on. Bought in that order, the AI removes real work. Bought first, it just automates your blind spots at machine speed.

Accenture didn’t spend $4.2 billion on a model. It spent it on the ability to see what’s there — because seeing what’s there is the part that was missing, and the part everything else depends on. That’s the layer worth fixing before you turn an industrial AI project on.