Ask almost any industrial AI company what protects it from competitors and the answer will include the word data. The claim is intuitive. A company that has seen more machines, more faults and more outcomes should build better models, which should win more customers, which should produce more data. In practice, that loop is weaker and slower than the pitch suggests, and data on its own protects less than many founders and investors assume.
This article looks at what can actually make an industrial AI business hard to copy. Drawing on public research and the economics of industrial deployment, it describes advantages a company could earn, without claiming that AFFERENT, or any other particular company, has earned them.
The case against easy data moats
The most widely cited critique is a 2019 essay by Martin Casado and Peter Lauten of Andreessen Horowitz, The Empty Promise of Data Moats. They argued that most so-called data network effects are really scale effects, and that data scale effects often weaken over time. The cost of acquiring new, useful data tends to rise, while the value of each additional piece of data tends to fall, because more of it overlaps with what a company already has. Data can also go stale. Their conclusion was that defensibility is not inherent to data itself.
Their analysis focused on enterprise software, but much of it applies to industrial AI. The first thousand examples of an imbalanced fan teach a model a great deal. The next ten thousand teach it much less. Meanwhile, competitors can buy similar sensors, read the same research and collect their own minimum viable dataset.
The same essay identified where data does protect a business: where sources are scarce, where owners are reluctant to share with more than one vendor, where cleaning and understanding the data needs specialized know-how, and where a vendor earns enough trust to be given sensitive data that others never see. Those conditions describe industrial infrastructure unusually well.
How a data lead shrinks: an illustration
A simple illustration shows why. Suppose detection accuracy for one fault type rises with confirmed examples along a typical diminishing curve. The figures below are invented for illustration, not measurements of any system.
| Confirmed examples | Illustrative detection accuracy |
|---|---|
| 50 | 70% |
| 500 | 85% |
| 5,000 | 88% |
| 50,000 | 89% |
With 500 examples, a company has a 15-point lead over a newcomer with 50. If the newcomer reaches 500 two years later while the leader reaches 5,000, the lead has shrunk to 3 points, even though the leader's dataset is ten times larger. Measured on volume, the leader's advantage grew. Measured on what customers experience, it nearly vanished.
The picture changes in the long tail. If the leader has examples of rare fault types, unusual machines or extreme operating conditions that the newcomer has never seen, it can detect problems the newcomer simply cannot. In industrial settings, those rare cases are often the expensive ones. Coverage of the tail, not the size of the head, is where data advantages last.
A framework: four sources of advantage
Looking across industrial technology businesses, four kinds of advantage come up repeatedly. None is sufficient alone. Together, they can make a position difficult to dislodge.
1. Data, but the right kind
Raw sensor data is the least defensible layer, since anyone with a similar device can collect it. The defensible layers are the ones that take time and cooperation to build: consistent measurements from a known instrument, detailed context about each asset, and confirmed outcomes linking signals to what technicians actually found. As we argued in the next data moat is physical, the record of what happened is scarcer than the signal itself.
2. Deployment
Industrial AI depends on devices installed on real machines. Each installation takes a site visit, a safety induction, physical mounting, network approval and integration work. That effort is slow for everyone, including later entrants. Once a device is installed and trusted, there is little reason for a customer to tear it out. We explored the economics of this in why the next great infrastructure company may start with hardware.
3. Workflow
A finding that appears on a separate dashboard is easy to replace. A finding that arrives as a work order in the maintenance system a plant already uses, with the right parts and priority attached, becomes part of how the plant runs. Workflow integration builds switching costs that have nothing to do with model quality.
4. Trust
In industrial settings, trust is earned one correct call at a time and lost quickly. A system that cries wolf trains technicians to ignore it, as we described in our article on false positives. A system that has been right repeatedly earns the right to change maintenance plans, capital budgets and safety decisions. That track record cannot be bought.
What makes industrial data valuable
If data is to contribute to defensibility, it needs properties that competitors cannot easily reproduce.
| Property | What it means | Why it is hard to copy |
|---|---|---|
| Fidelity | Sample rate, bandwidth and mounting good enough to capture early faults | Moderate: depends on hardware design and installation practice |
| Consistency | The same instruments and processing across the fleet | Requires control of the measurement over years |
| Context | Asset type, speed, load, environment and maintenance history | Must be captured at installation and kept current |
| Outcomes | Confirmed faults, repairs and false alarms linked to signals | Needs cooperation from maintenance teams and time |
| Rights | Contractual permission to learn from the data | Set customer by customer; increasingly regulated |
The last row deserves attention. In Europe, the Data Act entered into application on 12 September 2025. It gives users of connected products, including businesses, the right to access the data produced through their use of machines and devices, and to share it with third parties such as repair and maintenance providers. Rules like this make it harder for a vendor to treat customer data as its private property. Any data advantage has to rest on value delivered and permissions granted, not on locking data away.
Trust has a formal side too
Trust is not only a matter of reputation. Buyers increasingly ask how AI systems are governed. The U.S. National Institute of Standards and Technology published its AI Risk Management Framework to help organizations manage risks from AI systems, including validity, reliability, safety, security, transparency and accountability. For an industrial AI vendor, that translates into practical questions. Can the vendor explain why a finding was raised? Does it measure and publish its own false alarm rates? Does it tell customers when a model changes? Vendors that answer these questions well build trust faster than those that rely on accuracy claims alone.
How advantages erode
Each source of advantage can weaken. Data loses value when formats change with each hardware generation or when rights to use it are unclear. Deployment loses value when devices are unreliable or installation is costly, because customers then welcome an alternative. Workflow advantages fade when integrations break or lag behind the customer's systems. Trust collapses after a run of false alarms or a missed failure that the system should have caught.
The common thread is that defensibility in industrial AI is maintained, not owned. It requires continuing investment in hardware quality, model validation, integration and customer relationships. Any business that treats its early lead as permanent is likely to lose it.
A scorecard for evaluating claims
Investors, partners and customers hear defensibility claims constantly. A few questions separate substance from narrative.
- Outcomes: How many confirmed fault cases, by type, are linked to sensor records? How are they collected?
- Consistency: How much of the data comes from a single, controlled instrument design?
- Rights: What do customer contracts allow, and how would regulation like the Data Act affect that?
- Retention: What share of installed devices are still active after two years?
- Workflow: What share of findings become work orders in the customer's own systems?
- Trust: What are the measured false alarm and missed detection rates, and are they shared with customers?
Credible answers to most of these suggest something real. An answer that rests on the size of a dataset alone probably does not.
Implications for strategy
For builders, the lesson is to invest where advantages compound and are hard to copy: reliable hardware that stays installed, structured outcome capture, deep workflow integration and transparent performance reporting. Data then becomes a result of doing those things well, rather than a goal in itself. That is the core of any serious industrial AI platform strategy.
For investors, the lesson is patience and skepticism in equal measure. The advantages described here take years to build and show up first in retention, workflow adoption and measured accuracy, long before they show up in market share. Claims about data scale that are not backed by those operating measures deserve hard questions.
For buyers, the lesson is to look past claims about data volume and ask how a vendor earns and keeps trust. The vendor that measures itself honestly is more likely to still be useful in ten years.
