Read or skip? For enterprise architects and delivery leads deciding where physical AI lands first. If you’re tempted to pilot whatever sits at the top of Deloitte’s chart, read on — the ranking and the deployment order are not the same list. If you’ve already mapped your physical-AI work as a dependency graph, you can skip this one.
Deloitte’s State of AI in the Enterprise asked 3,235 business and IT leaders a deceptively simple question: which area of physical AI will have the greatest impact on your industry? The answers came back as a tidy leaderboard (Figure 11). Intelligent security systems and smart monitoring led on 21%, collaborative robotics a point behind on 20%, digital twins on 19%, then IoT-driven retail (16%), autonomous logistics (13%), smart materials (7%) and a residual 3%.
The instinct in most programmes is to read that list top-down and pilot the winner. That instinct is wrong — not because the data is wrong, but because the question measures belief, and belief is not a delivery plan.
Belief, readiness and sequence are three different axes
The survey asked what leaders expect to matter. It did not ask what is ready to deploy, and it certainly didn’t ask what order to deploy it in. Those are three separate measurements, and they don’t line up:
- Belief tells you where attention and budget will flow.
- Readiness tells you what can survive contact with production today.
- Sequence tells you what has to come first so everything else has something to stand on.
Conflate them and you end up commissioning a collaborative-robotics cell that’s waiting on sensor data nobody instrumented, validated against a digital twin nobody built. The capex is spent, the cell is idle, and the post-mortem concludes that “physical AI isn’t mature” — when the real failure was sequencing.
There’s a readiness signal hiding inside Deloitte’s belief survey, and it’s worth pulling out. The leaders interviewed said smart monitoring and digital twins are already transforming operational processes — one example being 3D mapping of stores to drive both interior design and VR-based staff training. That’s not a prediction. That’s two of the seven categories already in production while the rest are still mostly aspiration.
Read the chart as a dependency graph
Group the seven categories not by score but by what depends on what. Three layers and a frontier:
| Category | Belief (greatest impact) | Delivery layer |
|---|---|---|
| Intelligent security / smart monitoring | 21% | Sense |
| IoT-driven retail | 16% | Sense |
| Digital twins | 19% | Simulate |
| Collaborative robotics | 20% | Act |
| Autonomous logistics | 13% | Act |
| Smart materials | 7% | Frontier |
| Other | 3% | — |
Source: Deloitte, State of AI in the Enterprise (Figure 11), n=3,235. Layer grouping mine.
Layer 1 — Sense
Smart monitoring, intelligent security and IoT-driven retail are the instrumentation. They turn the physical environment into a signal everything above them consumes. They’re also the most mature categories on the chart — the ones Deloitte’s interviewees say are already changing how work gets done. Low blast radius, fast payback, no moving part to crash into a person. This is where you start, regardless of where it ranks.
Layer 2 — Simulate
Digital twins. A twin is a test rig: a physics-grounded replica where you rehearse a change before it touches the floor. Deloitte’s store-scanning example — building an interior-design and training environment from an actual store — is a twin doing exactly this job. And a twin is only ever as good as the sensor feed beneath it, which is precisely why it sits on top of the Sense layer rather than beside it. No instrumentation, no useful twin.
Layer 3 — Act
Collaborative robotics and autonomous logistics. This is where physical AI moves mass in the real world, and where a mistake has a blast radius measured in injuries and downtime, not dashboard noise. Cobots need the Sense layer to perceive their surroundings and the Simulate layer to rehearse before they act. Deploy them without both and you are commissioning the moving parts of an engine before the wiring is run and the gauges work.
The frontier — Smart materials (7%)
Genuinely early. Worth watching, worth a research line if it’s core to your industry, not worth a place on the critical path. Treat the 7% as honesty, not pessimism.
The robotics trap
Here’s the non-obvious part. Collaborative robotics ranks second on belief — one point off the top. Read the chart top-down and it’s your second pilot. Read it as a dependency graph and it’s your third phase. The capability leaders are most eager to deploy is the one that depends on the two layers most of them haven’t finished building.
I’ve watched this pattern repeat across four platform shifts. In the move from mainframe to client-server, everyone wanted the rich graphical front end and ended up rebuilding the network and the database underneath it. In the shift to cloud, everyone wanted lift-and-shift and discovered they’d skipped the landing zone and identity model. The most visible layer is almost never the one you commission first. Physical AI has the same shape: robots are the turbines; sensors and twins are the wiring and the instrumentation. You don’t spin the turbine before the gauges read true.
Deloitte’s chart ranks physical AI by what leaders believe. Your delivery plan should rank it by what everything else depends on. Those are rarely the same order.
How to sequence on Monday
A practical reading of Figure 11 for whoever is holding the delivery plan:
- Instrument before you automate. Stand up the Sense layer first — monitoring, security telemetry, IoT feeds. It pays for itself and it is the data foundation everything above it consumes.
- Build the twin before you trust the robot. No autonomous action reaches the floor until a physics-grounded twin can rehearse it. The twin is your safety case, not a nice-to-have.
- Treat robotics as a phase-three capability, whatever its survey rank. Gate it explicitly on the maturity of the two layers beneath it.
- Ring-fence the frontier. Smart materials and the long tail get a research budget and a watching brief, not a delivery milestone.
- Re-run the question inside your own walls. Don’t sequence off a cross-industry average — ask your own operators which layer is weakest today. The 21% headline belongs to someone else’s industry; your constraint is local.
The leaderboard isn’t useless. It tells you where the market’s confidence sits, and confidence moves budgets. But confidence is an input to your business case, not to your build order. Sequence by dependency, gate by readiness, and let belief argue for the funding — in that order.
The takeaway
- Figure 11 measures belief, not readiness or sequence — three different axes that don’t align.
- Read it as a dependency graph: Sense → Simulate → Act, with smart materials as a frontier bet.
- Sensing (21%, 16%) and simulation (19%) are mature and foundational. Start there.
- Collaborative robotics (20%) is second on belief but third in delivery order — it stands on the layers below it.
- Sequence on your own weakest layer, not a cross-industry average.
Source: Deloitte, “State of AI in the Enterprise: The untapped edge,” Deloitte AI Institute, January 2026. Figure 11, “Types of physical AI expected to have the greatest impact,” n=3,235. The survey reached 3,235 business and IT leaders across 24 countries and six industries (August–September 2025). Available at deloitte.com/us/state-of-ai.
Leave a Reply