Most companies bought the top layer and never built the foundation. Their customers can tell.
You have seen the diagram. Layers of AI stacked like a wedding cake: Classical AI at the bottom, machine learning, neural networks, deep learning, generative AI, agentic AI on top. Most companies read it backwards. The layers are not a ranking. They are a load-bearing structure — and most enterprises bought the top of it while skipping the bottom.
In a meeting not long ago, a CEO walked me through his company’s AI strategy. It was one thing: a copilot. When I asked what sits underneath it — what the model reads, what rules it obeys, what happens when it is wrong — the room went quiet. Someone looked at the CIO. The CIO looked at the slide.
The numbers back that silence. McKinsey finds 62 percent of organizations are piloting AI agents, but in any given business function no more than 10 percent say they are actually scaling them. The bottleneck is not ambition, and it is not the models — everyone rents those from the same providers. Enterprise infrastructure was built for humans to operate, and almost nobody went back and re-laid the foundation.
Skip a layer and you do not get a weaker result. You get a confident system that is confidently wrong. In B2B, where one account is worth millions and a bad answer travels through a procurement committee, that is not a bug. That is a lost renewal.
And B2B is where the gap is widest. In Deloitte‘s 2026 research across more than 1,000 US suppliers and buyers, 72 percent of suppliers said their sales processes were mostly or highly automated. Only 47 percent of their buyers agreed — and buyers were six times more likely than suppliers to call the process mostly manual. Read that again, because it is the whole problem in one line: internal automation has not translated into the external buyer experience. You automated your side. Your customer still feels the friction.
Layer 1 — Classical AI: the rules that still run your business
What it is, and why it survives
Symbolic AI, expert systems, business rules. It is deterministic: same input, same answer, and you can explain why to an auditor or an angry procurement director. It quietly decides contract pricing eligibility, credit limits, warranty entitlement, approval thresholds, and who gets called first when a customer’s line goes down. In regulated B2B that is not nostalgia. It is the reason the system is allowed to run at all.
Example: SAP anchors its agents in a knowledge graph
SAP did not ship agents and hope. Its agent architecture sits on a knowledge graph that gives agents a structured map of business entities, processes and relationships across the customer’s landscape — so an agent’s action stays tied to real business semantics instead of text that merely sounds right. Its CEO put the stakes plainly: for mission-critical customer processes, “almost right” is not good enough.
The failure mode: rules only answer what someone thought to ask
Every rigid decision tree eventually meets a customer it was not designed for. Then it does the most damaging thing software can do: it says no, politely, for a reason nobody can defend.
Example: Salesforce blocked its own agent from helping a customer
In a rare piece of honest reporting, Salesforce described running Agentforce on itself. It had a hard rule — never discuss competitors — with a block list of rival names. A customer then asked a legitimate question about integrating Microsoft Teams with Salesforce. The agent refused, because Microsoft was on the list. The fix was to replace the rigid rule with a goal: act in the customer’s best interest. Their conclusion is worth pinning above your desk: agents perform best when you tell them what to achieve, not how.
Layer 2 — Machine Learning: where service stops being reactive
What it changes
Supervised and unsupervised learning, classification, regression, anomaly detection. It stops asking what the rule says and asks what the data suggests. This is the layer that converts unplanned downtime — the most expensive event in industrial CX — into a scheduled one.
Example: Siemens Senseye at BlueScope — 53 lines that did not stop
Steel maker BlueScope, running Siemens‘s Senseye predictive maintenance, saved roughly 2,000 hours of unplanned downtime across three years and prevented 53 complete process interruptions by detecting early degradation and intervening before failure. Fifty-three times, a production line did not stop. That is not a report. That is a supplier who changed what it feels like to be their customer.
Layer 3 — Deep Learning: teaching the machine to perceive
What it adds
Perceptions, back-propagation, convolutional neural networks, transformers. Classical machine learning needs a human to define the features. Deep learning learns them from the raw mess: images, audio, sensor streams, free text. It lets a system perceive rather than merely compute.
Example: Schneider Electric cut false rejects by 70x
When Schneider Electric replaced its rule-based vision tools with AI inspection, it doubled production yield and cut false rejects by a factor of seventy. Sit with that second number. A false reject is a good part thrown away, a shipment short, a customer waiting. The old system validated what it was told to look for. The new one learned what a defect looks like.
Layer 4 — Generative AI: brilliant at interpretation, dangerous as the source of truth
The distinction that costs the most money
Summarising two years of account history in four sentences. Turning a dealer’s faxed purchase order into a structured ERP entry. Drafting an RFQ response. Every one of those is interpretation, not decision. The moment it becomes the system of record, it destroys value — because a fluent, confident, wrong answer on a price or a spec is worse than no answer, and your customer cannot tell the difference.
Example: ABB uses it to compress deployment, not to decide
ABB applies generative AI to cut robot-vision training and deployment time by up to 80 percent — and, more importantly, to let integrators and end users retrain the system themselves rather than wait on a specialist. Deployment drops from weeks to hours. That is generative AI doing what it is truly good at: collapsing the distance between a customer’s problem and a working answer, without ever being the final authority on the answer.
Layer 5 — Agentic AI: an assistant informs, an agent acts
The results are real
An assistant tells you something. An agent holds context across steps, decides what to do, calls the systems that do it, and confirms the result. A different reward, and a different risk.
Example: ServiceNow agents that close a sales quote, not just a ticket
At ServiceNow‘s 2026 conference, the company reported that AI specialists across its enterprise customer base already resolve 91 percent of cases without reassignment. The detail that matters for B2B: those specialists span customer relationship management and procurement, and they can close a sales quote autonomously, with a full audit trail. That is not a help desk. That is your commercial motion.
Example: an RFQ assistant that produced $1.8M in quotes in four weeks
The hard numbers in B2B sit in the commercial workflow, not the contact centre. McKinsey’s distribution research documents a water technologies company that built a generative RFP and RFQ assistant in six weeks; within four weeks of launch it generated 1.8 million dollars in quotes across 45,000 customers. A B2B petrochemical firm captured roughly 100 million dollars in additional earnings across six business units with machine-learning dynamic pricing. An industrial distributor surfaced more than two billion dollars in white-space leads. Quoting, pricing, RFQ response — this is where B2B AI actually pays.
The part nobody demos: governance is architecture
McKinsey’s prescription is blunt: every agent needs a digital identity, a defined scope, a named owner, and a logged, auditable trail — with high-impact actions requiring human approval and a mechanism to pause or override the machine. Somebody must be able to stop the machine at two in the morning. That is the price of letting software quote a price in your customer’s name.
Layer 6 — The industrial layer everyone leaves off the diagram
Why the diagram is incomplete
That diagram stops at software. In industrial B2B the customer experience is physical. Underneath the models sit sensors, edge compute, and connected equipment producing telemetry every few seconds. Above them, the digital twin.
Example: Rolls-Royce sells availability, not engines
Rolls-Royce calls its version the IntelligentEngine: an engine that is connected, contextually aware and comprehending, linked to a digital twin and to every other engine in the fleet. What that produces for an airline is not a maintenance report. It is availability — aircraft that fly when the airline says they will. The company is even developing snake-like robots to crawl inside engines and inspect places a human never could.
The IIoT layer supplies the signal. Machine learning finds the pattern. Deep learning reads the parts that are not tidy. Generative AI explains it. An agent acts. Your customer experiences none of that as a stack. They experience a supplier who knew the machine was failing before they did, arrived with the right part, and never let the line stop.
Layer 7 — When the customer becomes a machine
One more layer is forming. As agents mature they stop serving your customers and start being them: procurement systems that evaluate suppliers, APIs that reorder without a human, agents that shortlist vendors before a person sees a name. I have argued this is a structural break in customer experience, not a new channel. Machine buyers do not care about your brand story. They parse structured data, consistent performance, verifiable claims. If your pricing, specs and trust signals are not machine-readable, you are not losing the deal. You are not in it.
So what does a full stack actually look like?
Notice the pattern. Not one of them led with the chat window.
SAP shipped the knowledge graph before the agents. Siemens earned the right to a generative copilot only because years of sensor data and anomaly models already sat underneath it. ServiceNow released its AI Control Tower — which discovers agents as they appear, risk-scores them and enforces least-privilege access, even for agents it did not build — in the same box as the autonomous workforce. They shipped the brakes with the engine. That is what makes 91 percent defensible to an auditor.
And Salesforce, running its own agents, learned it the expensive way: the failure was never the model. It was the rule underneath it.
The honest test — three questions for your team this week
- Where are our business rules actually written down, and who owns them?
- Can our AI cite the system of record for the answer it just gave a customer?
- If an agent does something wrong at 2 a.m., who is paged, and what do they press?
If those produce a confident answer, you have a stack. If they produce a slide, you have a demo.
Classical AI gives you defensibility. Machine learning gives you foresight. Deep learning gives you perception. Generative AI gives you fluency. Agentic AI gives you action. The industrial layer connects you to the physical world your customers live in. Remove any one, and the layer above it stops being a product.
The question is not which layer to buy next. It is which layer you skipped — and whether your customers have already figured it out.
👉 Stay ahead of CX, AI, and innovation — Subscribe to my weekly LinkedIn Newsletter “CX Insights by Ricardo S. Gulko.
Share it if it’s useful — and let’s connect for more. Ricardo Saltz Gulko
My columns in several respected CX publications.
- My recent articles on Eglobalis: https://www.eglobalis.com/blog/
- My recent articles on the European CX Organisations ECXO.org : https://ecxo.org/blog/
- My recent articles on CMSWire: https://www.cmswire.com/author/ricardo-saltz-gulko/
- My German articles on CMM360: https://www.cmm360.ch/author/ricardo/
- My old articles on CustomerThink: https://customerthink.com/author/rgulko/
