AI Infrastructure for CX
For forty years, the layer that turned engagement into action was a person. That layer is becoming software, and it will not belong to the vendor that owns your system of record.
For about forty years, enterprise software has been organized around a three-part loop that nobody ever bothered to name properly, because one of the three parts was not software at all. It was us.
The loop went like this. A customer, an employee, or a partner engaged through some channel: a phone call, an email, a portal form, later a chat window. A human being read that interaction, interpreted it, pulled up whatever history existed in the system of record, made a judgment, and executed an action. Then, if the process was working, that human recorded what they did back into the system of record so the next human would have continuity and context.
Every enterprise category we take for granted was built to serve one leg of that loop. Geoffrey Moore’s 2011 white paper for AIIM gave us the vocabulary that stuck: systems of record hold the durable truth, systems of engagement mediate the interaction. CRM, ERP, ITSM, and HRIS vendors sold the first. Contact center, collaboration, and portal vendors sold the second. Workflow and RPA vendors automated the mechanical parts of the third.
What none of them sold, because it could not be bought, was the interpretation layer in the middle. That was a person with a headset, a browser with eleven tabs open, and enough tenure to know that this account escalated last quarter for a reason that never made it into a ticket field.
That is the layer AI is now taking over. And once you see it that way, the entire stack rearranges.
We tell ourselves that the system of record is the source of truth. It is not. It is a lossy, delayed, human-authored summary of the truth, structured to fit fields that somebody defined in a requirements document years ago.
Consider what actually happens in enterprise support. A customer sends a four-paragraph email describing a production issue, mentions in passing that their VP is asking questions, and notes that this is the second time this month. An engineer reads all of that. What lands in the CRM is a priority value, a product component, and a resolution note that says “config issue, advised customer.”
The signal was in the email. The record got the summary. Everything that made the interaction predictive of churn, of escalation, of expansion, was discarded the moment a human compressed it into a form.
This is why support leaders are moving to a CRM-less architecture. The ticketing system was never the intelligence. It was the filing cabinet where the residue of intelligence got stored, and it was only ever as good as the human who filled it in at the end of a long day.
A system of intelligence does not wait for a human to summarize. It reads the raw interaction directly, at the source, in full fidelity, across every channel, and it maintains its own representation of what is going on.
That is a much stronger claim than “AI reads your tickets.” The important property is not extraction, it is persistence. A system of intelligence builds and maintains contextual intelligence along four axes at once, and this is precisely what humans were doing badly and what point-solution AI still does not do at all.
Across the session window
A model call has a context window. A customer relationship does not. State has to outlive any single inference, any single conversation, and any single model generation. When the frontier model changes next quarter, the context has to survive the swap.
Across communication channels
The customer who filed the calm ticket is the same customer who was furious on the phone yesterday and vented in the Slack Connect channel this morning. Three systems, three vendors, three data models, one relationship.
Across the timeline
Sentiment at a point in time is nearly worthless. Sentiment trajectory, the second derivative of a relationship, is where the predictive power lives. That requires continuous state, not a classifier invoked on demand.
Across people and organizations
The end user, the admin who opened the case, the executive sponsor who is quiet right now, the account team, the engineer on the third handoff. Enterprise context is a graph of people and their positions, not a single customer ID.
This is the difference between generic AI and contextual intelligence, and it is why a well-built intelligence layer starts to subsume the system of record rather than sitting beside it. Once the intelligence layer holds richer, fresher, more complete state than the CRM, the CRM’s role quietly demotes to what it was always best at: a transactional ledger and a compliance artifact.
Here is the framing that I think matters most for anyone architecting this today.
Borrow the language from networking. Every serious distributed system separates the data plane, which moves packets, from the control plane, which decides where packets should go and under what policy. The data plane is fast and dumb by design. The control plane holds the topology, the state, and the rules.
Systems of action are a data plane. RPA bots, workflow engines, agentic executors, API calls into a fulfillment system: all of them are extremely good at doing a thing reliably, at volume, with an audit trail. None of them are good at deciding whether the thing should be done, to whom, right now, under these conditions, given what happened last quarter.
That decision is a control plane function, and it needs exactly the four axes of context described above. The system of intelligence is that control plane. It answers the questions that come before execution:
Gartner’s guidance points the same way from a governance angle. It warns that applying uniform governance across AI agents will lead to enterprise AI agent failure, because agents differ wildly in autonomy and risk. Differentiated governance is a control plane responsibility. You cannot put it inside each executor and expect coherence.
The clearest live demonstration of a system of intelligence is not in the enterprise. It is in your pocket.
At WWDC in June 2026, Apple introduced Siri AI with what it described as deep, system-wide understanding of personal context and on-screen awareness, shipping with iOS 27 this fall. The demonstrations were exactly the four-axis argument in consumer form: find the email, respond to it, notice the gathering mentioned in Messages, create the calendar event, know which person is which in Photos. Nothing about that is a chatbot. It is a persistent personal context graph acting as a control plane over App Intents, which are the system of action.
It is worth remembering how long this took. Apple announced a personalized Siri in 2024, delayed it repeatedly, and spent two years rebuilding the foundation. The hard part was never the model. It was the context layer.
Apple can build it because Apple has an unusual privilege: the walled garden. Your mail, your messages, your calendar, your photos, your contacts, your location, your app activity. One vendor, one identity, one device graph, one privacy boundary. For consumers, the price of admission is lock-in, and it is a modest price. Most of us pay it happily. I do.
Now try to run the same play inside a company.
No enterprise of any size is a single-vendor shop, and no vendor sells the whole loop. A typical mid-to-large B2B support organization is running Salesforce or Zendesk for cases, ServiceNow for IT and change management, Jira for engineering escalations, Genesys or a similar platform for voice, Slack or Teams for internal and Connect channels, Gong or Chorus for revenue conversations, Confluence and a knowledge base for documentation, Snowflake or Databricks for the warehouse, and Workday for the people graph. That is nine systems from eight vendors before anyone has said the word AI.
Every one of those vendors is now shipping an intelligence layer. Salesforce has Agentforce. ServiceNow, Zendesk, and the rest all have their own. Each is genuinely useful, and each has the same structural limit: it can only be as smart as the slice of the customer relationship that flows through its own product. A support AI that cannot see the sales conversation, the voice call, the Slack thread, and the engineering ticket is reasoning from a fraction of the evidence and presenting the result with full confidence.
This is the structural argument for a neutral, third-party system of intelligence. Not neutral as a marketing posture, but neutral as an architectural requirement:
The protocol layer is arriving to make this practical. MCP’s 2026 specification work and its published roadmap have turned “connect an agent to a data source” from a bespoke integration project into a standard. That standard is what makes a neutral intelligence layer economically viable, because the integration tax that used to protect incumbents is collapsing.
Jerry Chen argued in The New Moats that systems of intelligence would become the next defensible business model, because the moat is not the software, it is the accumulated, workflow-specific context. That was 2017, written before the model layer commoditized. The commoditization only sharpens his point. If everyone can call the same frontier model, the model is not the moat. The proprietary context you feed it is.
Follow the money through the loop and the migration is obvious.
Systems of record lose pricing power.
If the intelligence layer holds the richer state, the record becomes a ledger. Ledgers are valuable, but they are not strategic, and they do not command seat-based pricing forever. This is a multi-billion dollar repricing across CRM, ITSM, and adjacent categories.
Systems of engagement commoditize.
Channels become transport. Voice, chat, and email infrastructure are increasingly a cost line, not a differentiator.
Systems of action get more valuable, not less.
This is the part people get wrong. UiPath, whose platform is explicitly organized around business orchestration, Automation Anywhere, and Pega all become more useful when something upstream is making better decisions about what to execute. An execution layer starved of judgment automates the wrong things faster. Pair a strong system of action with a strong system of intelligence and the returns compound: better decisions, executed reliably, with the outcome fed back as training signal.
The system of intelligence captures the coordination rent.
It is the only layer that sees everything, holds state across everything, and decides. Gartner projects that agentic AI will autonomously resolve 80 percent of common customer service issues by 2029. Note the word “common.” The residual 20 percent is where enterprise value actually concentrates, and it is exactly the set of cases that require cross-system, cross-timeline, cross-people judgment. That residual is a control plane problem.
Customer experience is where this pattern shows up first, for a reason. Support is the highest-volume, highest-signal, most cross-functional unstructured data stream in the enterprise, and it is the one place where getting the judgment wrong is immediately expensive.
At SupportLogic we have built the architecture the argument above describes, so let me be concrete about what a neutral system of intelligence looks like in production:
The write-back matters as much as the read. The system of intelligence does not ask you to abandon your CRM. It integrates with what you run and pushes structured outcomes back into it, which is exactly the continuity job that humans used to perform manually and inconsistently at the end of every interaction.
If you are assessing this category, these are the questions that separate a real control plane from a feature with an AI label.
The enterprise stack is rearranging around a layer that never existed as software before, because for forty years it was staffed by people. Systems of engagement will keep collecting the interaction. Systems of action will keep executing. Systems of record will keep the ledger, and will keep charging less for it than they do today.
The layer that reads everything, remembers across sessions and channels and years and people, decides what should happen, and dispatches it: that layer is new, it is neutral by necessity, and it is where the next generation of enterprise value is going to sit.
We started in customer experience because that is where the signal is richest and the cost of bad judgment is clearest. The pattern does not stay there.
A system of intelligence is the software layer that reads raw interactions across every channel, maintains persistent context about customers, people, and events over time, and decides what action should be taken. It sits between systems of engagement, which capture interaction, and systems of action, which execute. Historically this interpretation role was performed by humans rather than software.
A system of record stores structured, human-authored summaries of what already happened, such as a CRM ticket with a priority field and a resolution note. A system of intelligence reads the full-fidelity raw interaction directly, maintains continuous state rather than point-in-time entries, and produces decisions rather than storage. As the intelligence layer matures it tends to hold richer and fresher context than the record, which demotes the record to a transactional ledger.
Because no enterprise runs a single vendor’s stack. Customer context is spread across CRM, ITSM, voice, chat, collaboration, engineering, and data warehouse systems from many different vendors. A vendor whose primary business is one of those systems can only see its own slice, and has a commercial incentive to keep its slice central. Apple can build a personal system of intelligence because it owns every source inside its walled garden. Enterprises have no equivalent garden.
No, it makes them more valuable. Those platforms are the system of action, the data plane that executes reliably at volume. The system of intelligence is the control plane that decides what should be executed, for whom, when, and with what confidence. Execution without judgment automates the wrong things faster. The two layers compound.
Contextual intelligence is context maintained simultaneously along four axes: across the session window so state outlives any single model call, across communication channels so the same relationship is understood wherever it appears, across the timeline so trajectory rather than snapshot drives prediction, and across people and organizations so the full graph of stakeholders is represented.
Enterprise customer experience. Support generates the highest volume of cross-functional unstructured signal in most companies, and the cost of misjudging an escalation is immediate and quantifiable, which makes it the natural proving ground for an intelligence control plane.
See the architecture in production.
Walk through the control plane with our team, or read how enterprise support organizations are already running it.
Want the latest B2B Support, AI and ML blogs delivered straight to your inbox?