Your AI agent’s rulebook is already written. It sits in the event history of the processes the agent is about to join, and process intelligence can now read it back to you. That is one of the most practical ideas in enterprise AI this year, and it arrives at a good moment. At Dreamforce in mid-September, Salesforce made Agentforce Coworker available and launched AIforce under the line that AI replaces the user interface. Agents are taking on more of the work people used to do on screen, so the useful question for most CRM teams has moved on from what an agent can do. It is now what an agent should do on its own, and where it should stop.
The encouraging part is that this boundary does not have to be invented in a workshop. It can be discovered from how the process has actually run, which is faster, more specific and far easier to defend. This is the third post in a short series on process intelligence, the discipline Salesforce brought in-house when it completed its acquisition of Apromore in November 2025. The first, See the Process Before You Automate It, covered choosing a first agent. The second, Process Mining Puts a Number on CRM Debt, turned the same instrument on technical debt. Earlier posts on approval gates, formal AI reviews and a single agent catalogue looked at how to control agents once the rules exist. This one is about where the rules come from.
What is agentic AI in CRM, and where do its boundaries come from?
Agentic AI in CRM refers to AI agents that pursue a goal across several steps and take actions in the CRM on their own, rather than only drafting text for a person to review. A CRM agent might qualify an inbound lead, update an opportunity, prepare a quote, triage a service case or chase an overdue renewal, using the records, permissions and automation the platform already holds. Salesforce Agentforce, Microsoft Copilot Studio agents for Dynamics 365 and HubSpot Breeze agents all work this way. What separates an agent from an assistant is that it acts. That makes the definition of its scope, meaning what it may decide alone, what needs approval and what must always escalate, the most important design decision in any deployment.
The analyst Kai Waehner, in his Process Intelligence Landscape 2026 published on 22 July, describes these as “operational boundaries that define what an agent can decide on its own, what needs human approval, and what must always escalate.” A workshop produces the process as people remember it. An event log records the process as it ran, and the boundaries an agent needs are mostly already being applied by people every day, just not written down. If every discount above a certain level has in practice gone to a sales manager, if credit checks are always repeated for one region, or if a particular case type is always handed to a specialist, the event history shows it plainly, with volumes and timings attached. Those patterns are the agent’s boundary in raw form. Process intelligence extracts them, and the team then decides which to keep, which to tighten and which to retire.
There is a reliability benefit too. When Salesforce announced the Apromore deal, Holger Mueller of Constellation Research observed that “process intelligence graphs are a proven and successful approach to make sure agentic AI does not hallucinate and it’s usually more reliable than data based RAG.” Retrieval gives an agent facts. A process model gives it sequence, the knowledge of what normally comes next and what never should. Boundaries discovered this way are also easier to defend to an auditor or a works council, because every rule can point back to the cases that produced it.
What is process conformance for AI agents?
Process conformance for AI agents is the practice of checking an agent’s actions against a reference model of how the process should run, and flagging every step it takes outside that model. It applies conformance checking, explained in Process Mining Puts a Number on CRM Debt, to a new kind of actor. The reference model is usually derived from historical event data and then approved by the process owner. Once the agent is live, each of its actions is logged like any other event and replayed against the model, so skipped approvals, out-of-sequence updates and unusual loops surface automatically. The same model that defines the boundary also measures whether it holds.
The tooling for this is arriving from several directions. Apromore’s Compliance Center, co-created with two large banks, manages risks, obligations and controls and checks them against process data in real time. Waehner describes a decision gate that “checks every automated action and AI recommendation against business rules, confidence thresholds, and constraints before it has consequences.” Celonis launched Agent Mining, which traces the reasoning behind agent decisions, and took its Microsoft Agent 365 integration into private preview on 1 May 2026. Three vendors, one direction: the process model is becoming the governance artefact.
What is shadow AI, and how does process data bring it into view?
Shadow AI is the use of AI tools, models or agents inside an organisation without the knowledge or approval of the people responsible for governing them. It began with staff pasting customer data into public chat tools. It now increasingly means automations and agents with real system access: a script calling the CRM through an integration user, a low-code agent built by one team, or an assistant connected through a personal API key. Shadow AI Has Outgrown Your Policy looked at why policy alone cannot see it. Most shadow AI is well-intentioned, built by people solving a genuine problem faster than the formal route allowed.
Event data is one of the most effective ways to see it, because every write to the CRM is recorded against a user, and automated actors behave differently from people. A step that suddenly completes in seconds rather than hours, a new integration user updating opportunities, or a cluster of records changed outside business hours all stand out in a process view. The constructive response is to treat each discovery as a candidate for the catalogue rather than a breach to close. Teams that do this find useful automation they can scale, and the people who built it become natural champions for the governed version.
Why is CRM data quality important when the process sets the rules?
CRM data quality is important because every decision a person or an agent makes from the CRM inherits the accuracy, completeness and consistency of the records underneath it. Poor data produces confident but wrong recommendations, misrouted leads and forecasts nobody trusts. Good data does the opposite: it lets an organisation automate with confidence, measure what changed and explain why. When agents act on CRM records, data quality stops being a reporting concern and becomes an operational one, because an agent will repeat a data error at machine speed. Investment in data quality therefore pays back twice, once in better human decisions and again in safer automation.
For discovered boundaries there is a specific layer worth attention, which is the quality of the event history itself. Field history tracking on the key objects, consistent status values and reliable timestamps are what make a process model trustworthy. Most platforms keep much of this by default, and switching on history for a handful of extra fields is quick work that improves pipeline reporting whether or not an agent ever runs. Teams that make this small investment early find the boundary-setting step goes faster, because the evidence is already clean enough to act on.
Which steps should you redesign before an agent takes them on?
Waehner makes an observation every agent programme should hear early: “A process built for human execution is rarely the right foundation for an autonomous one.” That sounds like a warning, but it is really a map. The recorded process shows exactly which steps exist only because a human was doing the work: waiting for someone to read an email, re-keying values between systems, an approval loop that is granted almost every time. Those are the steps to redesign first, and usually the ones where an agent creates the most value once they are simplified. Steps with high variation and genuine judgement show up just as clearly, and they are where an agent prepares the work and a person decides.
A practical sequence fits into a month. In the first week, choose one high-volume process and extract around ninety days of event history. In the second, derive the reference model and note the thresholds people already apply. In the third, translate those into the agent’s instructions, permissions and approval gates, and agree them with the process owner. In the fourth, run the agent alongside the team and check its actions for conformance before widening its scope. At the end you have a working agent, a documented boundary and a baseline that proves the improvement.
How does a clear boundary improve CRM adoption rates?
Improving CRM adoption rates comes down to making the system genuinely useful to the people asked to use it. The most effective levers are removing low-value data entry, building screens around how each role actually works, involving users in design decisions, giving managers reports they rely on so they coach from the CRM, and measuring adoption through behaviour such as timely stage updates rather than log-in counts. Adoption rises when people feel the CRM saves them time and reflects their work accurately. Agents can help on every one of those levers, provided users understand and trust what the agent is doing on their behalf.
A discovered boundary helps directly with that trust. When the rules an agent follows come from the team’s own recorded work, people recognise them. A sales rep can see that the agent escalates exactly the discounts a manager would have wanted to review, and a service lead can see that complex cases still reach the specialist. That recognition turns a new agent from something done to a team into something built from its expertise. It also gives managers a clear, shared answer to the question every user asks first, which is what the agent will and will not touch.
The Sirocco perspective
We find this one of the most encouraging developments in agent governance, because it replaces a slow drafting exercise with a fast discovery one. We work across Salesforce, HubSpot and Microsoft Dynamics 365, and the same thing is true on all three: the rules an organisation needs for its agents are mostly already being followed by its people. Apromore makes the discovery step native for Salesforce customers, Microsoft offers Power Automate Process Mining for Dynamics 365 estates, and Celonis remains a strong choice where processes run across several platforms. The route depends on your landscape. The evidence is worth gathering on every one of them.
Our advice is to let the process write the first draft of the rulebook. Start with one process, read what it has been doing, and agree the boundary with the people who run it. Agents built this way earn wider scope quickly, because every expansion can be justified with the same evidence that set the first limit. That is a confident footing for the next year of agentic work. If you would like help reading your agent boundaries out of your own process data, schedule a consultation with our team.
Get in Touch
If you are deciding what your next CRM agent should be allowed to do on its own, we would be glad to help you discover that boundary from your own process data.
