AI agents need boring rules to do useful work

The autonomous enterprise still needs a rulebook

Many vendors talk about the autonomous enterprise, where workflows inside the enterprise software stack run with little or no human intervention.

The phrase has obvious appeal. It suggests speed, efficiency and less manual work. It also gives vendors a clean way to describe a future in which AI agents, assistants and copilots move work through systems without people having to click through every step.

But as many organizations found during the initial AI fever, the concept can become, at best, a gross misnomer and, at worst, marketing gobbledygook if it suggests that cheaper machines can simply take over and human roles will no longer be needed.

AI agents can summarize information, extract data, route work, generate recommendations, draft responses, identify compliance gaps and automate routine tasks. But automation does not erase the need for control, observability, testing, governance, authentication and human judgment. If anything, it makes those old enterprise disciplines more important.

The enterprise version of agent autonomy does not mean "allow the agent to take whatever actions it believes are best." It means "allow the agent to work within a clearly defined process that the business can explain."

That is the difference between autonomy as a marketing promise and autonomy as something an enterprise can actually use.

Autonomy still needs control

Salesforce's recent Agentforce updates point in that direction. The company is still promoting agentic AI, but it is also adding the less glamorous machinery enterprises need to manage agents once they go live: dashboards, testing tools, orchestration features, human checkpoints, governance controls and authentication for higher-risk workflows.

The company's own read from early Agentforce use is that customers want more granular control over their agents. That matters because agents are not traditional software or human workers. They can reason, generate and act in ways that are useful, but also less predictable than a conventional application following fixed code.

That is where the distinction between probabilistic and deterministic work becomes important.

LLMs are powerful because they can generate flexible responses. That flexibility is useful for tasks that require judgment, variation or interpretation -- such as summarizing customer feedback, drafting outreach, identifying patterns in support tickets, surfacing themes from internal communications, interpreting sentiment or suggesting next-best actions.

But many tasks asked of enterprise systems depend on repeatable outcomes. Claims processing, payments, approvals, compliance actions, employee records, inventory reservations and financial close tasks cannot simply improvise their way to a different result each time. In those cases, the business needs consistent, auditable steps. A plus B still has to equal C.

So, when it comes to agents, there is a range. On one end are tasks where agents have room to interpret, summarize, recommend or improvise. On the other end are processes where there is little to no room for creative action and the agent must follow a fixed path.

Most enterprise work will fall somewhere in between.

Rules need business context

That makes boring rules more important than the AI sales pitch might suggest. But rules alone are not enough. Agents also need enough business context to understand what the rules mean.

Many software vendors -- which, by now, means pretty much every vendor in the enterprise software stack -- have woken up to the idea that AI needs more than model access and a few agent features. It needs traffic cops to prevent agents from conflicting with each other. It needs a common data foundation. It needs governance, security and compliance baked into the lower levels of the enterprise software stack.

One concept that straddles those needs is context.

The rules that govern agents cannot just be technical guardrails when AI operates across ERP, CRM, HR, customer service, collaboration and data systems. Once agents cross those boundaries, rules alone are not enough. An agent also needs to understand what the rule applies to, which system serves as the source of truth, what each business term means, and which policy or compliance limits apply.

That means an agent needs more than a login to the right system. It needs some understanding of how the business uses the information in that system.

A customer record is a good example. In one system, an active customer might mean a company with a current contract. In another, it might mean a company with recent buying activity, an open support case or an upcoming renewal. A supplier might be approved for one region but not another. A candidate in one system might already be an employee in another. A customer might appear ready for renewal in the CRM but still be on a compliance hold elsewhere.

If an agent does not understand those differences, it can follow a rule and still do the wrong thing.

Raw data access is not context. Giving an agent a database connection is not the same as telling it what the business does, which definitions matter and what policies should govern the action.

That is why context engineering is becoming a more important enterprise software discipline. As AI agents move across ERP, CRM, HR, customer service and collaboration systems, enterprises need a way to make business meaning available to the AI: definitions, relationships, source-of-truth logic, access policies, freshness signals and provenance. Otherwise, the agent might retrieve the right data and still apply it to the wrong business situation.

A rule only helps if the agent understands what it applies to.

AI still needs the old enterprise stack

That is why SAP's agentic AI pitch is revealing. SAP is promoting the idea of the autonomous enterprise, but it is also making the case that the underlying enterprise architecture still matters.

To SAP, ERP is the trusted system of record for running companies. Whether that is universally true or not, SAP makes a useful point when it comes to AI use, whether probabilistic or deterministic: AI requires structure and rules to function.

For SAP, ERP provides that structure through trusted records, business processes, knowledge graphs, identity rules and authorization controls that enable AI to operate across workflows.

Look beyond ERP, and the same idea applies across the wider enterprise software stack. The foundation includes governance, compliance, data security, identity, access controls, business definitions, workflow logic and systems of record across ERP, CRM, HR, customer service, collaboration and other platforms.

Any enterprise AI layer, agentic or otherwise, operates within that software environment. And as the lines blur between enterprise software silos, the work AI is asked to do will increasingly span and intersect those systems. That makes the common foundation more important, not less.

There is a counterintuitive point here: The more autonomous the AI interface becomes, the more important the underlying systems become.

Records, permissions, process logic, identity and authorization might seem old-fashioned and basic. But without them, AI cannot come close to doing what it is being asked to do now, let alone what it will be asked to do in the future.

The agent may become the new front door to enterprise work. But the underlying systems still define what the business knows, what the agent can access, what the user is allowed to see and which actions are compliant.

So, counterintuitively, the autonomous enterprise is not really all that autonomous. It just seems that way.

The most useful enterprise agents might not be the ones that act the most human. They may be the ones that know when not to improvise.

James Alan Miller is a veteran technology editor and writer who leads Informa TechTarget's Enterprise Software group. He oversees coverage of ERP & Supply Chain, HR Software, Customer Experience, Communications & Collaboration and End-User Computing topics.