Every technology wave creates the same temptation: use the new thing for jobs the previous tool already handled well.

AI agents are no exception.

A team looks at a process and asks: “could we put an agent here?”

Invert the question: which part of this work contains enough uncertainty to justify an agent?

RevOpsHubs thesis

The more deterministic the decision, the fewer reasons there are to use an agent. Autonomy creates value when the work requires interpretation, path selection or coordination across tools.

Four execution modes

Most Revenue Operations work can be decomposed into four modes.

ModeBest fit
WorkflowThe rule is known, the input is structured and the action is predictable.
AI action inside a workflowThe flow is deterministic but one step must classify, summarise or interpret unstructured input.
AgentThe system must choose steps, gather context, select tools or adapt the path.
Human approvalThe impact is high, the action is hard to reverse or it creates an external commitment.

These are not maturity levels. They are different execution mechanisms.

The Decision Uncertainty Matrix

Use two axes: decision uncertainty and action impact.

QuadrantDefault
Low uncertainty · low impactWorkflow
Low uncertainty · high impactWorkflow + validation/approval
High uncertainty · low impactAgent with bounded autonomy
High uncertainty · high impactAgent-assisted human decision

What does “uncertainty” mean here?

It is not uncertainty about whether the system will work. It is uncertainty about which decision should be made from the available context.

Low

The rule fits in one sentence

“If the deal is Closed Won, create an onboarding task.” “If the form comes from Spain, assign it to the ES team.”

Medium

The path is known but the input is ambiguous

“Classify this email reply.” “Summarise the call and extract next steps.” An AI action inside a workflow may be enough.

High

The system must decide the next step

“Research this account, decide which sources to inspect and choose what information matters for the rep.” That is multi-step reasoning and tool selection.

Revenue Systems Brief

One useful idea. One system to improve.

A concise weekly field note for people responsible for revenue, operations and growth.

Six B2B examples

  1. Route a lead by country and segment. Workflow.
  2. Classify an email reply as interest, timing, objection or unsubscribe. AI action inside the workflow.
  3. Research an account before a meeting and choose which sources to inspect. Agent.
  4. Update Lifecycle Stage because an observable event occurred. Workflow.
  5. Prepare a personalised follow-up after a meeting. AI action or agent depending on how many sources and decisions are involved. Sending automatically is a separate risk decision.
  6. Delete, merge or bulk-change critical records. Human approval even if an agent prepares the action.

Why an agent can be the worse architecture

More autonomy creates more operating surface: more context, more tools, more possible paths and more behaviour to evaluate.

HubSpot separates workflows, which execute actions on enrolled records, from custom agents that can analyse data, generate output and take actions using instructions, inputs and tools. OpenAI recommends approvals for sensitive tool calls and trace evaluation to verify whether agents chose the right tool and path.

There is also an economic dimension. Some AI and agent actions consume credits. A powerful agent can still be an unnecessarily expensive architecture for a binary rule that a workflow executes predictably.

Do not use reasoning where a rule is enough. Do not force a rule where the work needs judgement.

The four-question test

  1. Is the rule explicit and stable? Start with a workflow.
  2. Is the only uncertainty unstructured input? Try an AI action inside the workflow.
  3. Does execution require choosing steps, tools or sources? That is a real agent case.
  4. Is the action external, hard to reverse or materially risky? Add human approval regardless of the mechanism.

One prerequisite: stable meaning

Even the right execution mechanism fails if the states feeding the decision are ambiguous.

That is why Semantic Debt comes before this decision. A workflow codifies a rule. An agent interprets context. Both depend on sufficiently stable semantics.

If you choose an agent, the next step is to define its limits in an Agent Operating Contract.

What to do Monday morning

Pick one process where someone has suggested “adding AI”. Break it into decisions. For each decision, mark:

  • explicit rule or judgement?
  • structured or unstructured input?
  • one step or adaptive multi-step path?
  • reversible or difficult-to-reverse action?
  • internal or external impact?

You may find that the process does not need “an agent”. It needs a workflow, one or two AI actions, perhaps an agent in one narrow zone and human approval at the end.

RevOpsHubs principle

The best architecture does not maximise AI. It minimises unnecessary uncertainty and reserves autonomy for decisions where it creates value.

Public sources

Keep going

Semantic Debt →Agent Operating Contract →Shadow Agents →HubSpot Credits economics →