Do You Need an AI Agent or Workflow Automation?

01 /

Write the job in one sentence

Describe the input and the expected output without naming a tool. “When someone books a call, update the opportunity and create a preparation task” is a clear job. “Transform our sales with agents” is not yet a specification.

02 /

Identify the predictable steps

Matching a known identifier, checking a required field, assigning a task, and moving a record after a confirmed event are usually rule-shaped problems. Keep those rules explicit so people can inspect and test them.

03 /

Find the interpretation step

A free-text enquiry may need to be categorised. A long conversation may need a summary. A reply may need a draft based on approved facts. Those are candidates for bounded AI assistance, provided uncertainty has somewhere to go.

04 /

Give uncertainty an owner

Decide what happens when the input is ambiguous, the relevant document is missing, or two sources disagree. A human review queue is a valid output. Forcing a confident answer in every case creates a hidden failure path.

05 /

Define what the agent can touch

Reading an approved document is different from sending a customer message or changing a financial record. Give each action its own permission boundary. Treat incoming text as content to interpret, not instructions that can expand those permissions.

06 /

Evaluate completed work, not impressive replies

Test examples that resemble the real job, including failures and misleading input. Count tasks completed to your agreed standard, the corrections required, and the cost of operating the process. The useful system may combine simple rules, a small AI step, and a person.

The takeaway

Use predictable rules for predictable work. Give AI a bounded interpretation job and a clear fallback.

Explore the related service

Workflow, AI step, or agent

Choose the lowest level of
necessary intelligence.

There are three different implementation shapes hiding behind many AI proposals: deterministic workflow automation, an AI-assisted step inside a workflow, and an agent that can choose among tools or actions. Treating them as interchangeable creates unnecessary cost and risk.

The simplest design that satisfies the business need is usually easier to explain, evaluate, secure, and maintain. Complexity should be earned by a measurable improvement that cannot be achieved with stable rules or a reviewed AI output.

Use deterministic automation

Choose explicit rules when the trigger, required fields, conditions, and actions are predictable. Validation, calculations, routing, reminders, record updates, approvals, and scheduled reports often belong here. The system is easier to test because the same input should produce the same result.

Use an AI-assisted step

Add AI when the workflow must extract details from varied text, classify a request, retrieve relevant knowledge, summarise a conversation, or draft content. Keep the surrounding process explicit and give a person the source, proposed output, and ability to correct or reject it.

Use an agent

Consider an agent when the task genuinely requires multi-step judgment, choosing among approved tools, observing results, and adapting the next step. Define a narrow goal, limited permissions, maximum steps or spend, stopping rules, and escalation before granting action access.

Reject automation for now

Pause when nobody can define success, the source data is unreliable, the process changes every time, the decision is too consequential, the required access is inappropriate, or the expected value is smaller than the cost of building and supervising the system.

The agent decision test

Ask what decision the system makes, what evidence it may use, which tools it needs, what can go wrong, how the error is detected, who receives the escalation, and what the business gains from autonomous action rather than a reviewed draft.

Budget for evaluation and operation

Agent cost includes model calls, tool calls, retries, latency, logs, security controls, test cases, monitoring, human review, and behaviour changes after updates. Token pricing alone does not describe the operating cost.

Enough circling back.

Give busywork its notice.

One conversation. One workflow. A clearer way forward.

Let’s kill the busywork