Agentic AI or Automation? How to Tell What Your Workflow Actually Needs

For most of the last decade, automating a workflow meant writing down the rules. If the invoice matches the purchase order, approve it. If the ticket mentions a refund, send it to bill. Workflow engines and robotic process automation did this reliably, and nobody called it intelligent.

The agents changed the pitch. Almost every automation proposal now arrives labeled agentic, promising a system that plans, decides and acts on its own. Some workflows need that. Most do not, and the ones that don't pay for autonomy in cost, speed and risk without getting much back.

Use the least autonomous system that does the job.

This piece is part of our guide to why AI pilots stall before production. Choosing the wrong level of autonomy is one of the quieter reasons they do. Let's cover how to tell rules from judgment from genuine multi-step reasoning, how to break a workflow down step by step, and what to ask before anyone builds an agent.

The Agent Label Tells You Very Little

The market has made the word hard to trust. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 as costs climb, value stays unclear or risk controls fall short. It also warns of agent washing: vendors rebranding assistants, chatbots and robotic process automation as agentic AI solutions without adding real agentic capability. By Gartner's estimate, only about 130 of the thousands of vendors making agentic claims are the real thing.

The Gartner analyst behind the forecast put the core problem in one line: "Many use cases positioned as agentic today don't require agentic implementations."

Intent is running well ahead of readiness. In its 2026 Hype Cycle for Agentic AI, Gartner places agentic AI at the Peak of Inflated Expectations. Its 2026 CIO survey found that only 17% of organizations have deployed AI agents so far, while more than 60% expect to within two years. Gartner adds that fully autonomous agents are not yet ready for most enterprise use cases.

So the real agentic AI vs automation question is not whether a system is agentic. It is how much autonomy each step of the work needs.

Three Kinds of Work, Three Kinds of System

Most enterprise workflows are made of three kinds of work. Each has a natural fit, and the fit is decided by the work, not by what is newest.

Gartner's own starting advice is similar: give routine workflows to automation, simple lookups to assistants, and save agents for work that needs decisions. Anthropic's guidance on building agents draws the same line between workflows, where the steps are fixed in code, and agents, where the model decides its own steps. Its advice is to start simple and add complexity only when results demand it, which can mean not building an agent at all.

The middle row is the one most often skipped. Many proposals sold as agentic are really a judgment step inside a path that never changes. One well-scoped model call, with the rest of the workflow in ordinary code, is what the market calls intelligent process automation. It delivers most of the value at a fraction of the cost and risk.

Decide Step by Step, Not Workflow by Workflow

Workflows are rarely one kind of work. The common mistake is to pick one system for the whole thing, either forcing rules onto the parts that need judgment or handing an agent the parts that follow rules.

Take supplier invoice exceptions, a frequent candidate for an agentic rebuild. Broken into steps, it looks like this.
Six steps. One is agent-shaped, and one stays with a person whatever the technology can do. That is precision in practice. Automation is cheaper and auditable where rules apply, AI does the reading and drafting where judgment is needed, and the agent's cost and risk are confined to the one step that pays for them.

Five questions sort almost any step.

1. Can a Subject Expert Write the Rule Down? If yes, code it. It will be faster, cheaper and easier to audit than any model.
2. Is the Input Messy but the Decision Single? Use a model for that step and keep a fixed path around it.
3. Does the Next Action Depend on What the Last One Found? Only then does an agent's ability to plan earn its place.
4. What Does a Wrong Answer Cost? The higher the cost of an error, the less autonomy and the more review the step should get, however capable the model.
5. Can the Result Be Checked? Agents work best where progress can be verified: a test passes, a record reconciles, a case closes. Without a check, autonomy is a guess.

What Autonomy Costs

Autonomy is not free, and the bill arrives in more than one place.

1. Usage Cost: Agents call a model many times per task. In Anthropic's own data, agents used about four times as many tokens as chat interactions, and multi-agent systems about fifteen times as many. At workflow volumes, that multiplier is the business case.
2. Speed: Every planning step and tool call adds time. A rule returns almost instantly; an agent loop can take minutes.
3. Predictability: The same input can take a different path on a different day. That is useful for open-ended work and a problem for anything an auditor or regulator will ask about.
4. Compounding Errors: A mistake early in a multi-step run is carried into every step after it. Short, checked steps contain errors; long, unchecked chains spread them.
5. Governance: Every tool an agent can use is a permission to define, monitor and revoke. Gartner's 2026 Hype Cycle tracks agentic AI governance, security and FinOps for agentic AI as disciplines of their own, and says the controls for risk, trust and cost are still catching up.

None of this argues against agents. It argues for paying for autonomy only where it returns more than it costs, and for measuring cost per completed task rather than cost per call. Building each layer to production standard, from rules to model steps to agents, is the focus of our AI engineering work.

When an Agent Is the Right Call, Build It Narrow

Some work genuinely needs an agent: incident investigation, research across many sources, or cases that cross several systems where the next move depends on what turns up. When it does, the design choices decide whether it reaches production. Most of the effort in AI agent development goes into these controls, not the model.

1. Scope It Tightly: One job, a short list of tools, and read-only access wherever writing isn't required.
2. Surround It With Deterministic Steps: Validate inputs before the agent sees them, and check outputs against rules before anything is committed.
3. Put People at the Points of Consequence: Payments, customer messages and changes to records go to a person for approval, with the agent's reasoning shown alongside.
4. Set Budgets: Cap steps, time and spend per task, so a confused run stops instead of looping.
5. Log Every Action: Record each tool call and decision, so any outcome can be traced and explained.
6. Earn Autonomy With Evidence: Start with a person reviewing every output, then move to spot checks once the error rate is known.

On The Pivot, each step gets its system in the Define stage, and autonomy is revisited in Evolve, raised only where error rates and costs are proven. Once a system is live, cost per outcome, error rates and drift need watching like any production system. That is the work of AI operations.

Six Questions to Ask Before You Approve an Agent

These work for a proposal built in-house or one pitched by a vendor. Each is far cheaper to answer before the build than after it.

1. Which Steps Follow Rules, and Which Need Judgment? If nobody has broken the workflow down, the proposal is a guess.
2. Could One Model Call Inside a Fixed Workflow Do This? If so, the agent adds cost, not capability.
3. What Does One Run Cost, and What Will It Cost at Full Volume? Price the multiplier before the pilot, not after the first invoice.
4. What Happens When It Is Wrong, and Who Notices? Name the check and the owner.
5. Which Actions Can It Take Without a Person? And which need approval, with the reasoning shown?
6. How Will We Know It Is Still Worth It in Six Months? Agree the measure, such as cost per resolved case, before launch.

If the answers point to rules, build automation, and keep the budget for the step that needs more.

Choose Precision Over Autonomy

The goal is not the most capable system. It is the right system for each step. In most workflows that means mostly rules, a few model steps, and perhaps one agent where the work truly calls for it. That is not a smaller ambition. It is the version that reaches production and stays there.

If you have a workflow on your agentic shortlist, start with a workflow assessment. You get every step mapped to the system it needs, with a run-cost estimate for each. Book a workflow assessment or see how we build AI for production.