AI Agents, Workflow Automation, or RPA?

The wrong question

Automation programs are being asked to choose between three things that are often presented as generations: robotic process automation (RPA) as yesterday, workflow automation as today, and AI agents as tomorrow. Vendors encourage the framing. In June 2025 Gartner coined a term for the result, "agent washing": the rebranding of existing assistants, RPA and chatbots as agentic AI without substantial agentic capability. Gartner estimated that only about 130 of the thousands of vendors claiming agentic products offered real agentic features, and predicted that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.

The three are not generations. They are different machines for different parts of a process:

• Workflow automation (BPM engines, orchestration platforms, integration flows) executes a defined sequence of steps, rules and handoffs. The same input produces the same path every time.
• RPA drives user interfaces the way a person would: clicking, typing and copying between screens. It is an integration technique for systems that have no usable API.
• AI agents use a language model to decide which steps to take and which tools to call, adapting the path to the input.

The thesis of this article: the unit of decision is the step, not the process. Use deterministic workflow for the spine of a process, RPA as an adapter of last resort, and agents only at the steps where judgment over unstructured or variable input creates value. In most enterprise processes, that means agents working inside workflows, not workflows rebuilt as agents.

What each engine is good at

The distinction that matters most is who decides the path. In a workflow, the designer decides it in advance, which is why it can be tested, audited and explained. In an agent, the model decides it at run time, which is why it can handle cases nobody anticipated, and why it can also do things nobody anticipated.

Anthropic's engineering guidance draws the same line. It defines workflows as systems where models and tools follow predefined code paths, and agents as systems where the model directs its own process and tool use. Its advice is to find the simplest solution that works, and to add agentic complexity only when it demonstrably improves results, because agents trade latency and cost for flexibility and their errors can compound.

RPA belongs in a different conversation. It does not decide anything. Its weakness is fragility: a changed screen, field or login flow breaks the bot. Its strength is that it can reach systems nothing else can. That makes it an adapter, and the right long-term question for each bot is whether an API or integration can replace it.

What the reliability evidence says

The case for agents rests on capability; the case against rests on reliability. Independent benchmarks of realistic office work give a sober picture.

TheAgentCompany, a benchmark from Carnegie Mellon University and collaborators, simulates a small software company with 175 tasks across engineering, finance, HR and administration. In the latest version of the paper (September 2025), the best agent completed 30.3% of tasks fully, with a partial-credit score of 39.3%, at an average cost of about US$4.20 per task. CRMArena-Pro, from Salesforce AI Research (May 2025), found leading agents succeeded on about 58% of single-turn business tasks and about 35% in multi-turn settings, with near-zero awareness of confidentiality unless specifically prompted. Workflow execution was the exception, above 83% for the best model.

Consistency is a separate problem from capability. Sierra's τ-bench (June 2024) introduced the pass^k metric: the probability that an agent succeeds on the same task in all of k attempts. In the original results, a leading function-calling agent succeeded on fewer than half the tasks, and its pass^8 in the retail domain was below 25%. An agent that solves a case once may not solve the same case the same way next time.

Three caveats apply. These benchmarks measure models from 2024 and 2025, and newer models score higher. Benchmark tasks are harder and less structured than many production steps. And the vendors building agent products will report better numbers on their own tasks. But the pattern across independent studies is consistent: agents are now useful on bounded steps, and still unreliable on long, multi-step processes run end to end.

Why long processes punish agents

The reliability gap compounds with process length. If each step an agent takes is right 95% of the time and errors are independent, a 10-step process completes correctly about 60% of the time, and a 20-step process about 36% of the time.

This is the arithmetic behind the thesis. A deterministic workflow does not eliminate errors, but its errors are fixed and testable: the same defect appears in the same place every time and can be found and fixed once. An agent's errors vary from run to run. The shorter and more bounded the agent's part of the process, the less that variability costs.

Cost and auditability

Run cost. A workflow step costs close to nothing to execute. An agent step consumes model tokens on every run, and the cost grows with the number of reasoning and tool-use turns. TheAgentCompany's figures, from about US$0.30 to US$6.80 per task across the models tested, are a reminder that agent costs per transaction can rival the cost of the human work they replace when the task is long. Measure cost per completed case, including retries and human rework, not cost per call. The value tests in From AI Experimentation to Enterprise Value apply to agents as much as to any other AI system.

Change cost. Workflows are expensive to change when the process itself changes often or has many variants. That is where agents earn their keep: one agent step can absorb variation that would otherwise need dozens of branches.

Auditability. A workflow can be explained from its definition. An agent's behavior can only be explained from its traces. In regulated processes, such as credit, claims, benefits, or anything with a statutory decision, that difference often decides the design: the decision itself stays deterministic or human, and the agent prepares, extracts and recommends.

Exit cost. Gartner added a further warning in September 2026, predicting that by 2028, 70% of enterprises will abandon agentic AI built for them by vendor forward-deployed engineers, trapped by rising costs and unable to evolve it themselves. Agent logic that lives in prompts and vendor-managed configuration is harder to own than a workflow definition in your own repository.

The hybrid pattern: agents as steps

The pattern that holds up in production is a deterministic workflow that calls agents for bounded tasks and checks their output before acting on it.

The workflow owns the sequence, the state, the business rules, the approvals and the audit trail. The agent is invoked for a specific job, such as reading a free-text email and classifying intent, extracting fields from a non-standard document, drafting a response, or reconciling records that do not match exactly, and returns structured output. The workflow validates that output against a schema and business rules and routes low-confidence results to a person. RPA, where it remains, is a connector the workflow calls, not a process owner.

This inverts the "agentic orchestration" pitch, where an agent sits on top and calls deterministic tools. That inversion is right for genuinely open-ended work, such as research, investigation or software engineering (see our comparison of coding agents), where the path cannot be known in advance. For a known business process with exceptions, the process should own the agent, not the other way round. The guardrail architecture in our companion article, Designing Enterprise AI Agents With Guardrails, applies in both cases.

Klarna's experience is a useful reference point. In February 2024 the company reported that its AI assistant handled two-thirds of customer service chats in its first month, the equivalent of 700 full-time agents, while customers could still choose a human (company-reported figures). By May 2025 its chief executive was telling Bloomberg that cost had weighed too heavily in the design, quality had suffered, and the company would make sure a human was always available. The lesson is not that the assistant failed. It is that the boundary between automated handling and human handling is a design decision that needs to be measured and moved, not set once.

A decision table by step

Apply this per step, not per process. Most processes will end up with all three.

Before choosing an agent, try the simpler option first: a single model call inside a workflow step often delivers most of the value with a fraction of the variability. For prioritizing which processes to automate at all, see How to Select Enterprise AI Use Cases.

What to do in the next 90 days

1. Map three candidate processes step by step and label each step as rules, integration, judgment over unstructured input, or regulated decision.
2. Audit your RPA estate for bots that break often or duplicate an available API; these are integration debt, not automation assets.
3. Pilot one hybrid process in which a workflow calls a single agent step with schema-validated output and a confidence-based route to a person.
4. Measure reliability with repeats. Run each agent step on the same test cases several times and report consistency, not only average accuracy.
5. Measure cost per completed case, including model spend, retries and human rework, and compare it with the current process.
6. Ask every "agentic" vendor to show which decisions the model makes at run time, how behavior is traced, and what happens when the model is wrong. Treat vague answers as agent washing.
7. Keep ownership. Hold workflow definitions, prompts, evaluation sets and traces in your own repositories, whoever builds the first version.

Frequently asked questions

Will AI agents replace RPA?

Partly. Agents can operate interfaces and absorb variation that breaks RPA bots, but they are slower, costlier per run and less predictable. For stable, high-volume tasks, APIs and workflows remain better. Many RPA bots should be replaced by integrations, not by agents.

When should I use an AI agent instead of workflow automation?

Use an agent for steps that require judgment over unstructured or highly variable input, or where the path cannot be designed in advance. Use workflow automation for the overall process, its rules, state, approvals and audit trail. Most good designs combine both.

How reliable are AI agents for business processes?

Improving but uneven. In independent 2025 benchmarks, the best agents fully completed about 30% of realistic office tasks in TheAgentCompany and about 35% of multi-turn business tasks in CRMArena-Pro. Reliability is much higher on short, bounded steps than on long processes run end to end.

What is agent washing?

Gartner's term for vendors rebranding existing products, such as AI assistants, RPA and chatbots, as agentic AI without substantial agentic capability. In June 2025 Gartner estimated only about 130 of thousands of vendors claiming agentic AI offered genuine agentic features.

Sources and further reading

• Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027, Gartner, 25 June 2025; vendor estimate as reported by MarTech.
• Gartner on vendor forward-deployed engineering, Gartner, 29 September 2026.
• TheAgentCompany, arXiv, version 3, 10 September 2025.
• τ-bench, arXiv, 17 June 2024, and Benchmarking AI agents for the real world, Sierra.
• CRMArena-Pro, Salesforce AI Research, arXiv, 24 May 2025.
• Building effective agents, Anthropic, 19 December 2024.
• Klarna AI assistant handles two-thirds of customer service chats, Klarna, 27 February 2024 (company material); Klarna CEO reverses course, Entrepreneur, 9 May 2025.