Services Case Studies Impact About Insights Contact
Book a Discovery Call
Insights Enterprise AI Architecture
Series · Part 1 of 3 Published · June 2026 7 min read

AI Workflows vs AI Agents: Where to Set the Dial And where most enterprises set it wrong

The choice between a workflow and an agent is not primarily a technical decision. It drives your token cost, your reliability, your EU AI Act exposure, and whether your team can diagnose a failure at 2am on a Sunday. Get it right and the rest of the project becomes engineering. Get it wrong and no amount of prompt tuning will save you.

Part 1 of The Enterprise AI Architecture Series
Kamlesh Kshirsagar
Kamlesh Kshirsagar
Founder & Chief AI Architect, ProDataAI

Every enterprise AI conversation in 2026 starts with the same word: agent. But the real architectural decision is not whether to build one. It is how much autonomy your use case actually needs - and most organisations are buying more than they can govern or afford.

Core Thesis

The workflow-vs-agent choice is not binary. It is a dial. Turn it too far left and you ship rigid automation that breaks on the first edge case. Turn it too far right and you ship a system you cannot price, cannot debug, and cannot defend to a regulator. Most enterprise AI projects fail not because the model was wrong. They fail because the dial was set wrong.

1. The decision nobody frames correctly

In most boardrooms, the AI conversation skips straight from "impressive demo" to "let's build an agent." The middle step - deciding how much autonomy the use case actually requires - is missed entirely.

That middle step is the first real architecture decision. It drives your token bill, your reliability, your EU AI Act exposure, and whether your team can fix the system at 2 a.m. on a Sunday. Get it right and the rest of the project becomes engineering. Get it wrong and no amount of prompt tuning will save you.

The instinct to default to "agent" is understandable. The word carries the prestige of frontier research, the marketing weight of every major AI lab, and the implicit promise that the system will figure things out so the team does not have to. But prestige is not architecture. Architecture is the deliberate placement of a dial.

2. Workflows, agents and hybrids - in plain business terms

Anthropic's working definition - from their Building Effective Agents guide (2024) - is the cleanest available: "Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage."

In business language:

SystemWhat it isBusiness analogyWhen it wins
WorkflowA scripted sequence of steps. The LLM does a job at each station; the developer controls the route.A recipe followed by a competent cook.Predictable inputs, regulated outputs, high volume, thin margins.
AgentA goal-driven loop. The LLM decides which tools to call and when to stop.A chef given a kitchen and a brief.Open-ended tasks, novel inputs, where the path cannot be known in advance.
HybridA workflow that hands off to an agent for the one or two steps that genuinely require judgement.An assembly line with a specialist station.Most real enterprise use cases.

The honest answer for nearly every European enterprise we assess is the third row. What that hybrid actually looks like in production - the five patterns that ship in 95% of European enterprise AI work - is the subject of Part 2 of this series.

3. What "agent" actually means (and what it doesn't)

Before we discuss why the dial is mis-set, it is worth stating what a real agent is - because most of what is sold as one is not.

A real agent has three properties. It pursues a goal, not a script. It chooses which tools to call and in what order. It decides when it is done. Remove any one of these and you have something else.

In practice, four things get rebranded as "agents" in enterprise procurement and they should not be.

A chatbot is not an agent. A chatbot answers a question and stops. It does not pursue a goal across multiple steps. A retrieval-augmented chatbot that fetches one document and summarises it is a workflow with one LLM call. Calling it an agent is marketing.

RPA is not an agent. Robotic process automation follows a fixed sequence of UI actions. The path is hard-coded. Adding an LLM to extract data from a form does not change that - it is a workflow with an LLM step. Useful, often the right choice, but not an agent.

Retrieval is not an agent. A system that searches a vector database and returns relevant chunks is doing retrieval. Even if the retrieval is multi-hop, if the retrieval logic is coded by a developer, it is a workflow. Agency requires the model itself to decide what to retrieve next based on what it just found.

A workflow with a long prompt is not an agent. A 4,000-token system prompt that lists thirty tools does not make a workflow autonomous. If the code dispatches the next step, the code is in charge, not the model.

Gartner's analysts call this pattern "agent washing" and estimate that only a small fraction of the thousands of agentic vendors in the market are doing something that meets the definition. The implication for procurement teams: when a vendor says "agent," ask which of the three properties - goal, tool choice, termination - their system actually owns. If the answer is none, the price tag should reflect a workflow, not an agent.

This matters because the cost, governance and reliability arguments below all change shape depending on which one you are actually buying.

4. Why the dial is usually set too far toward "agent"

Three forces push organisations rightward before the business case justifies it.

The cost reframing

Cheaper tokens do not mean cheaper AI. Production data shows agents use roughly 4× more tokens than a single LLM call, and multi-agent systems around 15× more. At enterprise volume this is a budget-line decision, not a technical footnote. A workflow that costs €0.005 per task at pilot becomes €0.005 per task in production. An agent that costs €0.05 at pilot can cost €0.20-€0.50 at production scale, because production inputs are messier and the loops run longer.

The reliability reframing

A workflow is deterministic - same input, same output, traceable when it breaks. An agent is non-deterministic - it can solve the same problem two different ways. That is powerful for research and triage. It is dangerous for claims processing, credit decisions, or anything a regulator can ask you to reproduce. Non-determinism is not a bug - it is a feature you may not want. A 2025 security survey of LLM-based agents (arXiv:2506.23844) formalises this as a risk gradient: L1–L2 systems can rely on static safeguards; L3 and above require adaptive oversight - which is precisely what EU AI Act Article 14 mandates.

The market reframing

Gartner research on agentic AI projects warns of widespread project abandonment, citing "escalating costs, unclear business value or inadequate risk controls." In the deployments we assess, the failure is rarely the model. It is autonomy granted where it was never needed.

The failure is rarely the model. It is autonomy granted where it was never needed.

5. The CLEAR Autonomy Test

The CLEAR test is our own synthesis. The five-position dial is grounded in a converging body of published frameworks: the Cornell/Knight First Amendment Institute's user-role taxonomy (Archipelago et al., arXiv:2506.12469, 2026), the ASDLC L1–L5 autonomy scale (2026), and Anthropic's Building Effective Agents (2024). The "levels" metaphor originates with SAE J3016 driving automation - a lineage shared by all the frameworks cited below.

Before you build, run the use case through five questions. The more "yes" answers, the further right on the dial you belong. Anything else defaults to a workflow or a hybrid with a human checkpoint.

C
Can the path be scripted?

If a competent engineer can draw the decision tree on a whiteboard, build a workflow.

L
Latency and cost tolerance?

High volume, thin margins, sub-second response - stay left.

E
Exceptions and variability?

If edge cases dominate and inputs are unpredictable, autonomy starts to earn its keep.

A
Accountability and audit needs?

Regulated decisions favour deterministic, auditable workflows. The EU AI Act, Article 14, requires high-risk systems to be designed so they "can be effectively overseen by natural persons," with oversight "commensurate with the level of autonomy." Interface EU's autonomy classification (2025) maps this to liability: above L3, liability shifts progressively from deployer to developer.

R
Reversibility of actions?

If a mistake is expensive or hard to undo, constrain autonomy and add a human-in-the-loop.

The dial that follows from the test has five positions, not two:

The Autonomy Dial: five positions from Workflow through Workflow + LLM step, Hybrid + human-in-the-loop (sweet spot), Supervised agent, to Autonomous agent

Most enterprise systems we ship sit in positions two and three - what the ASDLC scale terms L2–L3 (Conditional Autonomy) and what the Cornell/Knight Institute framework describes as the Collaborator-to-Consultant range. Position five - a fully autonomous agent acting on consequential decisions without human review - is rare in regulated European environments and, in our view, should be.

The one-sentence test: if you can write down every step in advance, build a workflow. If the path cannot be known until the task is underway, that is where an agent belongs - and even then, put a human at the checkpoint.

6. What this looks like in European enterprises

Three short examples, each showing a different point on the dial.

A claims-triage system for a DACH insurer

90% of the work is a workflow - intake, classification, fraud checks, payout calculation - with a single LLM call to summarise free-text damage descriptions. An agent here would burn tokens for no measurable gain and complicate the BaFin audit trail.

Architecture diagram: claims-triage workflow with LLM summarisation step

A supplier-onboarding system for a Mittelstand automotive group

Multi-format documents arrive in five languages from hundreds of suppliers. A workflow handles intake and OCR; a routing step dispatches to specialised processing; a human approves anything that fails policy checks. The Betriebsrat signed off because the human checkpoint is real, not theatre.

Architecture diagram: supplier-onboarding hybrid with human-in-the-loop approval

A market-research assistant for the same group entering a new vertical

The path genuinely cannot be scripted - the next search depends on what the last one found. A supervised agent, with a human approving the final synthesis before it informs any commercial decision, earns its cost.

Architecture diagram: supervised agent with human synthesis approval for market research

Same company. Three use cases. Three settings on the dial. That is what deliberate placement looks like.

7. Three questions to ask your AI team this week

  1. For our current agent project, could 80% of it be a workflow with one or two LLM calls? What would we lose?
  2. What is the token cost per task today, and what does it become at 10× volume?
  3. If a regulator asked us to reproduce the system's decision from last Tuesday, could we?

Key takeaways

Workflows and agents are not rivals. They are points on an autonomy dial.

Most of what is sold as an "agent" is a chatbot, RPA, or a workflow with a long prompt. Ask which of goal, tool choice, and termination the system actually owns.

Start as far left as the use case allows. Add autonomy only when the business case demands it.

Cost, reliability, and EU AI Act oversight all favour the middle of the dial for most enterprise work.

The failure mode is not picking the wrong technology. It is not deciding deliberately.

Not sure where your use case sits on the autonomy dial? We help European enterprises place the dial deliberately.

Book a 30-min call →

References

  1. Archipelago et al. - Levels of Autonomy for AI Agents. Cornell / Knight First Amendment Institute, arXiv:2506.12469 (2026).
  2. ASDLC - L1–L5 AI Agent Autonomy Scale (2026).
  3. Zhihao et al. - A Survey on Autonomy-Induced Security Risks in Large Model-Based Agents. arXiv:2506.23844 (2025).
  4. Interface EU - An Autonomy-Based Classification of AI Agents (2025).
  5. Cloud Security Alliance - Levels of Autonomy for AI Agents (2026).
  6. Anthropic - Building Effective Agents (2024).
  7. SAE International - J3016: Taxonomy and Definitions for Terms Related to Driving Automation Systems (2021). The origin of the "levels" metaphor borrowed by all frameworks above.
The Enterprise AI Architecture Series
Part 1 · You are here
AI Workflows vs AI Agents: Where to Set the Dial
The CLEAR test, agent washing, the cost reframe, and why most enterprises set the dial too far right. · 7 min read
Part 2 · Coming soon
5 Workflow Patterns That Solve 95% of Enterprise AI Problems
The production-proven patterns - prompt chains, routing, parallelisation, orchestrator-subagent, and evaluator loops.
Part 3 · Coming soon
When You Genuinely Need an Agent: A Production Readiness Checklist
The 12 questions to answer before granting autonomous decision-making to any system.

Not sure if your AI project is a workflow, an agent, or a hybrid?

We work with DACH and European enterprises to make this decision before the architecture is locked - building workflows where reliability matters and agents only where autonomy earns its cost. A 30-minute conversation costs nothing. If we are not the right fit, we will say so.