In early October, Google introduced a "universal agent" that you give goals instead of instructions. It is impressive, and for many tasks it is the future. One detail from the same presentation interested me more. DBS Bank, one of the showcase customers, prepares its credit memos with fixed chains of 70 to 80 specialised agents. Each step has a clearly defined job, and the order is set in advance.
That matches what I see in projects. Most AI solutions that take work off people's plates in a company are fixed workflows with a few smart steps. That is exactly why they go live quickly.
Two ways to put AI to work
Anthropic, the maker of Claude, drew a clean line between the two back in late 2024:
- A workflow follows a path someone defined in advance. At individual points a language model does the work: it reads a document, classifies a request or writes a draft.
- An agent decides for itself which steps to take, in which order, and which tools to use.
After working with dozens of teams, Anthropic advises starting with the simplest solution and only adding complexity when it is needed. Workflows bring predictability for well-defined tasks. Agents bring flexibility when the path cannot be fixed in advance, and pay for it with more compute and higher cost.
An example: the order that arrives by email
Take incoming orders. An order arrives by email, as a PDF attachment.
As a workflow, it looks like this:
- A language model reads the PDF and extracts customer, items, quantities and delivery date.
- Code checks item numbers and prices against the ERP. That is an ordinary database query.
- If everything matches, the order is created. If something does not, the order goes to the sales office with a note.
Three steps, one of them with AI. The language model does exactly what was impossible a few years ago: understanding a document in any format. Everything else is code that works the same way every time.
As a free agent, the task would read "process incoming orders". The agent then decides for itself whether to search the ERP first or ask the customer first, and whether to pick the closest match for an unknown item number. That can work. But it is harder to predict, harder to measure and harder to explain when the sales office asks why the order looks the way it does.
Why the workflow goes live faster
You can measure it. Take 200 orders from last quarter, run the workflow over them and compare the result with what the sales office booked at the time. After an afternoon you know how often it gets it right.
Errors have an address. When something is wrong, you see which step it happened in. Usually it is reading one particular supplier's format, and that can be improved in a targeted way.
It is frugal. One reading step per order needs little compute. An agent that thinks, searches and thinks again needs many times more.
The team understands it. A workflow fits on one page. That helps with trust, and later with every change.
Where the agent shines
Agents play to their strengths where the path cannot be fixed in advance: research across many sources, programming, or questions about your own numbers where every question needs a different analysis. "Why was September revenue below last year?" has no fixed workflow. Here the agent finds the right data itself, calculates, checks and recalculates.
Often the best solution is a mix: a fixed workflow as the frame, and an agent at the point that cannot be predicted. The DBS chains are built that way too, as a fixed workflow in which specialised agents each handle one piece.
Three questions for your task
- Could you write the steps down on one page for a new colleague? Then it is a workflow.
- Does the next step depend on the result of the previous one in ways you cannot list in advance? Then that point needs an agent.
- How would you know the result is right? If you can answer that, you can measure it, workflow or agent.
The good news: most of the tasks that cost a business time every day fall under the first question. Nobody has to wait for the next model generation for those.
In an AI Pilot I take a task like this into production in four to eight weeks, measured against your own cases.
Sources
- Anthropic: Building effective agents, 19 December 2024
- Google Cloud: Gemini at Work 2026, 8 October 2026