Which Steps Actually Deserve AI Workflow Automation

Which Automation Pattern Does Your Process Need?

Ask two things about the task, not one. Is it well defined, which points at a workflow. And are flexibility and model-driven decision making needed at scale, which is when the guidance says an agent is the better option.

That settles more arguments than it looks like it should, because both halves are properties of your process rather than preferences about tools.

In my experience the step everyone complains about is not necessarily the one either half of that test picks out. Loudest and most automatable are different properties, which is why I do not start from the complaint.

The Five Patterns, And What Each One Suits

Quoting the conditions directly, because the exact wording is the useful part.

1. Prompt chaining. Ideal "for situations where the task can be easily and cleanly decomposed into fixed subtasks".
2. Routing. Works well "for complex tasks where there are distinct categories that are better handled separately".
3. Parallelisation. Effective "when the divided subtasks can be parallelized for speed, or when multiple perspectives or attempts are needed for higher confidence results".
4. Orchestrator workers. Well suited "for complex tasks where you can't predict the subtasks needed".
5. Evaluator optimiser. Particularly effective "when we have clear evaluation criteria, and when iterative refinement provides measurable value".

Read those as five different shapes of problem rather than five tools. Two of them turn on whether you can predict the subtasks, and they point opposite ways. Prompt chaining wants fixed subtasks you can name in advance. Orchestrator workers wants the case where you cannot predict them.

A Method I Use, Which The Source Does Not Propose

What follows is not in the guidance. Anthropic states when each pattern suits a task; it does not propose running the five as a selection procedure. This is my own habit, offered as one way to structure the conversation rather than as a method with authority behind it.

Take one process. Read the five conditions in order. For each, answer yes or no about that specific process, in writing.

Can the task be cleanly decomposed into fixed subtasks? Are there distinct categories better handled separately? Can subtasks be split for speed, or would multiple attempts raise confidence? Is it a case where you cannot predict the subtasks? Are there clear evaluation criteria, with iterative refinement providing measurable value?

Three things usually come out of that, in my experience.

Some processes satisfy one condition cleanly, and the shape of the build is then decided rather than debated. Some satisfy several, which usually means the process is more than one process and is worth splitting before anything is built. And some satisfy none, which is the most useful outcome of the three, because it is cheap and it arrives before the budget does.

On my reading, a process satisfying none of the five is not un-automatable. It means none of these named shapes fits, so anything you build is bespoke, and bespoke should be a decision rather than a surprise.

Context Is A Finite Resource

One constraint belongs in the selection conversation rather than the build, because it changes what is worth attempting.

Anthropic's context engineering guidance states that context must be treated as a finite resource with diminishing marginal returns. It describes context rot: as tokens increase, a model's accuracy in recalling information decreases, which it attributes to the transformer architecture's pairwise relationships being stretched across longer sequences.

My own reading is that a process needing a great deal of material held at once therefore carries a risk a process needing a little at the right moment does not, even where both look similar on a process map.

The same guidance names three techniques for long horizon work: compaction, meaning summarising and restarting; structured note taking, meaning maintaining external memory; and sub agent architectures, meaning delegating focused work to specialised agents. It favours managing context during execution over compartmentalising a single task upfront.

My own reading is that the presence of those techniques is a signal about scope. If a process only works with compaction and external memory, that is more machinery than a step which fits in one pass, and worth knowing before scoping.

When It Should Not Be An Agent At All

Worth stating plainly, because in my experience the word agent gets used for both.

The guidance draws the line at predictability. Workflows offer predictability and consistency for well-defined tasks. Agents are the better option when flexibility and model-driven decision-making are needed at scale.

So where a task is well defined, the guidance points at a workflow rather than an agent. If you have already decided you need an agent, the production concerns are covered separately in What Changes When An AI Agent Leaves The Demo For Production.

Example: Splitting An Invoice Process Across Patterns

This is a constructed example, not a client engagement, assembled from the conditions above.

An operations team handles inbound supplier invoices. Someone opens each one, reads it, decides which cost centre it belongs to, checks it against a purchase order, and either approves it or queries it with the supplier.

The complaint is the reading. It is dull and there is a lot of it.

Reading the five situations across it, as I would, the process comes apart into pieces with different answers.

Extracting fields from an invoice satisfies the prompt chaining condition. The subtasks are fixed and nameable: read the document, pull the fields, normalise them.

Deciding the cost centre satisfies the routing condition. There are distinct categories and they are better handled separately.

Checking against a purchase order satisfies none of the five. It is a lookup with a definite answer, which is a rules problem, not a model problem.

Querying a supplier is the one that moves money and contacts a customer, so it belongs behind a checkpoint regardless of which pattern it fits.

The useful result is not that the whole thing gets automated. It is that four steps have four different answers, and the purchase order check turns out not to need a model at all.

That decomposition is my judgement applied to a constructed case. The conditions are quoted; how you cut a given process against them is a call you have to make with the people who run it.

Signs A Process Is Ready For Automation

Five signals, drawn from processes I have looked at.

1. Your automation shortlist was written from complaints rather than from a process map.
2. Nobody can say which step in the process actually takes the time.
3. A step everyone calls complicated turns out to be a lookup with a definite answer.
4. The same process description gets different answers from two people who run it.
5. You are being quoted for an agent before anyone has established that the task is not well defined, or that flexibility is needed at scale.

Signal three is the one that saves the most money, because a rules problem solved with a model is expensive twice: once to build and again every time it runs.

If any of those apply, you can Request The Free Scoping Audit. The Agentic AI and What We Build sections are on the main site.

Common Questions About AI Workflow Automation

What Is AI Workflow Automation?

Anthropic distinguishes workflows, where models and tools are orchestrated through predefined code paths, from agents, where models dynamically direct their own processes and tool usage. Both put a model inside a process. The guidance points at a workflow for well-defined tasks, and at an agent where flexibility and model-driven decision making are needed at scale.

Which Process Should I Automate First?

Establish whether the task is well defined, and whether flexibility and model-driven decision making are needed at scale. Those are the two halves Anthropic's guidance turns on. It also states the situation each of five patterns suits. Matching a process to one of those is my own habit, not the guidance's.

What Are The Five Workflow Patterns?

Prompt chaining, routing, parallelisation, orchestrator workers and evaluator optimiser. Anthropic names each with the situation it suits: cleanly decomposed fixed subtasks, distinct categories handled separately, subtasks split for speed or confidence, unpredictable subtasks, and clear evaluation criteria with measurable iterative refinement.

When Is A Workflow Better Than An Agent?

Anthropic's guidance is direct: workflows offer predictability and consistency for well-defined tasks, whereas agents are the better option when flexibility and model-driven decision making are needed at scale. A well defined task with a known route points at a workflow on that basis, which is a predictability argument rather than a cost one.

Does The Size Of A Task Matter?

Yes. Anthropic states context must be treated as a finite resource with diminishing marginal returns, and describes context rot, where a model's recall accuracy decreases as tokens increase. That is worth weighing when a process would need a great deal held at once rather than a little at the right moment.

Part of The Agentic AI Guide. See also What Changes When An AI Agent Leaves The Demo For Production and The Build Or Buy Decision.