AI workflow

An AI workflow is a set of steps you define in advance, with AI doing one or more of them. You decide the path when you build the system, rather than asking the model to choose the whole process while the work runs.

That gives you a written process you can test, a place for review, and a result you can trace back through the steps. You can still use an agent for one flexible decision while keeping the path around it fixed.

Across the 150+ franchise units in the delivery data I track, about 15% run workflows that start from a trigger. From my perspective, this is a useful middle step to understand before you decide you need an agent.

What an AI workflow looks like in delivery

Take a proposal process, where a call ends and the transcript arrives. The first step pulls out what the client asked for and the concerns they raised, and the second drafts a proposal using the firm’s pricing rules and past work.

The third step sends it to a named person for approval, and the fourth files the approved version and updates the CRM. In this example, two steps use AI and two move the work between systems, with a person checking it before it goes further.

The model changes what goes into the proposal, but it doesn’t choose a different path through those four steps. That’s what makes this a fixed workflow.

How it starts matters to the way your firm works too. A file arriving, a call ending, a schedule, or a status change can start the run, though a person can also start a workflow manually.

Starting on its own is part of the operating role I describe as an AI employee, not a requirement for software to count as a workflow. You still need to ask whether someone has to carry the work between steps.

Anthropic’s engineering guidance defines workflows by predefined code paths, while agents choose their process and tools as they go. The distinction is who controls the path, not whether someone clicked a button to start.

Workflow or agent: how do you choose?

Two questions settle the choice for most delivery work.

Are the steps the same every time? If you can draw the sequence on a whiteboard and it holds for the next hundred runs, start with a workflow. If covering normal cases would take twenty branches because the next action depends on what the last step found, an agent may fit better.

That’s the tradeoff in Anthropic’s guidance, predictability for a well-defined job versus flexibility when the model needs to choose.

Can you check the output, and can you undo it? Ask that before either build, because it matters more than how clever the model sounds. It’s the test in delegate the inputs, own the outputs.

If the work is hard to check and hard to reverse, keep a person at a named decision step, whatever a vendor calls the product.

A workflow that fails is often easier to investigate because you know the path and can look for the faulty step. An agent can reach the same wrong result through different paths, so finding the fault becomes part of its cost.

In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, citing cost, unclear value, and weak risk controls. That’s a forecast, not a measured cancellation rate.

Gartner also said many use cases called agentic don’t need agentic implementations. That supports checking whether a simpler workflow can do the job, without assuming it would have saved every canceled project.

Where firms get workflows wrong

They automate a process nobody wrote down. If the process exists only in your most experienced person’s head, the build has to discover those missing steps while client work is moving. Start by recording someone doing the job and use that as a first draft of the steps, then check it with the people who handle the exceptions.

They turn skilled people into approvers. A bad design puts the model at every production step and leaves the person approving or rejecting its work at the end. The process may move faster while the person’s job gets worse, which is the problem I describe in the output-validator workflow.

Put people where their judgment changes the outcome, often in scoping and client context, rather than only at the end with a yes-or-no choice.

They count workflows instead of results. You can build twelve workflows and still need the same amount of work to deliver something a client can use. The number of builds doesn’t tell you what changed for the business.

An April 2026 US Census Bureau working paper found that 66% of firms reporting an AI-related effect on worker tasks used AI only to help with existing tasks. A separate February 2026 NBER survey of nearly 6,000 senior executives at US, UK, German, and Australian firms found nine in ten reported no AI impact on employment or productivity over the prior three years.

Those are different surveys, and neither measures whether your workflow works. They give you a reason to check the business result rather than assume building something changed delivery.

They skip the workflow rung. An agent can look more impressive in a demo, but you still need to write down what the job requires. Across the 150+ franchise units I track, roughly 80% still run chatbots, about 15% run trigger-based workflows, and about 5% run autonomous agents defined by conditions.

If you’re moving from the first group toward the third, don’t skip the written process just because the model can choose its own steps.

What does it cost to keep one running?

The expensive part is maintenance, not the first version. Every workflow depends on the systems it touches. Someone renames a CRM field, the proposal step starts filing to the wrong place, and nobody notices until a client does.

Give each workflow a named owner and a check that alerts them when something fails. Monitor the final business record, because every software step can return a green status while the proposal still lands on the wrong account.

The other cost is people’s time to learn the new way while still delivering the old way. In the delivery data I’ve reviewed, training that sticks runs 8 to 10 hours per person over about 3 weeks, mixing video, in-person sessions, live Q&A, and technical support.

That’s an observation from that work, not a fixed requirement for every firm. But I’d make room for learning beyond one afternoon, because getting a workflow started and building a habit are different things, and the new process can get quietly worked around.

That’s why I’d start with one workflow, assign its owner, watch the result, and repair the weak steps before adding more. Six announcements at once divide the same people’s time between learning new work and maintaining it.

Where this sits on the Delivery Model Ladder

A rigid workflow that fires on a trigger is Stage 2, Augmented, on the Delivery Model Ladder. At this stage, the firm decides what delivery staff no longer do by hand, what the review checks, and who remains in the chain.

Using a chatbot to draft the same proposal faster is Stage 0 or Stage 1. The task moved faster while its shape stayed the same. The workflow is the jump to Stage 2 because the firm rebuilt a part of delivery around a defined path.

Stage 3, AI-native delivery, is where agents take defined slices of work and revenue stops tracking headcount in the old way. Firms reach that stage by compounding sound Stage 2 decisions. They do not get there by skipping the written process.

Quick answers

Can a workflow contain an agent? Yes. Many production systems use a fixed path and place an agent inside one narrow step with a defined output. That gives the work flexibility where it needs it and predictability everywhere else.

How is this different from ordinary automation? Ordinary automation follows rules someone wrote, often to move structured data between systems. An AI workflow adds at least one model step, for example reading a call transcript, email thread, or scanned document whose variation would be hard to handle with fixed rules alone.

What should a business build first? Start with the process your most experienced person runs often and describes the same way twice. Choose work that is high volume, repeatable, and checkable by someone else. Do not start with the process that would impress a client or move money.