Why working workflows are often already aging — and what to do about it
There is a piece of received wisdom that governs most of how organizations think about their deployed systems: if it isn’t broken, don’t fix it. Leave well enough alone. Stability is the goal. The workflow is running, the metrics are acceptable, the customers aren’t complaining loudly enough to trigger an incident review. Move on to the next project.
This is a reasonable philosophy for a static technology environment. It is the wrong philosophy for AI.
Here is the problem in plain terms. An AI workflow deployed today is built around a specific model, a specific set of tool integrations, a specific prompt architecture, and a specific understanding of how customers phrase their needs. Six months from now, every one of those things may have changed. The underlying model may have been updated — sometimes significantly — on a schedule controlled by the model provider, not by the organization that deployed the workflow. New capabilities may have become available. Customer language patterns shift. The business itself evolves. Staff changes. Processes that the workflow depended on quietly get modified by operations teams who didn’t know the AI was depending on them.
A workflow that was the best available design at launch can, without ever “breaking,” drift to a point where it is delivering meaningfully worse outcomes than a freshly designed workflow would. Not broken. Just aging. Slowly, invisibly, while the metrics dashboard shows acceptable performance and nobody is looking at the question nobody is asking.
What maintenance misses
Most organizations that have deployed AI workflows have some version of maintenance in place. Someone is watching error rates. Someone gets paged when a tool call fails. There’s a process for rolling back a prompt change that breaks something. This is all legitimate and necessary.
But maintenance is reactive. It responds to events that have already occurred. A broken tool call, a spiking error rate, a customer complaint that escalated — these are the signals maintenance is listening for. And by the time those signals appear, the damage has already happened.
More importantly, maintenance has no mechanism for asking the question that matters most in a changing AI environment: is this still the best way to do this? Maintenance can tell you whether the workflow is functioning. It cannot tell you whether the workflow is still the right workflow.
That question requires a different orientation — one that I think of as stewardship, to distinguish it from the more reactive mode of maintenance.
A steward does not wait for something to break. A steward asks, on a regular schedule, whether the workflow still serves the purpose it was built to serve, whether that purpose is being served as well as it could be given what is now available, and whether the assumptions the workflow was built on are still valid.
The assumptions buried in every workflow
Every AI workflow contains a set of embedded assumptions that are rarely written down and almost never reviewed.
There is an assumption about what the customer will say and how they will say it. There is an assumption about which model capabilities are worth relying on and which aren’t stable enough yet. There is an assumption about which integrations exist and how reliable they are. There is an assumption about how staff will respond to escalations and notifications. There is an assumption, often implicit, about what the business process looks like on the other side of the AI — who receives the handoff, what they do with it, how long it takes.
When the workflow was designed, these assumptions were reasonable. They were grounded in current reality. But current reality is not static.
The model provider releases a new version with significantly improved reasoning. The assumption that certain multi-step logic had to be handled outside the model may no longer be true. A new tool architecture becomes available that could eliminate three steps from the current workflow. The business adds a new product line, but nobody updates the AI’s knowledge of what the company offers. A key staff member who was the de facto authority for a particular escalation path leaves the company, and the escalation now goes to a queue nobody is monitoring.
None of these changes “breaks” the workflow in the traditional sense. The error rate stays flat. The tool calls complete. The AI keeps talking. But the outcomes — the actual resolved conditions on the other side — quietly deteriorate.
What a stewardship review looks like in practice
A stewardship review is not a full redesign. It is a structured conversation with a specific question at its center: if we designed this workflow today, from scratch, knowing what we now know, would we design it the same way?
That question, asked honestly, tends to surface the things that maintenance misses. The tool that was state-of-the-art at launch but has since been superseded. The routing path that made sense when the team was structured a certain way but no longer reflects how decisions actually get made. The prompt instruction that was added to work around a model limitation that no longer exists. The status message that was accurate when it was written but now describes a process step that operations quietly changed six months ago.
A stewardship review should happen on a regular cadence — quarterly is a reasonable starting point for most production AI workflows — and also after any significant event: a model upgrade from the provider, a meaningful change to the underlying business process, or any incident where the AI produced an outcome that surprised the team.
The goal of the review is not to find something wrong. It is to make a deliberate, current-state decision: keep the workflow as designed, because it still best serves the purpose; adjust specific components because better options now exist; or redesign a section because the underlying assumptions are no longer valid.
Any of those three conclusions is a good outcome. The bad outcome is the one that happens when nobody asks: the workflow keeps running, the assumptions keep aging, and the gap between what the AI is delivering and what it could deliver keeps growing — invisibly, acceptably, until it isn’t.
The deeper shift this requires
There is a disposition underneath this that is worth naming directly, because it runs against a natural instinct.
Designers get attached to their designs. This is understandable. A well-designed workflow represents real work — research, analysis, negotiation, careful prompt construction, edge case consideration, testing. Walking back into that workflow and asking whether it should be redesigned can feel like questioning the original work. It can feel disloyal to the effort that went into it.
The shift is to be more committed to the outcome than to the mechanism. The outcome — the customer whose problem was genuinely resolved, the staff member who got a useful escalation at the right moment, the business that retained a customer it might otherwise have lost — that is the thing the workflow was built to produce. The workflow itself is the mechanism, and mechanisms can be improved.
A steward who is genuinely committed to the outcome will ask the redesign question without ego and will welcome a better answer even when that answer makes the previous design look limited. That is not a failure. That is what a design discipline that takes its own purpose seriously looks like.
Since February 5, 2026, I’ve had the opportunity to study feedback from more than 18,000 complicated AI integrations running in production across 43 projects.
That work has taught me something important: AI workflow design is not just automation. It is a new discipline.
Humans do not arrive as clean tickets, categories, forms, or process steps. They arrive unfinished. They speak, the problem emerges, and the workflow designer has to understand what solution the human is actually reaching for.
I wrote this textbook to help other workflow designers think more clearly about that process.
The framework is built around five movements:
Emergence gives you the solution. Hallucination gives you the ideal. As-is discovery gives you reality. To-be selling gives you transformation. Stewardship gives you renewal.
The textbook is free to download below.