Most small businesses we talk to now have at least one AI workflow already running. They tried something. It mostly works. Then a second one. Then a third. None of them feel terrible. None of them feel essential either.
That is the quiet problem with the current AI moment. It is easy to add a workflow. It is much harder to know which ones are pulling their weight.
This is a short note on the four properties we look for when deciding which AI workflows earn their place inside a business, and which ones are candidates for removal rather than renewal.
Does this workflow still earn its keep?
Twelve months into running real AI tools, the useful question is no longer "what should we try?" It is "what should we stop?"
In our own studio we now keep a long list of scheduled agents and workflows. Some catch and sort live software errors. Some draft project-health reports. Some triage incoming work. Some classify new entries against a shared memory store so the rest of the team's AI tools can read them. Each one started as a "let's see if this is useful" experiment. Some of them earned their place. Several did not, and we have quietly retired more of them than we have kept.
The ones we kept all share the same shape. The ones we retired all failed in the same handful of ways.
1. It handles a real repeated job
The first question we ask is whether the workflow is attached to a job that already happens. Not a job somebody might do once, or a job somebody has been meaning to do, or a job that exists because the new tool exists. A job that is already happening in the business every week.
Our scheduled error-sorting agent passes this test. The error file is going to land whether the agent runs or not. Sorting it is real work that somebody used to do by hand. The agent is doing the same job, faster, and the team's time goes somewhere more useful.
Our experiment with auto-summarising every internal chat channel did not pass this test. There was no person whose job it was to read every channel. The summaries piled up in a shared document that nobody opened. We were generating output, not removing work.
A workflow that is not attached to a job somebody was already doing is producing reports about nothing.
2. It has a named reviewer
The next question is whether somebody specific is responsible for what comes out of it. Not a team. Not a shared channel. A name.
This is the property most AI workflows fail on quietly. Output piles up. Nobody opens it. Eventually somebody asks whether the thing is still running, and nobody is sure.
The workflows we kept have a clear answer to "who reads this?" Our daily inbound triage has a specific person who picks it up before standup. Our project-health report lands at a specific time and a specific person decides whether anything in it changes our plan that week. The code-review assistant drafts comments for a named senior engineer who decides whether to keep, edit, or discard them.
A useful rule of thumb: if you cannot say who reviews the output without checking, the workflow does not have a reviewer. It has a recipient list.
3. It leaves a useful record
The third property is harder to spot, but it is the one that catches the most expensive failures.
A useful AI workflow leaves something behind that you can read later. A categorised incident in your error tracker. A row in your customer record. A line in a project report. An annotated pull request. Something that a human could pick up next month and understand without asking the agent for its working.
Workflows that do not leave a record become invisible. We had one a year ago that helpfully replied to every internal note with a brief comment. The comments were fine. None of them survived. Six months later we could not tell which decisions had been informed by the agent and which had not. We turned it off. Nothing changed, which was the point.
If the output is only useful in the moment, the moment is the only time the workflow earns its keep. That is sometimes the right answer for short, throwaway tasks. It is the wrong answer for anything you would want to defend in a year.
4. Its cost is known
The last property is the most often skipped.
Every AI workflow has a cost. Some are obvious: a subscription, a usage charge, a per-request cost. Some are quieter: the time the reviewer spends checking output, the time the rest of the team spends arguing about whether the agent was right, the time you spend trying to remember which workflow produced which thing.
A workflow that nobody can put a rough monthly cost against is a workflow that will quietly grow until somebody notices. We keep a small list of figures against each of our scheduled agents. Not just the usage bill. Reviewer time, expected output volume, and the stop condition that tells us when the agent should be paused, fed a smaller prompt, or pointed at a narrower source.
The phrase we use internally is "expected volume, expected cost, and what triggers a review." None of those have to be precise. They have to exist.
Failures need a clear review path
When an AI workflow stops earning its keep, it almost always fails on one of those four properties first, not on the model.
If the work is no longer happening, the agent has lost its job. The team moved on, the process changed, the source dried up, and the workflow is now generating output about a question nobody is asking.
If the reviewer changed jobs, took on different work, or got too busy to read the output, the workflow now produces unread content. That is not nothing, because there is still a cost. Output nobody reads is some of the most expensive output you have.
If the workflow leaves no record, you cannot audit it, defend it, or learn from it. That is fine for very low-stakes tasks. It is not fine for anything that touches customers, finance, or commitments.
If the cost grew while the output stayed the same, the usage bill is rarely the worst part. The cost in reviewer attention is.
Check the job, reviewer, record and cost
The same four properties work in reverse when somebody is selling you an AI workflow.
Before the proposal arrives, write down the answer to these four questions. If you cannot answer all four, the workflow is not ready to be commissioned. It is a demo waiting for a job.
- What repeated job, already happening this week, will this workflow take on?
- Who, by name, will review the output?
- What useful record will it leave behind?
- What is the expected monthly cost in subscription, usage, and reviewer time?
A workflow that passes those four questions has a real chance of being useful in a year. One that fails any of them is a candidate for being switched off later, no matter how impressive the first run looks.
Useful workflows still need maintenance
This is not a way to tell whether a workflow will be successful. It is a way to tell whether it is the right shape to even try.
Some workflows that pass all four still fail. The team changes. The source data drifts. The reviewer disagrees with the agent enough times that they stop reading the output. We have had agents we built carefully that died for reasons we did not predict. Useful work, but the business outgrew the question it answered.
What the four properties give you is a way to make those failures visible early, while the cost is still small. The risk is not that an AI workflow stops being useful. That is normal. The risk is that you do not notice for six months because nobody had named what success looked like.
Review the workflows you already use
Make a list. For each one, ask the four questions. Pause the ones that fail two or more.
This is uncomfortable. Some of those workflows belong to somebody on the team who built them, asked the owner to sign off on the subscription, and is still using them privately for one or two useful things. That is fine. Move the useful task into a smaller workflow that has all four properties, and switch the original one off.
In our own studio, the workflows we still run are not the most exciting ones we tried. They are the ones that fit a job that was already happening, that have a person reading the output, that leave a record we can defend, and that we can put a cost against without spreadsheet gymnastics.
That is the only shape we have found that survives a year of weather. The rest is theatre.
