Recurring AI Workflows Need an Expiry Date
Recurring AI Workflows Need an Expiry Date
Say you run sales for a small team, and a few months ago you set up an AI workflow to prepare a pipeline brief every Monday. It reads the CRM after Friday's updates, flags the deals that look at risk and has the brief ready before the pipeline meeting.
For a while it's the first thing everyone opens. Then the team changes how it works. The Monday meeting turns into one-on-ones spread across the week. Sales redraws which accounts count as mid-market. The brief keeps arriving every Monday, accurate and well formatted, flagging deals by a segment definition nobody uses anymore. Someone still skims it. Nobody makes a decision with it.
Nothing broke, so nothing got anyone's attention. The workflow is still running on a decision someone made months ago, and the schedule can't tell you that decision is out of date.
A schedule says when work runs. It says nothing about whether the brief still serves a meeting that happens, or whether the segments it sorts by are still the ones sales uses. Those answers need writing down, with a date when someone checks them again. We call that a Cadence Contract.
Why recurring work stops getting looked at
Setting up a workflow is a decision someone makes. Keeping it running usually isn't. Once the schedule works, each good run looks like a reason for the next one.
What made the Monday brief useful was its judgment about which deals looked at risk. Anthropic's guide to building effective agents separates workflows, which run "through predefined code paths," from agents, which "dynamically direct their own processes and tool usage." The more a recurring job leans toward the second, the less a good first run can vouch for every Monday after it, once the CRM, the segments and the meeting have all moved on.
That leaves two questions. Monitoring can tell you whether this Monday's run did what it was designed to do. Whether the brief still deserves its Monday slot is something a person has to decide, and nobody does until a date makes them.
What a Cadence Contract covers
A Cadence Contract is a short set of written answers about one recurring workflow, with a date when its owner decides again. When that date comes, nothing gets deleted. The owner looks at how the runs went and decides whether to continue, revise, pause or retire the work.
It answers seven questions:
- Purpose: which repeated decision does it improve? For the brief, which deals need attention this week.
- Trigger: what starts a run, and what happens if one is late, skipped or doubled? Friday's updates, ahead of Monday's meeting.
- Inputs: which sources, how fresh, and which wins when they disagree? This week's CRM updates and the current segment list.
- Action boundary: which tools, recipients and actions are allowed, and where must a person approve? Nothing leaves the sales team.
- Exposure limit: how much money, reach or irreversible consequence can one run, or a month of runs, create? For an internal brief, very little.
- Health evidence: what shows it's still accurate, used and worth the review time? Whether anyone acts on the deals it flags.
- Renewal and exit: who owns it, when is it reviewed, what stops it early and how is it retired?
The workflow keeps running while those answers hold. If one stops holding, the next run waits, someone reviews it or the recurrence ends.
Every recurring workflow is a subscription to its own assumptions. The brief subscribed to a Monday meeting and a segment list that have both changed, and it renewed itself every week.
A schedule answers when, not whether
"Every Monday at 9" isn't a complete trigger. It should also say why that timing fits, whether running twice is safe, what happens after a missed run and whether new data should start the work instead of the calendar.
How often it runs should follow the decision it supports. The brief made sense on Monday because the pipeline meeting was on Monday. Once the meeting turned into one-on-ones through the week, the Monday run was a habit.
Missed and repeated runs matter differently depending on the job. The scheduling chapter of Google's Site Reliability Engineering book uses three examples. A garbage-collection job that runs every five minutes may be able to skip one launch. A monthly payroll job should not be skipped. A newsletter job should not be launched more than once. Google says it favors skipping a launch over risking a double launch, because undoing one, like a newsletter sent twice, may be difficult or impossible.
For the brief, a skipped week is an annoyance and a second copy is harmless while it stays inside the sales team. That changes the moment it starts emailing reps or editing CRM records, so write the rule down while the stakes are low.
"Run weekly" is configuration. "Run before the Monday pipeline meeting once Friday's CRM updates are in, skip rather than run twice, and send nothing outside the sales team without approval" is an operating rule.
Check the inputs before every run
The brief shouldn't run just because the clock fired. Before it flags anything, it should check that Friday's updates actually arrived. If the sync failed, the brief should say so instead of reporting a quiet week. When a source is missing, stale or contradicted, it should hold, narrow what it claims or ask someone.
This is where schedules quietly decay. The instructions stay the same, so it's easy to assume the context has too. Meanwhile a segment gets redrawn, a CRM field gets renamed, the model is upgraded and the person who set it up moves to another role.
So the inputs clause names the sources, how fresh each has to be, which one wins in a conflict and what happens when data is missing. For the brief, fresh means this week's CRM updates, and the segment definition should come from wherever sales keeps the current one, not from a copy pasted into the instructions at setup.
NIST's AI Risk Management Framework says risk management should be continuous, timely and carried out throughout an AI system's lifecycle. A weekly brief goes through a small piece of that lifecycle every Monday, so each run checks its sources again and reuses last Monday's method, not its conclusions.
Measure whether it's still useful
A workflow can complete every run while quality slips, corrections pile up, people stop reading the output or reviewing it takes longer than doing the work by hand. So the evidence has to cover more than whether the run finished. For the Monday brief, it might be:
- missed or late runs, and runs that started on stale data
- corrections you had to make, and how often
- flagged deals someone overruled or ignored
- the time it takes to review each week
- what each brief costs, set against the briefs someone actually used
- whether anyone acted on the deals it flagged
Pick the smallest set that could change your decision to keep it running. One weekly brief doesn't need its own dashboard.
Anthropic's guide to evaluating AI agents is plain about the limits of automated evals: they need ongoing maintenance as the product and model change, and they "can create false confidence" if they don't match real usage patterns. It pairs them with production monitoring and periodic human review.
Recurring work has its own version of that problem. A workflow can keep passing a check that no longer describes the job. The Monday brief would pass "flags at-risk deals by segment" every week while using segments the team retired.
NIST's report on challenges to monitoring deployed AI systems lists the right cadence for monitoring among its open questions, along with the burden monitoring puts on users. With no settled answer to borrow, you pick the brief's review date yourself, and the evidence above is what you look at when it arrives.
Renewal has to be earned
Expiry is the point where the workflow stops renewing itself without anyone deciding. The contract needs a planned expiry and early stop conditions. A planned expiry could be after the first three runs, at the end of a campaign, at the next quarterly review or when the decision the workflow supports is due. For the brief, the end of the quarter is a natural point, since that's when the sales team plans the next quarter anyway.
Any of these should stop it early:
- you change roles or leave
- the CRM data it depends on stops arriving, or the segment list is replaced
- your corrections pass a level you agreed on
- it starts doing anything outside the sales team
- nobody acts on several briefs in a row
- the cost or your review time outgrows what the brief is worth
- an incident shows a failure your current checks don't cover
In the example, two of those would have fired well before the quarter ended. The segment definition the brief relied on was replaced, and nobody acted on several briefs in a row.
At renewal, the owner answers five questions:
- Is the purpose still active?
- Does the cadence match how often there's something new to know?
- Are the inputs, permissions and limits still current?
- Do the runs so far support trusting it for another period?
- Is the current owner willing and able to renew it?
Why an expiry date lets the work run more freely
Once the contract is written, you can stop approving each Monday's brief by hand. It already says what the brief may touch, what would show a problem and when the decision comes back to you, so ordinary Mondays carry on and only the exceptions interrupt anyone. The expiry date is what makes that safe. You're approving the brief until the end of the quarter, not every brief it will ever produce.
Different workflows deserve different amounts of room:
- An internal summary with no outside action, like the Monday brief, can run on its own with spot checks.
- A customer newsletter can be prepared automatically and approved before it's sent.
- A workflow that updates CRM records can change a few named fields and send anything else to a person.
- A payment or deletion workflow needs explicit approval for every outside action, however often it repeats.
If the brief ever starts updating deal stages, it has moved down that list, and its action boundary, exposure limit and contract should change with it.
NIST's framework also calls for ways to "supersede, disengage, or deactivate" AI systems whose performance or outcomes don't match their intended use. A renewal date is a plain way to build that in. On a day everyone knows, the person who can switch the brief off has to decide whether to.
How to write one
Write the contract when you put a proven piece of work on a schedule. Read the first few briefs closely. Renew only if they flag the deals you'd have flagged yourself and reviewing them takes less time than building the brief by hand.
Purpose:
Trigger, and what happens after a missed or duplicate run:
Inputs each run needs, how fresh, and which source wins:
Allowed actions, tools, recipients and approvals:
Limits per run and per period:
Health evidence and how often it's reviewed:
Owner:
Expiry date or maximum number of runs:
Early stop conditions:
Renewal decision: continue, revise, pause or retire
For the Monday brief this fits on half a page. The owner is you, the expiry is the end of the quarter and the first stop condition is a change to the segment definition.
Start with work that has already gone well by hand. If the instructions, inputs and review standard change a lot every time, you don't have a recurring workflow yet. You have a recurring need that still needs designing.
Running it on Supanova
Say the Monday brief ran on Supanova. You'd write the instructions once and set the work to repeat every Monday. After Friday's updates, you attach that week's pipeline report, and your atoms prepare the brief from it. Finished work comes back for your review, so each brief lands with you, and you can read it the way the contract asks: is it still flagging the right deals, by the right segments, for a meeting that still happens?
When the quarter ends, whether the brief keeps its Monday slot is your call, and the briefs you reviewed along the way are the evidence.
Pick one recurring workflow your team already runs. When did someone last decide it should keep running? If nobody can say, give it an expiry date.
If you want a workforce to take on your recurring work while you keep the decisions that matter, start at supanova.team.
Sources
- Building effective agents, Anthropic Engineering (2024).
- NIST AI Risk Management Framework 1.0, Section 5: AI RMF Core, National Institute of Standards and Technology (2023).
- Challenges to the Monitoring of Deployed AI Systems (NIST AI 800-4), NIST Center for AI Standards and Innovation (2026).
- Demystifying evals for AI agents, Anthropic Engineering (2026).
- Distributed Periodic Scheduling with Cron, Štěpán Davidovič, in Site Reliability Engineering, Google (2016; free online edition 2017).