A customer can adopt your platform and still trust their old way of doing the work more. The spreadsheet that pre-dated you. The Slack thread that still tracks the workflow. The manual export run every Friday. Those parallel workflows are what The CS Cafe calls the shadow SLA: a working system the customer trusts, running alongside yours, that your dashboard cannot see. While it runs, the customer has not migrated. They have added you. The renewal decision is the moment they choose between the two.
This matters more for AI agents than for any feature before them, because an AI agent's job is to act, not just to be present. An assistant that takes actions on data the customer no longer fully trusts will be tolerated, then routed around, then quietly disabled. The customer keeps the spreadsheet alive because the spreadsheet does not surprise them. Your AI does — sometimes well, sometimes not — and the moment it surprises them badly, the parallel workflow becomes the master record again.
What a shadow workflow looks like in practice
It rarely looks like resistance. It looks like a small, careful, repeated act the customer does outside your product, that lets them stay calm about what is happening inside it. Three shapes are common:
One, the export. Every Friday, someone pulls the data into a spreadsheet and runs a calculation that the platform also runs. The platform's number is presented to the team; the spreadsheet's number is what the operator believes. If the two ever disagree, the spreadsheet wins because the operator can see how it got the number.
Two, the side log. A separate Slack channel, a private Notion page, a shared Google Doc, where the team writes down what happened in language the platform's records do not capture. The platform's records read clean; the side log knows what actually went wrong, who actually owns the next step, and which customer is about to escalate.
Three, the fallback workflow. When the AI agent makes a decision they want to override, they do not click "override" in your interface. They run the manual process they trusted before the AI was installed, route the case around your agent, and let your agent's record sit untouched while the work happens elsewhere. Your usage metrics still count their session. Your outcome metric — if you have one — sees nothing.
Why this is the AI agent problem specifically
Older platforms could afford a shadow workflow. A reporting dashboard is fine alongside a spreadsheet — the dashboard just gets read less. The cost of being co-present is low.
An AI agent does not have this option. An AI agent is being asked to do the work. If the customer's true workflow runs in a parallel channel, the AI agent is acting on the version of the world the customer no longer trusts. Worse, the agent's actions accumulate in the platform's record, while the customer's reality accumulates in the side log. The two diverge. After ninety days the platform record and the customer's mental model are describing two different accounts.
This is the precise mechanism behind a pattern Decoding Customer Experience has been writing about all month: customers experience the operating model first. The customer is not measuring your feature. They are measuring whether the work, end to end, gets cleaner or messier with you in it. A shadow workflow is the operator's vote — quietly — that the answer is messier.
What to measure when the dashboard cannot see it
Adoption rates and usage frequencies will not catch a shadow workflow, because the customer is still using your product. They are using it and the spreadsheet. The measurement has to ask a different question: how much of the customer's actual decision-making your platform can see.
Three signals, in order of cost to collect:
One. Override rate trending toward zero is not a good sign. If a feature lets the operator override your AI agent's decision and the override rate is near zero, the most common explanation is not that the AI is always right. It is that the operator is doing the override outside your product, in a parallel workflow, so the agent's record never gets touched. Zero overrides on a still-young AI is a red flag, not a green one.
Two. Asking, with permission. A short structured ask, once per quarter, to the operator: "When you make a decision this AI is supposed to assist with, where does the final number live? In our product? In a spreadsheet? In both?" The answer maps directly to how much trust has migrated.
Three. The Friday-export interview. One CSM, one customer, thirty minutes. Ask to be walked through what the customer does on Friday mornings related to the workflow your AI assists. The spreadsheets and the side logs show up in the first ten minutes if they exist. The follow-up question is the one that matters: "What would have to be true in our product for you to stop running this on the side?"
What this changes about AI agent design
Two design choices follow from taking shadow workflows seriously, and they are choices most AI-agent platforms are not making yet.
The first is that the AI agent should make its reasoning visible by default. Not as a UX flourish, but as a precondition for the operator to trust it without keeping a parallel record. If the operator can see why the agent did what it did, they do not have to maintain a separate ledger to check it. The shadow workflow then has no job to do.
The second is that the system should explicitly support migration of the existing workflow, not bypass it. If the customer has a spreadsheet that calculates X, the right first step is for your platform to calculate X the same way the spreadsheet does, by default, with the option to change later. The wrong first step is to compute X your way and call the spreadsheet a legacy artefact. KD Be Schemin made the parallel point for support bots: closing the ticket is not the same as solving the customer's problem. Replacing the spreadsheet is not the same as earning its job.
The renewal is the test
Every shadow workflow is being run because the customer is hedging. At renewal, they choose. If they choose your AI agent, the spreadsheet stops being maintained — that is the migration signal nobody measures and everybody could. If they choose the spreadsheet, the renewal does not get signed, or it gets signed for less, or it gets signed and quietly does not get used the next year. The dashboard either way reads adoption. The shadow workflow has been telling you the truth all along.