AI Automation

Who Owns an Automation After the Builder Leaves?

Written by
Pravin Kumar
Published on
Sep 15, 2026

Who owns an automation after the person who built it leaves?

Whoever holds the credentials and receives the alerts. Not whoever is named in the handover document, not whoever the org chart says, and not the department the automation serves. If no single named person has both access to stop it and notification when it breaks, that automation is unowned regardless of what any document claims.

I build automations for clients, which means I am frequently the person who leaves. That gives me an unusually direct view of what actually survives a handover and what quietly does not, and the answer has almost nothing to do with documentation quality.

The uncomfortable version of this is that most organisations have automations running right now that nobody owns. They work, so nobody looks. They will keep working until they do not, and the discovery process at that point is archaeology.

Why is a handover document not the same as ownership?

Because a document describes a system and ownership is a relationship with it. You can write a perfect runbook, hand it to someone, and change nothing about who notices when the thing fails, which means you have transferred understanding without transferring responsibility.

The test is embarrassingly simple. If the automation silently stopped working this afternoon, whose Tuesday gets worse? If the honest answer is "nobody, for about three weeks", the document did not do what you think it did.

This is not an argument against documentation. Documentation is necessary and I write it, and the case for doing so properly is in writing automation handoff documentation clients can actually use. It is an argument that documentation is the second step, and people keep treating it as the only one.

What are the two things ownership actually consists of?

Access and consequence. Access means a named human can log in, see what the automation is doing, and turn it off without filing a request or hunting for a password. Consequence means that same human finds out when it misbehaves, ideally before a customer does.

Both are required and neither is sufficient. Access without consequence produces somebody who could help and does not know they need to. Consequence without access produces somebody who knows there is a problem and has to go find the person who can do anything about it, which is the worst of the available outcomes.

When I hand over work, those two things are what I actually transfer. Everything else is supporting material. If I have moved the credentials and redirected the alerts to a named person who has acknowledged receiving them, the handover has happened. If I have not, I have just written an essay.

What happens to credentials when someone leaves?

This is where most of the real damage lives. Automations authenticate as somebody, and if that somebody was the person who left, the automation dies the moment their account is deprovisioned. It is often the single most predictable failure in the whole category and it still catches organisations constantly.

The fix is to never let an automation authenticate as an individual in the first place. Use a dedicated account that belongs to the organisation rather than to a person, keep its credentials in whatever password manager the team already uses, and make sure the recovery address is a shared one rather than one human's inbox.

Whether the connections live in Zapier, Make, n8n, Airtable, or HubSpot, the principle is the same, though the mechanics differ enough that I would read each vendor's current documentation on transferring ownership rather than assuming. What does not differ is that this is much easier to arrange on day one than during an offboarding.

Who should get the alert?

One named person, with a second named person as backup, and neither of them should be a distribution list. A channel that everyone can see is a channel nobody reads, and an alert that arrives to twelve people arrives to zero.

The named person does not have to be technical. They have to be able to recognise that something is wrong and know who to call, which is a much lower bar than being able to fix it. Diffusing that responsibility across a team feels safer and is the opposite.

Route the alert somewhere they already look, which in most organisations means Slack or email rather than a dashboard they would have to remember to visit. An alert in a system nobody opens is indistinguishable from no alert at all, and it is worse because it lets everyone believe they are covered.

Does it matter if the builder was an agency or a contractor?

It matters enormously, and the direction surprises people. External builders are often better at this than internal ones, because everyone knows from the start that the builder is leaving. That knowledge forces the ownership conversation to happen while there is still time for it.

Internal builds go wrong more often precisely because nobody thinks of them as temporary. The person who built it is sitting right there, so the ownership question never gets asked, and the automation accumulates undocumented dependency on one employee for three years until they resign on a Friday.

Across more than seventy projects for over twenty-five clients, the thing I have become most insistent about is naming the receiving owner before the build starts rather than at the end. If nobody on the client side will put their name on it, that tells me something important about whether the automation should exist at all.

What about automations built by an AI agent?

The same rules, with one addition. A Claude Code agent orchestrating steps through an MCP server still authenticates as something, still fails in ways somebody has to notice, and still needs a human whose Tuesday gets worse. Nothing about an agent changes the ownership question.

The addition is that the logic may now live in a prompt rather than in a visual canvas, which is genuinely better for handover if you treat the prompt as code and keep it somewhere versioned. It is genuinely worse if the prompt lives in somebody's local setup and nobody else has seen it.

There is also a new failure mode worth naming, which is that an agent-built automation can be harder for a non-technical owner to reason about, because the behaviour is described rather than diagrammed. That raises the bar on the alert being legible, since the owner cannot easily inspect the logic themselves.

How do you assign ownership to someone non-technical?

Give them three things and nothing else. What this automation is for, in one sentence about the business rather than the system. How they will know it has stopped. And exactly who to contact, with a name and a contact method, not a role.

Resist the urge to teach them the system. They do not need to understand the branching logic, and attempting to explain it usually convinces them that ownership is beyond them, which is the opposite of what you wanted. Ownership here means noticing and escalating, not diagnosing.

What I do in practice is write the one sentence with them rather than for them. If the owner can say in their own words what breaks in the business when this stops, they will recognise the failure when they see it, even if nobody ever sends them an alert. That is a kind of resilience a runbook cannot produce, and it complements the sort of failure planning in keeping an automation alive after a tool changes underneath it.

What does an unowned automation actually cost?

More than a broken one, because a broken automation announces itself. An unowned automation degrades. It keeps running with a stale assumption, writing slightly wrong data into a system of record, and the cost accumulates quietly in a place nobody is looking.

Volume makes this sharper. The Airtable and WhaleSync automation I run for Ajust has delivered more than 25,000 cases and helped over 400,000 people, saving upwards of 50,000 hours, and at that scale a silent wrong assumption is not a small problem. The HubSpot automation I built for Kismet Health runs through Zapier into a CRM, where wrong data is expensive in a different way because people act on it.

In both cases what makes them safe is not that they never fail. It is that when they do, a specific person finds out. That is the entire difference, and it is worth more than any amount of build quality.

What should you do next?

Pick your three most consequential automations and, for each, write down one human name for access and one human name for alerts. If you cannot fill in both, you have found an unowned automation and you now know more than you did this morning.

Then check whose account each one authenticates as. If the answer is a person rather than an organisation account, fix that before you fix anything else, because that is the failure that arrives on a schedule you do not control.

If you have inherited a set of automations and you genuinely do not know what they do or who is responsible for them, reach out. Mapping that is unglamorous work and it is far cheaper than finding out during an outage which of them mattered. And if you are deciding what these things should be permitted to do in the first place, start with what an automation should never be allowed to do.

Get found, cited and the back office automated

Let's make your site the source AI engines quote and wire up the systems behind it.

Contact

Let's get your website found and cited by AI

Tell me what you're working on, whether AI search is skipping your product, your back office is buried in manual work, or you need a build that does both.

Got it, thanks. I read every message personally and reply within 1-2 business days.
Oops! Something went wrong while submitting the form.