Who Keeps Your AI Running After the Vendor Leaves?

Most AI rollouts have a launch owner. Far fewer have a running owner. In the accounts below, the trouble started after setup, when the person who built the tool moved on or a dependency quietly changed. The tool kept running, its output got worse, and nobody's job was to notice. What fixed it was rarely technical. It was a named person, time on a calendar, and a small check that runs every time.

Pattern 1: The tool outlives its owner

The clearest failures came when the launch owner changed jobs. Ace Zhuo, CEO of TradingFXVPS, described how it happened on his support team:

Our launch owner was a support leader and wrote out all the prompts for the AI to write out tickets in and also set up the knowledge base to sync. She is now a sales person and no one has touched the AI support for about 6 weeks.

In the meantime the knowledge base sync had ceased to function due to an expired API token. Subsequently generated AI replies would be sent to agents containing reference to old pricing and a location that was no longer offered.

A customer found it first. The tool now belongs to an ops person on the support team rather than an engineer. "It gets reviewed on her calendar for 90 minutes every Monday where she goes through 20 or so AI generated replies," and every product or pricing change carries a required "AI knowledge base updated?" checkbox.

Nicolas Falourd, CEO of Cyber Techwear, saw the same thing with a chatbot a vendor handed to a non-technical ops person:

We tested ownership when that ops person resigned: prompts had drifted, integrations missed schema changes, orders were mis-tagged. We responded by assigning a single owner (a part-time engineer shared with e-commerce), documenting prompts and integrations in a living runbook, and scheduling weekly 30-minute QA checks so small drifts get caught before they compound.

Pattern 2: Tools rarely break loudly

The failures here didn't look like outages. Jobs reported success, and records looked complete. Jens Hagel, co-founder of Yupana, caught one in a bank import:

Ownership got tested when a bank import reported success but had lost a field: from a newer CAMT version, the counterparty sits one level deeper in the XML. A simple count of recognised fields caught it, and that count now runs on every import. Tools rarely break loudly, so after launch somebody has to own the dull checks.

Jack Wright, Senior Vice President of Operations at HealthWright Technologies, runs AI on routine work across five small companies, each job following a written rulebook. One of those jobs lost its record:

One of our automated jobs was cut off mid run by a usage limit. It only wrote its record at the end, so that record was lost and the rerun had nothing to check against. It could have repeated work it had already finished. The fix went into that job's rulebook: it now records each send the moment it goes out and reads that record before it starts. For me, owning AI means deciding what has to be written down.

Pattern 3: The owner sets the limits

In one-person businesses there is no question of who owns the tool. The harder job is deciding what it may conclude or do without being checked. Aziz Chaaben, founder of Groundwork, found this out from his own agent:

In a one-person business the running owner is me, and the moment that proved it was when my AI agent graded a whole section of my site as failing. It had looked at two pages. Search Console showed that same section bringing in 43% of my clicks. Nothing had broken; the agent was confidently wrong on too little data, and it would have steered me away from the only part of the site that was working.

The agent must now print its sample size next to every claim about his numbers, and it may not grade anything below 30 pages.

Nikita Vorontsov, founder of CharGen, found that an outreach agent had sent two emails to addresses it guessed. "No step in the chain checked where an address came from." His sender now refuses any email without a verified address and the URL where that address was found. He also switched off automatic sending and moved 14 approved emails back for checks. Only 4 of the 14 were confirmed, and 10 were dropped.

Who should own it

The contributors who had been through this tended to agree on one thing: the owner should be the person closest to the output, whatever their title. Jonluke Davis, founder of Data's Alpha Circle, runs a member-facing assistant that handled 586 questions in its first month, and every bad answer landed on him:

If I were handing one of these to a support or ops team, I'd pick the person who reads its output every week and give that person edit access to the prompts and the sources. A vendor can set it up. Only someone reading the answers will notice when they start to slip.

The through-line

None of these fixes needed a new tool or a new hire. They needed one name, a recurring slot, and a check dull enough that nobody would run it by choice. Ace Zhuo had the line for it:

AI drift is like server drift. It's not a crash, it's a degradation. If no named human has it on a calendar, you are running unmonitored infrastructure and referring to it as automation.