What is the maintenance burden of internal AI tools and who carries it?

Interviewer's question: "What is the maintenance burden of internal AI tools and who carries it?" Marisela Odom handles IT and operations for Redbrick Performing Arts Trust, a network of venues that share an internal AI tool drafting show load-in logistics schedules for stage crews.

How to answer this, stage by stage

Nobody is grading whether your final number is exactly right. They're grading whether you show the arithmetic and name the one assumption the number actually depends on.

Let's learn

Say we build an internal AI tool. It ships, it works, everyone claps in the launch meeting. Then, six months later, who actually answers the email when it breaks?

Redbrick's load-in scheduling tool reads a show's technical rider and drafts a crew call sheet, times, positions, equipment, for a stage manager to confirm before a show. Before the tool, a stage manager built the same sheet by hand from the rider, about ninety minutes per show.

The tool cut drafting time to about fifteen minutes a show. Nobody budgeted any ongoing time for it at all, since it felt, at launch, like a thing you build once and leave alone.

Three real costs, and none of them were on anyone's calendar at launch.

The turn: the tool never broke in any dramatic way. It just quietly needed someone, every week, to notice when a vendor's model update changed how it read a rider, and someone, every time a crew hit something odd, to answer the email personally instead of a support queue catching it.

At its worst: no one is ever formally assigned this work, Marisela quietly absorbs it on top of an already full role, burns out within a year, and the tool's actual institutional knowledge leaves with her the day she does.

What I would leave alone: the tool's core scheduling logic doesn't need any ongoing rebuilding. This isn't a model problem, it's a staffing and ownership problem, and treating it as a technical fix wastes effort on the wrong target.

The lesson: an internal tool's maintenance burden is really a staffing question wearing a technology costume, and the honest range depends entirely on whether anyone was ever asked to own it.

Now here is the same thing as a story

The short version above is what you'd say to Redbrick's board. Read this one for how the actual number got found.

Every Thursday, Marisela Odom checks the scheduling tool's drift log before her official workday even starts, a habit she picked up without anyone asking her to.

Four steps, every single Thursday, and none of them are written into her actual job description.

Two venues emailed her directly in the same week, both with the same complaint: the tool had started drafting call sheets with the wrong crew size after a vendor model update quietly changed how it read certain riders. Neither venue had any other number to call. Marisela's name was just the one everyone had saved from the launch meeting.

She spent the rest of that week fixing both riders' prompts by hand, on top of her actual job, plus fielding a dozen more emails from crews who'd noticed something felt off but couldn't say exactly what. She finally sat down and logged every hour she'd spent on the tool that month: 34 hours, almost a full extra work week, none of it on any calendar or job description anywhere.

Marisela brought that number, 34 hours in a single month, to her director. Redbrick named her the tool's actual owner the following week, with two protected hours a month built into her calendar for it and a real ticket queue so crews stopped emailing her personally for every small thing.

BOUND, in one screenNot a staffing memo. BOUND is what turns "who carries this" into an equation you can actually check.

Only one corner is worth a week of anyone's time, and it isn't the corner most people reach for first.

The recap, one line per letter: break it down is the equation stated before any number, own numbers is naming where 13 and 2 hours actually came from, use a range is 14 to 38 hours instead of one, nail the sanity check is comparing it to similar teams' real staffing, and direction is naming ownership status as the assumption worth chasing.

And if you want to be sure it really works, try it somewhere elseSame five letters, an urgent-care clinic network instead of a performing arts trust. Nothing else about the two jobs is alike.

Cascade Urgent Care Network runs an internal AI tool that summarizes patient intake notes for clinicians across a dozen walk-in clinics.

Same equation, same swing, a completely different building.

Mapped onto BOUND: break it down is the same equation, monthly hours equal drift monitoring plus prompt updates after clinical-terminology changes plus ad hoc clinician support. Own numbers means naming that drift monitoring across a dozen clinics likely runs higher, around 18 hours a month, given more real intake volume to check. Use a range means acknowledging 20 hours a month with a named clinical-informatics owner, up to 50 without one, since clinicians reaching out directly scales with clinic count. Nail the sanity check means comparing that to how much time a hospital IT team typically budgets per clinical tool. Direction means naming ownership status as the assumption that moves the number most, exactly as it did for Redbrick.

He didn't write one number down. He wrote the range the honest answer actually lives in.

Swap the trigger and it still runs.
Speed: an interviewer caps you at thirty seconds. Say "14 to 38 hours a month, and it swings on whether the tool has a named owner, not on how many venues use it," and stop.
Cost: if there's no budget to formally name an owner, at minimum log the real hours being spent informally, so the true cost stops being invisible to whoever approves budgets.
The model gets better, for real: if the underlying model needs fewer prompt updates over time, the estimate still doesn't fall much, because ad hoc support was never really about the model, it was about the tool having nobody clearly responsible for it.

Three real tasks, and a fourth thing that isn't a task at all, just a gap nobody filled.

Where people run it wrong.
They budget for building the tool and assume upkeep will somehow take care of itself.
They track venue or clinic count as the driver of cost, when ownership status swings the number far more.
They let the most competent person quietly become the owner by default, instead of naming one on purpose.

How to use it live. When someone asks you about an internal AI tool's maintenance burden, ask yourself one question first: who, specifically, answers the email six months from now. If you can't name a person, the real cost is still invisible.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0