Leadership1 publisher3 min readPublished
A product chief retrains her AI agent every couple of months because its answers get worse
Sumaiya Noor, a UK product chief, says her AI customer service agent gets worse over time and needs retraining every couple of months. For teams adding agents, that upkeep becomes recurring manager time, and so far only individual accounts describe it.
The Board Room · Leadership desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Noor's agent is meant to pass her tasks it cannot handle, but over time it has taken on problems outside its scope and given customers confident, wrong answers.
- She manages three AI agents and seven people, and said deteriorating performance is a major difference between the two groups.
- In a Boston Consulting Group report published in June, 30% of respondents said their organisations had built AI agents into workflows, up from 13% a year before.
- A separate BCG study of 1,488 full-time US workers called the mental fatigue from excessive agent monitoring "AI brain fry".
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- cost Each deployed agent comes with a recurring maintenance block taken from its manager's own work, about six sessions a year at Noor's cadence, and a rollout plan has to pay for those hours.
- decision Leaders setting spans of control now have to decide whether an agent counts as a report, since on Noor's team agents make up three of ten and each draws on her time.
- exposure Customers meet drift before the manager does, because an agent answering confidently outside its scope produces the angry customer first and the correction later, and the manager answers for it.
- constraint Time that agents free up goes into oversight, so a business case that treats agent output as hours handed back to staff overstates the saving.
The evidence that agents decay is thin so far. It rests on one named manager, who said "Over the period of time, it gets worse rather than getting better" [5]. It also rests on workers who listed "rot" among their complaints, along with long-running tasks derailed by a single error and agents duplicating one another's work [8]. Business Insider's reporting does not say how fast performance falls or how long a fix takes, so we do not know yet whether drift is common to agents in general or limited to particular deployments.
Noor, chief product officer at a UK-based social-impact investment platform [1], pays for the drift out of her own working hours. Every couple of months she sets work aside to refresh the agent on what it is supposed to do and to train it on more edge cases [3]. If a couple means two, that is about six sessions a year for one agent [19]. She also runs two other agents, a junior product manager and an internal "chief of staff" [16].
The cost of overseeing agents is better documented than the drift. Sergio Freitas, an engineering director at Cisco, said agents have neither lengthened nor shortened his days, and that orchestrating them demands more task switching and more oversight [9]. "The fatigue comes from the ability to do more," Freitas said [10]. "I spend less time doing things and more time checking them," Dan Lewis, a newsletter author, told Business Insider [11]. Elijah Wee, an associate management professor at the Foster School of Business at the University of Washington, said part of the challenge is keeping up with agents that can operate around the clock [14]. "People burn themselves out signaling their worth," Wee said [15].
The obvious objection is that this is launch friction, and better tools and habits will remove it. Some of it is. Sebastian Gierlinger, vice president of AI and IT at Storyblok, said developers were "genuinely frustrated" by how much code they had to review after the company first rolled out AI tools [12]. He said review and testing practices now have to keep pace with how fast employees can build [12]. Storyblok's answer is a process change, made once and then maintained. Noor's problem is different: an agent trained to hand off what it could not handle started answering outside its scope [2], and her correction comes round again on a schedule [3].
The trade-off this quarter is between how fast agents go onto teams and how many manager hours are held back to maintain them. On BCG's survey figures, the share of organisations with agents in their workflows more than doubled in a year [18]. On Noor's team, three of ten direct reports are agents [17]. A leader who adds agents to a manager's span without removing other work is assuming the refresh cadence stays small. If it grows, next quarter's cost appears first in the customer queue. The manager carries it, because workers are on the hook when their agent gets something wrong [13].
What to watch
- A measured study of how agent performance changes over months of deployment would give leaders a decay rate to plan manager hours against.
- Whether agent vendors add scope limits or monitoring that stretch the interval between refreshes beyond the couple of months Noor describes.
- Whether companies start counting agents in managers' spans of control when they set team sizes and targets.