I’m worried that they put co-pilot in Excel

You don't have a human to manage. The relationship is completely one-sided, you can query a generative AI at 3 in the morning on new years eve. This entity has no emotions to manage and no own interests.

There's cost.

There's an implicit promise of improvement over time.

There's an the domain of expertise being inhumanly wide. You can ask about cookies right now, then about XII century France, then about biochemistry.

The fact that an average worker would be fired if they perform the same way is what the human actually competes with. They have responsibility, which is not something AI can offer. If it was the case that, say, Anthropic, actually signed contracts stating that they are liable for any mistakes, then humans would be absolutely toast.

The only relevant metric here is how often each thing makes mistakes. Programs are the most reliable, though far from 100%, humans are much less than that, and LLMs are around the level of humans, depending on the humans and the LLM.

While humans have the same risk factors, human oriented back-office processes involve multiple rounds of automated/manual checks which are extremely laborious. Human errors in spreadsheets have particular flavors such as forgotten cell, misstyped number, or reading from the wrong file/column. Human's are pretty good at catching these errors as they produce either completely wrong results when the columns don't line up - or the typo'd number is completely out of distribution.

An AI may simply decide to hallucinate realistic column values rather than extracting its assigned input. Or hallucinate a fraction of column values. How do you QA this? You can't guarantee that two invocations of the AI won't hallucinate the same values, you can't guarantee that a different LLM won't hallucinate different values. To get a real human check, you'd need to re-do the task as a human. In theory you can have the LLM perform some symbolic manipulation to improve accuracy... but it can still hallucinate the reasoning traces etc.

If a human decided to make up accounting numbers one out of every 10000 accounting requests they would likely be charged with fraud. Good luck finding the AI hallucinations at the equivalent level before some disaster occurs. Likewise, how do you ensure the human excel operator doesn't get pressured into certifying the AIs numbers when the "don't get fired this week" button is sitting right their in their excel app? how do you avoid the race to the bottom where the "star" employee is the one certifying the AI results without thorough review?

I'm bullish on AI in backoffice, but ignoring the real difficulties in deployment doesn't help us get there.

- Legacy systems typically have error modes where integrations or user interface breaks in annoying but obvious ways. Pure algorithms calculating things like payroll tend to be (relatively) rigorously developed and are highly deterministic.

- LLMs have error modes more similar to humans than legacy systems, but more limited. They're non-deterministic, make up answers sometimes, and almost never admit they can't do something; sometimes they make pure errors in arithmetic or logic too.

- Humans have even more unpredictable error modes; on top of the errors encountered in LLM's, they also have emotion, fatigue, org politics, demotivation, misaligned incentives, and so on. But because we've been dealing with working with other humans for ten thousand years we've gotten fairly good at managing each other... but it's still challenging.

LLMs probably need a mixture of "correctness tests" (like evals/unit tests) and "management" (human-in-the-loop).