Companies keep finding that an AI tool which looked fine in a pilot goes wrong when it reaches real work. The reasons are mostly inside the company: scattered data nobody will hand over, a risk team and an engineering team that do not trust each other, and a way of counting success that rewards the wrong thing. Each one looks different from the reader's side of the desk than from the chief information officer's.
A pilot avoids the folders that cause trouble
A pilot usually gives the assistant a small, hand-picked set of files. Microsoft's own rollout advice suggests starting with up to 100 popular, low-risk sites. That is a sensible way to learn, and it also means the first time the assistant meets the company's real permissions is the day it is switched on for everyone. We already wrote about why so many pilots never reach that day; the permissions are one reason.
Think of a 60-person distributor. In 2019 someone saved a salary spreadsheet in a shared folder and gave "everyone in sales" access so a colleague could open it once. Nobody has looked at it since. A sales rep who asks the new assistant "what do senior reps earn here?" gets a clear answer, with the file named underneath. The assistant did nothing wrong. It found something the rep was already allowed to open, and the rep never had to know where to look.
That is why permissions are the first job. The assistant does not break the rules the company set. It follows them exactly, and exposes the gap between the rules the company meant to have and the ones it has. For a small business this is an afternoon's work: list the ten folders you would hate to see read aloud, and check who can open each one. For a large company it is a project, and the sensible way through it is to fix access one business area at a time instead of waiting for a multi-year cleanup.
Waiting for zero risk and going around the rules fail the same way
Risk teams that demand a promise of zero errors will never approve anything, because no tool can make that promise. People with deadlines then use a public AI service on their own, and the company ends up with less control than it would have had by saying yes with limits. A department head can break that standoff by asking for a pre-approved pattern instead of a decision: internal search over approved folders is one pattern, a chatbot that talks to customers is another, with tighter limits. Barracuda's chief information officer built a standing AI council to deal with the same standoff.
Time saved is not money saved
Vendors count minutes. A recruiter who saves 15 minutes a day gets back about 55 hours over a working year, nearly seven days. That only becomes money if the extra hours go to something the business gets paid for, such as talking to more candidates. If the real wait is the hiring manager taking a week to schedule interviews, the recruiter has the same output and a quieter afternoon.
Employees understand this before executives do. Many suspect that a measured saving will be turned into a smaller team, so they use the tool lightly. I expect the companies that get real use to be the ones that say, before rollout, what the saved time is for, and measure the thing that matters, like how many days a customer complaint takes to resolve, instead of how many people logged in.
What this shows
A company's permissions, review habits and targets are its real organization chart, written down in settings nobody reads. AI assistants read all of them at once. The companies that do well will not be those with the best model, since everyone can rent the same one. They will be those whose own rules were already in good order when the assistant arrived.