AI is working at your company. Do you know where?

Human in Charge is a weekly newsletter about how businesses actually use AI and who stays in control. This is the third one, on the shadow AI your best people have already built without telling you.

Anna runs operations at a forty-person company. Every Monday she used to spend most of the day pulling numbers out of four systems and turning them into something the leadership team could read.

She doesn’t any more. Some time in the spring she worked out a sequence of prompts that does the dull part in about twenty minutes. She built it in her own account, on her own laptop, because nobody told her not to and nobody asked.

It is the single most valuable process improvement at that company this year. It is written down nowhere. Her manager does not know it exists. If Anna leaves in March, it leaves with her, and the Monday report quietly goes back to taking a day.

There is a name for this now. Shadow AI: the tools and the workflows running inside a business that the business has no record of. Some companies discover theirs by accident, maybe when somebody resigns, maybe never.

Shadow AI: the 10% who already worked it out

John Levitt has seen this across various companies, and he puts a number on it that is uncomfortably believable.

“We have a lot of people that are doing things successfully with AI, but they’re working as individuals on their own computers, and we don’t have a great way to take that work and make it collaborative and scale it the way that we need to.”

“Just taking that 10% of users who are the experts and can figure things out for themselves, and how do we scale that 10% to 50% or more of the company.”

The problem he describes is not that people are using AI badly. It is that a tenth of the workforce is using it well, privately, and the other nine tenths get none of it.

Meet John Levitt, co-founder of elvex

Before elvex, John and the people he now runs it with spent thirteen years at Parse.ly, a digital analytics company whose customers included the Wall Street Journal, Condé Nast, Bloomberg and the National Football League. They were early to this. They were integrating natural language processing into the product long before anybody called it AI.

What happened next is interesting, as it is the opposite of the story founders usually tell.

“As we started introducing these tools into our product for our customers, there was a lot of upheaval. People were very scared and nervous and had lots of questions, and asked us to roll things back.”

Customers asked them to undo it. Not because the technology did not work, but because nobody could tell them what it was doing or who was answerable for it. That request, to please put it back how it was, is what elvex was built in response to.

The demo day that surfaces shadow AI

Onboarding at elvex starts small. Ten or twenty people, the ones who already tinker, plus a stakeholder from each team. They build things for a few weeks. Then everyone presents what they made to the executives.

So, what do they do there?

“You have the executives who are naturally detached from that level of granularity of work, who just kind of sit there inquisitively and start trying to poke holes. ‘Well, why wouldn’t you just do this instead?’ And it’s like, well, because to do that there are all these dependencies that you are not aware of.”

The leadership team arrives expecting to evaluate. They end up being taught the shape of their own business by the people running it.

“People who are doing the work know where all the potholes are, and you can step over them.”

“Those conversations end with a level of respect that I think is hard to garner otherwise. And it’s really important to have that respect and support from the top down, so that the bottom up can feel confident in what they’re doing.”

Anna, in our imaginary forty-person company, never got that meeting. Nobody asked her to present anything, so her twenty-minute Monday stayed a private convenience instead of becoming how the company works.

Why locking it down makes shadow AI worse

The instinct, when a company realises how much is happening off the books, is to close it. Block the tools. Require approval. Put a policy in the handbook.

John’s view is that the fear driving that instinct is real but out of date.

“The biggest thing that people were terrified of several years ago was, well, what if my data leaks into a training model and then my competitor can see all of my customers and how much they’re paying us? I don’t think that’s the most likely fear at this point, but that is still where a lot of, especially larger enterprise conservative companies are.”

And the cost of acting on an out-of-date fear is not theoretical.

“We work with some companies where it takes six months to get approval to integrate elvex with another service. If you can’t connect the data and allow the agents to work with those information systems, then you’re really limiting what the capabilities of this technology are, in a way that is hurting your business.”

Six months of approval does not stop anybody using AI. It stops them using it officially. The work moves to the personal account, which is exactly where you cannot see it.

What he describes instead is control that is specific rather than total. An agent that can pull reports from Salesforce, and update records, save changes, and never delete one. Permissions by user or by group, set by an administrator, rather than a yes or no on the whole category.

“That level of governance and control gives companies a lot more confidence in rolling things out, versus an alternative which is just like, anyone can do whatever they want, and who knows what they’re going to do, or what their agents are going to do accidentally.”

After the call, John sent one more thought on this, and it is the sharpest version of the argument anyone has given us.

“People use their personal tools when the company approved platforms don’t really provide anything valuable that they can’t get elsewhere. Fear on its own is not a big enough incentive not to use it.”

“It’s only when the doc your team is working on, the settings for how your team works, the agents other people made that are really dialled in, all that good stuff other people figured out, are immediately and painlessly at your fingertips that people gravitate to the company tool first. And once you have a company platform that is more valuable than the personal plans, everyone will gravitate to it, and then you can easily see what they are doing.”

“Shadow AI is a product problem, not a policy problem.”

Back to Anna

The question was never whether to let her use AI. She already does, and she is better at it than anyone who would be asked to approve it.

But will anyone else ever benefit from what she worked out? Can the company describe it to a client, an auditor or her replacement?

“We need to get away from the ‘everyone has a personal assistant that only knows them and what they do’.”

Which is, when you take the technology out of it, an argument about working in silos. A company that cannot see its own best practices cannot repeat them.

The part that’s hard to plan for

I asked where elvex wants to be in two years.

“It is very hard to look more than three months out a lot of times, because you don’t know what’s around the corner.”

“It could be that a new model is released, and that model is so capable that some of the tooling and the harness no longer needs to be there. And so, great, we can cut out a bunch of code that doesn’t need to be there any more.”

Few founders will say out loud that a good enough model would let them delete part of their own product. Even if, deep down, most of the industry knows it.

Then he went further.

“We think in the next two years you’ll see a momentum shift where most work begins on the AI platforms, and most SaaS tools used today largely become systems of record.”

“Two years from now you and your whole team will be working inside your AI harness of choice, in organised groups and teams, working on collaborative projects, integrated with basically everything you currently have bookmarked in your browser, and assisted by proactive agents that bring things to you instead of only reacting to what you give them. And importantly, it will work well, be reliable, and be secure enough for business.”

Do the hard thing

At the end I asked what he would tell himself twenty years ago, before Parse.ly, before any of it.

“Do the hard thing. It can be very easy to sit back and do the easy thing, or follow the easy path. But especially in startups, doing the things that are really hard, doing the things that are really boring, doing the things that nobody else wants to do, that’s where a lot of the value is uncovered.”

“It may not be the most exciting or sexy thing to talk about all the time. But if you can find the right wedge, that is the thing to lean into.”

John’s answer to the invisible work done with AI is a platform they all work inside, where the building is visible and the permissions are specific.

Art50 does the other half, the duller half, which is a record of what is running, who is responsible for it, and who checked the output, sealed so it can be shown to somebody who was not there. Different jobs. You can do his without ours and ours without his, and most companies currently do neither.

John Levitt is co-founder and COO of elvex, an AI platform for enterprise teams. Before elvex he spent thirteen years at Parse.ly, which was acquired by Automattic in 2021.