The Shadow AI Finance Team: How Adoption Spreads

Every shadow AI finance team starts the same way. I used to work at much larger companies, and I remember how IT told our finance team we were “building unauthorized applications.”

How a shadow AI finance team starts

The applications? Excel dashboards. Financial models. A forecast with macros.

Their solution was an approval process. We submitted requests. Nothing got approved. Six months later we were still waiting on a ticket for a pivot table with a database connection.

So we kept building in Excel. Every finance team does.

That’s how shadow IT actually starts. Not with some rogue developer spinning up AWS instances.

With a finance analyst who needs an answer by Friday and can’t wait three months for IT to scope a solution.

Every company has it. Not the scary kind. The quiet kind.

The forecast model that runs payroll planning. The Access database from 2018 that nobody admits is load-bearing infrastructure. The Google Sheet with 67 tabs that’s basically your operating model.

These exist because the official process never worked. Because we had archaic tools that were not serving the purposes of the business.

Here’s what changed: AI made this significantly worse.

Teams aren’t just building Excel models now. They’re writing Claude prompts that pull data, building automation scripts that move information between systems, creating Notion workflows that replace what should be governed processes.

It’s a new layer of shadow IT on top of the old one. Smarter, faster, and much harder to audit.

At a bigger company, IT blocked everything. It left a bitter taste because the blocking never came with a better alternative. Just “no” and a ticket number.

Now I’m at a smaller company. And I still don’t have the answer.

But I have an obsession: my job is to enable my team to use these tools in the most secure way possible. That means removing access to private and confidential information. Making sure teams are properly trained. And involving IT for anything that could be compromised.

Including myself. I’m not exempt from this.

Full access means every team builds whatever they want. That’s a security and compliance nightmare. Lock it down and you kill the thing that makes teams fast and resourceful. The same tension I felt in finance fifteen years ago, just with better tools.

Twenty years of IT telling finance “you can’t build that” didn’t stop a single spreadsheet.

Blocking AI won’t work either.

The AI shadow role

A member of our leadership team calls me.

He has two org charts. One shows the holding company and the three owners. The other has every operating LLC, lines all messed up. He needs them merged into one clean ownership structure.

A few weeks ago he’d have sent it to our finance analyst. I offered exactly that.

He said no. He wanted to learn Claude.

So I stopped being the CFO and became something else. Walked him through Claude Desktop. Created a Cowork project. Dropped both PDFs in. Coached him through the prompt. By the end of the call, Claude had read both diagrams, identified the owners, traced every subsidiary, and produced the right chart.

His reaction: “It did analyze all that and come up with the correct org chart.”

Then the line I keep thinking about: “If this works as well as I hope, you will get me to convert and begin using it for other things.”

That’s the AI shadow role. Not replacing the expert. His decades of domain experience are what made the output recognizable as correct. Not making him learn AI alone either. You pair the expert with someone fluent for one tedious deliverable until the gap closes.

The on-ramp is a chore they already needed done.

AI doesn’t cross silos. You have to force it

Someone I respect watched me demo my AI setup last month.

Five Claude projects running at once. Dashboards, email triage, reporting, all built myself.

Then he said the thing I didn’t want to hear. “You’re automating things that sit on your desk. You’re automating for yourself.”

He was right.

I’d told myself we were building an AI company. What we’d actually built was one motivated person automating his own desk. I stopped asking my team for things and just did them myself. That isn’t progress. It’s a silo with better tools.

The real work still sits between functions. We’ve got 40 people whose whole job is negotiating files with law firms. AI would do a lot of that better. It’s still manual, because crossing that line means redrawing processes our old systems can’t change fast.

AI doesn’t cross silos on its own. It sits inside whatever box you drop it in. Org charts are built around human teams, not data, so the AI inherits the walls.

Getting it across isn’t a tech problem. It’s a management one. Someone has to redraw the process, not just the desk.

Before you build the agent, document the human

Our CEO spent 68 minutes last week pushing the team on the part of AI implementation everyone wants to skip.

Writing down what every role actually does.

Not the policies. Not the procedures. The actual user experience. Every seat in the org, internal and external. Intake. Risk. Case management. The law firms we work with. The healthcare providers.

His argument, close to verbatim: if you can’t tell the agent what the role is, it’ll make up assumptions, and they won’t be helpful. Skip this and you’re building on a weak foundation.

Almost nobody on LinkedIn talks about this part.

The market is loud about output. “Agents reviewed my site while I slept.” “Build an AI agency, charge $5K a month.”

Almost nobody mentions the input. What the agent’s been trained on. What it actually knows about the job it’s automating. Because that part is boring. Sitting with the people who do the work. Writing down every step. Capturing the wishlist they’ve never said out loud.

We left the meeting with three documents in Notion. One per audience. The agents we’re building this quarter will be loaded with them alongside the code and the SOPs.

If your team can’t describe what the role you’re automating actually does on a Tuesday afternoon, that’s the project. Not the agent.

The cheapest AI move is training your people

Everyone is asking the same question right now: how do we incorporate AI into the company.

Most of the answers are expensive. Build agents. Rewire Salesforce. Stand up an internal platform. Hire a head of AI.

The cheapest move is the one nobody leads with. Train your people. Then get out of the way.

Four months ago I sat down with someone on my finance team and walked him through Claude Code. He wasn’t an engineer. He just lived in the workflow.

Last month he showed me what he had built. Our cash receipts process used to take three hours a day. He fed Claude two weeks of his own training transcripts and it wrote the skill for him. Today the work takes minutes. The version we’re piloting now does it in seconds.

I’ve since run sessions for our marketing team, our HR team, our account managers, our legal team. I’ve done it for peer CFOs at other companies. Claude Cowork came out and made the onboarding bar dramatically lower, which means I can now train people who would never have touched Claude Code in a hundred years.

Same shape every time. Sixty minutes hands-on. Ninety days later, the wins aren’t where I expected them to be. They show up wherever someone closest to a tedious workflow finally has permission to fix it.

A 30-year compliance veteran I work with, the kind of person who’d never describe himself as technical, is now using Claude to pressure-test legal positions before he sends them.

A peer CFO I coached is generating investment memos in a fraction of the time his analysts used to take.

This isn’t a “no top-down” story. We picked the tool. We set up security. We provisioned the data layers. That part matters. But once the rails are in place, the wins don’t come from engineering. They come from the person closest to the grind finally having the keys to fix their own workflow.

The bottleneck in most companies isn’t the tool. It’s top-down work done properly, paired with real enablement. Everyone is hyping Cowork right now. Claude Code does the same thing for anyone willing to spend an hour with it. The brand is loud. What matters is whether your people actually use it.

If you’re an executive trying to figure out where to start with AI, start here. Train them. Then enable them. The ROI shows up faster than any platform project you’ll greenlight this year.

”It’s not our job to use AI”

Someone in the company put it in writing last week: “It’s not our job to use AI.”

Same week, the message from the top was the opposite. Use AI. Accelerate. The people who won’t are the ones who get left behind.

So you’ve got a mandate pushing one way and a team digging in the other, everyone waiting to see which signal is real.

Here’s what I’ve learned watching this play out: AI adoption almost never fails on the tool. It fails on two things.

First, mixed signals. Tell a team to adopt AI, then side with the person who refused the first time a deadline slips, and you just taught everyone the mandate was optional. People read what you reward, not what you announce.

Second, nobody showed the skeptic why. A memo saying “use Claude” isn’t training. It’s a to-do. So the person who’s done the job the same way for years reads it as tech telling them how to do their job.

The fix isn’t another email. It’s a white-glove session. Sit with the skeptic, pick one real task they own, run it through Claude together. Let them feel the time it saves before you ask them to believe in it.

Most “they won’t use AI” problems are really “nobody sat down with them” problems.

From AI curious to AI doer

One of our sales leaders built the app that ran our event last week.

Not a side project. The thing that ran the door. 112 people checked in. Every email captured to a Google Sheet that hit his inbox before I’d finished breakfast.

He’s not an engineer. He covers law firms. His description in our morning standup, verbatim: “I used Claude and it built most of it and gave me all the code. Then I deployed it with Vercel.”

I missed the event because I wasn’t feeling great. Saw this in the standup the next morning. On paper my job includes the AI rollout. And someone on our sales team built the thing the event ran on, in a day, while I was home with the flu.

Strange feeling. The good kind.

Here’s what changed. We stopped treating AI as a separate initiative and started baking it into the actual work. We train people. We hand them the tools. We get out of the way and let them build.

A year ago most of our team was AI curious. They’d ask what Claude could do. Now they’re AI doers. They open it, they build, they ship. The person who runs a workflow builds the tool for that workflow, because nobody’s faster than the operator who lives in it.

The magic isn’t one app working at one event. The magic is that reaching for Claude has stopped being a special occasion on our team. It’s just how the work gets done now.

None of this is “AI replaces developers.” All of it is “the gap between idea and working tool collapsed for the people closest to the work.”

AI adoption is now a retention signal

I ran an AI workshop for ten people on my team six months ago. Two picked it up. Eight didn’t.

One of the two was already on our worst manual process. Daily check processing, paper checks, allocated by hand. A month after that workshop, he’d built a tool with AI and OCR that automated the entire workflow. We connected him with the tech team. They turned it into an enterprise solution that runs at scale.

He didn’t wait for someone to redeploy him. He redeployed himself.

I was talking with another finance leader recently. He put it simply: “There are two types of employees. Ones that lean in and ones that wait for you to tell them to lean in.”

He told his 20-person team he wants it to be three. Not by firing anyone. By redeploying people elsewhere in the company. “What I won’t do is pay for manual work that adds no value.”

Nobody announced this. No policy changed. But the pattern is the same: the ones who lean in without being told end up on the higher-value work.

AI adoption isn’t a training checkbox anymore. It’s a retention signal.

Related

• More on the AI-native CFO
• AI readiness in the finance team
• How CFOs measure AI ROI
• Claude Cowork cheat sheet