Back to blogs

// What's on our mind

Governing AI when anyone in the business can build

Shadow AI and AI sprawl grow when employees build faster than governance moves, and a sanctioned path with tiered review brings that work into view.

Governing AI when anyone in the business can build

You govern AI well when the people building with it can do so in the open.

Shadow AI is already inside your business, and most of it is invisible to the teams accountable for risk. A global study of more than 48,000 people found that 57% of employees hide their use of AI at work and present its output as their own. That number describes a visibility problem before it describes a discipline problem. People are solving real work problems with the tool closest to hand.

Shadow IT taught us that prohibition slows nothing down. The same lesson applies now that a finance analyst can build a working agent before lunch. Governance has to keep pace with citizen building, which means giving people a sanctioned path that's faster than the workaround. We think the goal is simple. Bring the building into the light, and keep the momentum that made it happen.

What shadow AI means once every employee can build

Shadow AI is any AI tool, model or agent used for work without the knowledge or approval of the teams responsible for technology and risk. It covers a public chatbot holding pasted client notes, a browser extension that summarizes email, and a homegrown agent wired to a shared drive.

Picture a wealth operations analyst who spends 2 hours each morning reconciling account notes. She builds a small assistant on a personal subscription, feeds it yesterday's notes, and finishes the job in 10 minutes. Nobody approved it, and nobody knows which client details left the building. Her manager only sees that the work got faster.

That story explains why the definition matters. Shadow AI used to mean someone using a tool. Now it means someone building one, with logic, data connections and outputs that other people start to rely on. You're no longer tracking software purchases. You're tracking small products with no owner, no testing and no record.

Why shadow AI carries more risk than shadow IT

The main difference between shadow IT and shadow AI is what the unapproved tool does with your data. Shadow IT stores or moves information in a place you didn't sanction. Shadow AI reads that information, reasons over it and produces new output that feeds decisions, client messages and code.

Consider an unapproved file-sharing app in a commercial lending team. The exposure is contained, since you can find the files, close the account and move on. Now swap in an unapproved AI assistant that drafts credit memos from those same files. The data has left, and the output is already shaping a lending recommendation that no model risk team has reviewed.

‍

"You're tracking small products with no owner, no testing and no record."

‍

Three things are different. The output is persuasive and sometimes wrong, so errors travel further before anyone questions them. You can't always retrieve what someone has pasted into a public model. Regulators also expect you to explain how you reached a decision, and an analyst's personal assistant won't satisfy an audit. Shadow IT was mostly a security and cost question. Shadow AI adds accuracy, bias and accountability to the list.

How AI sprawl grows inside a regulated business

AI sprawl is the uncoordinated build-up of AI tools, models and agents across a business, each adopted for a good local reason. It grows through 3 doors at once, namely personal accounts, AI features switched on inside software you already own, and agents that employees build themselves.

The scale is hard to see from the top. A survey of 302 cybersecurity leaders found that 69% of organizations suspect or have evidence that employees are using prohibited public generative AI. Walk through a mid-sized insurer and you'll find the pattern. Claims uses one summarizer, underwriting pays for another, and marketing runs 3 agents doing overlapping work on separate budgets.

Sprawl costs you twice. You pay for duplicate licences and duplicate effort, and you lose the ability to answer basic questions. Which tools touch client data? Who owns the agent that emails brokers every Friday? What happens when its builder moves to another team? An inventory that can't answer those questions is only a list.

‍

Shadow AI suspected in most organizationsA survey of 302 cybersecurity leaders found that 69% of organizations suspect or have evidence that employees use prohibited public generative AI.69%of organizations suspect or have evidenceof prohibited public generative AI useA survey of 302 cybersecurity leaders found 69% of organizations suspect prohibited generative AI use.

‍

Why banning unsanctioned AI tools pushes the risk underground

A blanket ban on unsanctioned AI tools reduces what you can see without reducing what people do. Employees who've found a faster way to work will keep using it on personal devices and personal accounts. You end up with the same exposure and far less information about it.

A capital markets desk shows how quickly this happens. Compliance blocks public AI sites on the corporate network on a Monday. Two days later, analysts are pasting research notes into the same tools on their phones. Usage logs now show nothing, and the risk committee reads that silence as success.

Bans also send a message about trust. They tell your most resourceful people that initiative is a liability, and those are the people you want building. A more useful rule restricts specific behaviours and leaves the curiosity alone. Say plainly that client data, material non-public information and credentials never go into an unapproved tool. Then give people somewhere approved to take everything else. People follow clear lines and work around vague prohibitions.

‍

A ban pushes AI use out of sightWhen compliance blocks public AI sites, analysts move to their phones, the usage logs go quiet and the risk committee reads that silence as success.MondayCompliance blockspublic AI sitesTwo days laterAnalysts paste notesinto the same tools on phonesUsage logsnow shownothingRisk committeereads that silenceas successA ban reduces what you can see without reducing what people do.

‍

What a governed path for citizen builders looks like

A governed path is an approved way to build with AI that's quicker and easier than going around it. It gives every employee sanctioned tools, clear data rules and a short route to review. The controls sit inside the path, so following it takes less effort than avoiding it.

A path that works for a business analyst and a risk officer has 5 parts.

‍

"A blanket ban on unsanctioned AI tools reduces what you can see without reducing what people do."

‍

  • Approved tools with enterprise data protection, available to everyone on day 1.
  • A plain-language data rule that says what never goes into a prompt.
  • A register where builders log what they've made in under 5 minutes.
  • Review tiers that match the scrutiny to the risk of each use case.
  • Coaching from real people, sized to each team's daily work.

Coaching is where most of the budget belongs. Tooling is the cheap part, and a licence nobody understands just produces better-funded shadow AI. At Electric Mind, we embed engineers beside business teams so the guardrails live inside the tools people already use. A legal clerk learning to automate contract intake needs a person to ask, and a policy PDF won't do that job. It sounds slower. It ends up faster, since nobody has to rebuild anything after a failed review.

Which AI use cases need review before they scale

Review effort should follow the potential for harm, and most citizen-built tools carry very little. A personal drafting aid needs registration and nothing more. Anything that touches client data, influences a financial outcome or acts without a human check needs a named owner and a proper review first.

These 5 tiers give builders an answer before they start.

‍

What your team is building What it needs before it scales
A personal assistant that drafts or summarizes with no client data The builder registers it and carries on, with no review required.
A team tool that works on internal documents A team lead confirms the data source and names an owner who will maintain it.
A tool that reads client or personal data Privacy and security check where the data goes and how long it is kept.
An output that informs credit, pricing, claims or advice Model risk tests accuracy and bias, and a person signs off each decision.
An agent that takes action in a system of record Engineering rebuilds it on a supported platform with logging and a stop control.

‍

Tiering protects speed. An HR coordinator's interview-scheduling helper shouldn't wait in the same queue as an agent that amends payment instructions. When most ideas clear the first 2 tiers in a day, people stop hiding them. Your review teams also get their time back for the handful of tools that deserve it.

The tradeoff is that somebody has to keep the tiers honest. A tool that starts as a personal aid and ends up feeding a client report has moved up a tier. Its owner needs to say so, and the register needs to make that a 2-minute update.

How to bring shadow AI into the light

You control shadow AI when you make disclosure safe and the sanctioned route worth choosing. Start with an amnesty, ask people what they've built, and treat the answers as a product roadmap. The tools your teams made in secret show you exactly where the work is slowest.

Run the amnesty for 30 days and promise no penalties for anything declared. A regional bank that does this will hear about dozens of small tools, from a collections script to a board-pack summarizer. You'll retire some. You'll adopt, harden and share others. The builders become your first coaches, since they already know the workflows and the shortcuts.

None of it is glamorous. It's a register, a few rules, a review queue that moves, and people who'll sit beside a colleague and show them how. That's how we approach it at Electric Mind, starting with what your people have already built and engineering the path around it. Governance done this way doesn't ask your people to slow down. It asks them to build where you can see, and that's a trade most of them will take gladly.

Got a complex challenge? Let's solve it together, and for real.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Top