Your Employees Are Already Using AI. Do You Know What They’re Putting Into It?

Somewhere inside your business, an employee may already be using an AI tool you never approved.

Maybe purchasing receives a supplier document and uploads it to ChatGPT for a summary.

Someone in sales rewrites an awkward customer email with Claude.

Finance asks an AI assistant to explain a variance.

Operations uploads an Excel export because they need a formula or a quick analysis.

A manager drops meeting notes into an AI tool and gets an action list back thirty seconds later.

They probably are not trying to create a governance problem.

They have work to finish.

The AI makes that work easier.

So they use it.

This is Shadow AI.

And if management's first reaction is:

“We need to stop this.”

I would wait a moment.

Because two things may be happening at the same time.

Your employee may have found a genuinely better way to do part of their job.

And your company may have no idea what information is being shared, which tool is involved, whether the output is reliable, or whether an important business process now quietly depends on someone's personal AI account.

Both can be true.

That is why Shadow AI is much more interesting than a simple IT-security problem.

It is also about operations.

Behaviour.

Management.

And increasingly, about how work gets redesigned from the bottom up before leadership has decided what the future operating model should look like.

What is Shadow AI?

Shadow AI is the use of artificial-intelligence tools for company work without the organisation formally approving, managing or sometimes even knowing about that usage.

At the simplest level, that might mean an employee using a personal ChatGPT account to improve an email.

At the other end, someone may:

• Export company data
• Upload documents
• Run analysis
• Use a sequence of prompts they created themselves
• Manually correct the output
• Use the result in a recurring business process

Then repeat it every week.

At that point, something interesting has happened.

What started as a productivity experiment has quietly become part of the operation.

Except nobody officially designed it.

Nobody documented it.

Nobody decided who owns it.

And management may not know it exists.

That is where Shadow AI becomes important.

The issue is not simply:

“Did someone use AI?”

The more useful question is:

“Has AI become part of how this business works without the business actually knowing how?”

AI adoption is not waiting for your AI strategy

Management teams sometimes discuss AI as though adoption will begin once the company decides to launch an initiative.

Employees have a different timetable.

If something helps them get the job done, some of them will try it.

Research over the last few years has shown just how far ahead employee behaviour can move.

Netskope's 2025 telemetry found widespread generative-AI use through personal or unmanaged accounts.

IBM-sponsored workplace research similarly found heavy employee AI adoption while only a minority of users relied exclusively on employer-provided tools.

Deloitte has also documented employees using their own AI tools alongside company-provided ones and made a particularly useful observation:

Shadow AI can represent unmet demand.

I think that matters.

Because management can interpret the exact same behaviour in two completely different ways.

Version one:

“Employees are breaking the rules.”

Version two:

“Employees found a faster way to solve something before we provided a safe, organised way to do it.”

The first interpretation leads naturally to restriction.

The second still requires controls, but it also creates curiosity.

• What were they trying to fix?
• Why did the approved process feel inadequate?
• What can we learn from what they did?

That is where the conversation gets useful.

Why are employees using unapproved AI?

Usually because it saves them time.

This is the uncomfortable part.

If AI were useless, Shadow AI would be easy to manage.

Tell people not to use it.

Problem solved.

But these tools can take tasks that used to require thirty or forty minutes and reduce them to five.

Drafting.

Summarising.

Translating.

Comparing.

Cleaning information.

Creating formulas.

Extracting key points.

Preparing a first analysis.

Rewriting awkward text.

Turning messy notes into something useful.

Imagine a procurement employee receives a twelve-page supplier document.

Traditionally they might read it line by line, compare it with an older version, highlight changes and prepare a summary.

Instead, they upload both documents and ask:

“Compare these. Show me any changes to price, lead time, payment conditions, minimum quantities and cancellation terms.”

Thirty seconds later, they have a starting point.

From the employee's perspective, that does not feel reckless.

It feels productive.

And sometimes it absolutely is.

People route around friction

This behaviour is not new.

AI is only the latest version.

Employees have always created workarounds when official systems make a simple task unnecessarily difficult.

They build personal spreadsheets.

Write macros.

Create WhatsApp groups.

Send themselves files.

Use Dropbox.

Keep private notes.

Build an Access database nobody knows about.

Store information somewhere more convenient than the official system.

Sometimes those behaviours create risk.

But they also tell you something.

People route around friction.

If the official route takes twenty minutes and an obvious shortcut takes thirty seconds, sooner or later somebody will try the shortcut.

This is important because it changes the management question.

Instead of asking only:

“Why did they go around the system?”

ask:

“What made going around the system so attractive?”

Because banning the shortcut does not remove the reason people wanted one.

Your employee may be solving a productivity problem while creating a management problem

Suppose Sarah prepares a management report every Friday.

She exports data from Odoo.

Cleans it in Excel.

Checks several unusual transactions.

Calculates a few ratios.

Writes the management summary.

Emails it.

It takes two hours.

Then one week Sarah discovers an AI tool.

She builds a prompt.

Refines it over several Fridays.

Now she uploads the export and the AI identifies anomalies, structures the numbers and drafts the commentary.

Sarah reviews it, corrects a few things and sends the report.

Two hours have become thirty minutes.

Fantastic.

Management is happy.

Sarah is happy.

The report arrives earlier.

The business recovered time.

But management does not know that the process now depends on:

• Sarah's personal AI account
• A prompt nobody else has
• An external service nobody assessed
• Company information being uploaded somewhere
• Sarah knowing which parts of the output should not be trusted

What happened?

The business did not remove key-person dependency.

It modernised it.

That is where Shadow AI starts turning into something more serious: Shadow Operations.

When Shadow AI becomes Shadow Operations

This, for me, is the part that gets missed most often.

Shadow AI sounds like:

“Someone used ChatGPT.”

Shadow Operations is different.

It is when unmanaged AI becomes part of a process the company now depends on.

Sales has a personal prompt that prepares every proposal.

Purchasing uses a private AI workflow to compare suppliers.

Finance produces month-end analysis through a personal account.

An operations manager has built an assistant containing years of context.

Customer service has an undocumented process for drafting difficult replies.

Everything works.

Until the person who created it leaves.

Then management discovers the real process was never:

Odoo → Excel → report.

It was:

Odoo → Sarah → personal AI account → five prompts → Sarah's judgment → report.

That is not a controlled company process.

It is personal expertise hidden behind new technology.

AI did not eliminate fragility.

It made the fragility harder to see.

Shadow AI is not only about confidential information

Security matters.

But if that is the only lens management uses, it misses a much broader problem.

There are five risks I would pay attention to.

1. Data risk: what exactly are people putting into these tools?

Employees may paste or upload:

• Customer information
• Supplier pricing
• Contracts
• Employee data
• Product specifications
• Financial information
• Internal emails
• Commercial plans
• Management reports
• Other confidential material

KPMG research has found meaningful levels of unauthorised AI use at work and employees acknowledging that they have uploaded sensitive company information or intellectual property into public AI platforms.

Cisco research has shown a similar contradiction.

People say they are concerned about confidential information being exposed through generative AI.

Then some of those same people enter confidential information anyway.

That is not necessarily hypocrisy.

It is behaviour under pressure.

The risk is abstract.

The task is immediate.

The customer needs an answer.

The manager needs the report.

The contract needs reviewing.

Convenience wins.

The risk is not “AI steals your data”

We should be precise.

Different AI products, plans and providers have different retention policies, training policies, enterprise controls and security arrangements.

A managed enterprise AI environment is not necessarily equivalent to an employee signing up to a random AI service using a personal email address.

The real Shadow AI problem is:

Management may not know.

• Which tool?
• Which account?
• What information?
• What retention policy?
• What processing?
• Which third parties?
• Which controls?
• Was the employee even authorised to share that information?

NIST's Generative AI guidance highlights the risks associated with third-party AI systems, confidential data, privacy and intellectual property.

OWASP similarly identifies sensitive-information disclosure as an important risk around LLM applications.

The practical question for a business is therefore not:

“Is AI safe?”

That question is meaningless without context.

Ask:

“Is this information appropriate to put into this tool under these conditions?”

Much more useful.

2. Accuracy risk: where does the AI output go next?

An AI gives an answer.

Then what?

That's the question.

An employee asks the model to summarise a contract.

The summary misses a clause.

Purchasing trusts it.

Someone asks AI to analyse an operational spreadsheet.

One assumption is wrong.

The result goes into the management pack.

Customer service asks for help drafting a reply.

The model invents a detail.

The employee does not notice.

KPMG research has found employees acknowledging that they sometimes rely on AI output without thoroughly evaluating it.

That is hardly surprising.

If something is correct again and again, human beings naturally become less suspicious of it.

The management issue is therefore not simply:

“Can AI make mistakes?”

Of course it can.

The better question is:

“Where can an AI mistake become a business mistake?”

An internal brainstorming prompt?

Low consequence.

A customer contract?

Very different.

A banking instruction?

Different again.

The level of verification should follow the consequence.

3. Process risk: what happens if the person or the tool disappears?

This is the Shadow Operations problem.

An employee creates a brilliant workflow.

Nobody documents it.

Management begins relying on the result.

Then the employee leaves.

Now:

• Where are the prompts?
• Where is the conversation history?
• Which documents were uploaded?
• Which outputs were manually corrected?
• Which assumptions were based on experience?
• Which steps existed only in the employee's head?

Nobody knows.

A company process should not disappear when someone closes their personal account.

Once an AI workflow matters to customers, finance, management or operations, it should stop being only a personal trick.

4. Consistency risk: five employees, five versions of the process

Suppose five employees use AI for similar work.

One uses ChatGPT.

Another uses Claude.

Another Gemini.

Another found a specialised platform.

Another uses something they personally pay for.

They use different prompts.

Different source files.

Different judgement.

Different standards.

One removes names.

Another uploads everything.

One checks every figure.

Another assumes the output is correct.

Individually, all five people may be working faster.

Organisationally, the business now has five different operating methods.

And nobody consciously chose any of them.

That matters.

Because businesses eventually need consistency.

Customers should not receive completely different quality depending on who happened to use which prompt that morning.

5. Management risk: nobody actually knows how AI is affecting the business

Eventually the owner asks:

“Where exactly are we using AI?”

Marketing uses something.

Sales definitely uses ChatGPT.

Someone in finance mentioned AI.

Operations built an assistant.

Purchasing might be uploading supplier quotations somewhere.

What tool?

No idea.

What information?

No idea.

How important is it?

No idea.

Does anyone verify the output?

Probably.

That is the management problem in its simplest form.

If you cannot answer:

• What AI is being used
• By whom
• With what information
• For what purpose
• With what level of human checking

then you do not really have an AI operating model.

You have AI happening to you.

Why people know the risk and use the tool anyway

This behaviour can look irrational.

It usually isn't.

An employee has been told:

“Do not put confidential information into public AI tools.”

Fine.

Six weeks later, their manager wants a supplier-contract comparison by 2 p.m.

There are two forty-page documents.

The employee knows AI can compare them in seconds.

So they think:

“I'll just remove the supplier name.”

Or:

“Everyone uses ChatGPT.”

Or:

“This probably isn't sensitive.”

Or simply:

“I need this finished.”

This is an important behavioural point.

The risk is delayed.

The benefit is immediate.

If the official way of working is materially slower than the unofficial one, company policy is fighting the employee's day-to-day incentives.

And incentives usually win eventually.

If your AI policy is twelve pages long and the shortcut takes thirty seconds…

…I have bad news about which one is winning.

Policies are necessary.

But a policy is not an operating solution.

If management says:

“Do not use unapproved AI with company information,”

employees also need an answer to:

“What should I use instead?”

If there is no answer, the rule has created a gap rather than solving one.

People still have the job to do.

Some will go back to the manual process.

Some will find another AI tool.

Some will simply stop telling management.

None of those outcomes is particularly attractive.

Banning AI solves the tool problem. It may not solve the work problem.

There are absolutely situations where a tool or behaviour should be blocked immediately.

But suppose an employee was using AI because preparing a report manually takes three hours.

You block the tool.

Great.

The three-hour report is back.

If someone used AI because analysing supplier documents is painfully slow, the documents still need analysing.

If they used it because company knowledge is impossible to search, the information is still difficult to find.

The tool disappeared.

The operational problem did not.

This is why a good response to Shadow AI has two tracks:

Control the risk.

And:

Understand why the behaviour existed.

Do only the first and you may simply move the behaviour somewhere less visible.

Shadow AI may be the best process-discovery exercise nobody planned

This is where I think Shadow AI becomes extremely valuable.

If several employees independently start using AI for certain tasks, they are telling you where friction exists.

• Why is sales rewriting the same type of customer email?
• Why is purchasing comparing supplier documents manually?
• Why is finance analysing recurring spreadsheets through AI?
• Why is operations constantly creating formulas?
• Why are people summarising documents?
• Why are employees creating their own assistants?

Each use case is a clue.

The employee has identified work that feels:

• Slow
• Repetitive
• Information-heavy
• Difficult
• Unnecessarily dependent on them

That is valuable operational information.

Instead of asking only:

“How do we stop this?”

ask:

“What was this employee trying to make easier?”

That question can uncover a much larger improvement opportunity.

Some Shadow AI should disappear. Some should become official.

Not every unofficial AI use deserves the same response.

There are at least four categories.

Stop it

The use creates unacceptable data, legal, financial or security risk.

Fine.

Stop it.

Allow it with simple rules

Low-risk productivity uses may need little more than approved tools and basic guidance.

No need to create a committee.

Move it into a managed environment

The use is valuable but involves internal data or important business work.

Give people a safer, approved way to do it.

Turn it into a proper workflow

The employee has accidentally discovered a repeatable process improvement.

Now stop treating it as a personal trick and design it properly.

This last category is where operational value can become significant.

If employees are using AI to fix a process, investigate the process

Imagine purchasing uses AI every day to extract information from supplier documents.

Management could say:

“Stop uploading those.”

Maybe that is necessary.

But then keep going.

• Why does purchasing need to extract the information manually?
• How many documents are there?
• Are they similar?
• Where does the information go afterwards?
• Does someone type it into another system?
• Does someone check it again?
• Could the whole process be redesigned so employees do not need to manually upload anything at all?

Now the unofficial AI use has shown you something interesting.

A workflow worth fixing.

This is why Shadow AI can be a form of free operational research.

Employees are showing you which work they would happily remove from their day.

Pay attention.

A simple Green, Amber, Red approach is often enough to start

An SME does not necessarily need an eighty-page AI-governance framework.

Employees need to understand what is safe, what needs care and where they should stop.

One simple management approach is:

GREEN: Use it

Low-risk work.

Brainstorming.

Generic rewriting.

Creating templates.

Explaining formulas.

Structuring public information.

Tasks where confidential company data is not involved.

Let people experiment.

AMBER: Use it with controls

Internal documents.

Supplier information.

Customer correspondence.

Operational reports.

Commercial analysis.

Company data.

Use approved tools.

Understand what information is being submitted.

And verify important outputs.

RED: Do not casually put this into an external AI tool

Passwords.

API keys.

Bank details.

Highly sensitive personal information.

Privileged legal material.

Highly confidential contracts.

Restricted customer information.

Sensitive employee information.

Whatever else the company's risk profile makes inappropriate.

The exact boundaries depend on the company, contracts and jurisdiction.

This is not universal legal advice.

The principle is simply:

Make the rules understandable enough that an employee can actually use them on Tuesday afternoon.

Governance should follow consequence

Not every AI prompt needs approval from the CEO.

That would be an efficient way to make sure the policy gets ignored.

If somebody asks:

“Give me five possible headings for this presentation.”

Relax.

If somebody uploads payroll information and asks:

“Which employees should we fire?”

We have moved into a very different category.

Good governance is proportional.

The more sensitive the information and the greater the consequence of a wrong output, the stronger the controls should be.

The same is true for human review.

Generic internal copy?

Low consequence.

Contract interpretation?

Higher.

Customer pricing?

Higher.

Financial decisions?

Higher again.

One company-wide rule for every AI use is usually either too weak or too restrictive.

Experimentation and production are not the same thing

Employees need room to test.

Try prompts.

Experiment.

See where AI helps.

That is healthy.

But once an AI-assisted workflow becomes something the company depends on, it has crossed a line.

At that point ask:

• Can someone else reproduce it?
• Who owns it?
• Is the tool approved?
• What data does it use?
• What happens when the employee leaves?
• What happens if the AI service is unavailable?
• Which outputs require human verification?
• Can management see what is happening?

If you cannot answer those questions, the experiment may have quietly become infrastructure.

And infrastructure deserves more attention.

Turn good employee hacks into company capability

Suppose Sarah reduced a two-hour process to twenty minutes.

Excellent.

Do not punish Sarah for being clever.

Study what she built.

• Which steps genuinely require her expertise?
• Which are repetitive?
• Can the process use a managed company account?
• Can data flow directly rather than being manually uploaded?
• Can the prompt logic be standardised?
• Can other people use it?
• Can important exceptions be routed properly?
• Can the result be monitored?
• Can the process survive Sarah going on holiday?

Now the business has done something much more valuable than giving Sarah an AI subscription.

It has taken an individual productivity gain and turned it into organisational capacity.

That is the real opportunity.

Do not kill the behaviour you want

There is a cultural risk here.

An employee experiments with AI.

Finds a genuinely smart way to improve a process.

Management discovers it.

The response is:

“You violated policy. Stop.”

What does everybody learn?

Do not experiment.

Do not tell management.

If you discover something useful, keep it quiet.

That is probably not the culture you want.

Serious security or data breaches obviously need to be handled appropriately.

But management can separate intent from method.

A much better response may be:

“This is a clever idea. We cannot use company information this way. Let's find a safe way to achieve the same result.”

You kept the initiative.

And fixed the control.

Much better.

Your employees do not need to become your unofficial AI department

There is also a limit to bottom-up experimentation.

Employees should not have to decide:

• Which AI vendors are safe
• What data can be shared
• How prompts are stored
• Which outputs require checking
• What happens if an employee leaves
• How important AI workflows integrate with company systems

That is management's responsibility.

Employees can show you what is possible.

Management needs to decide what the company can rely on.

That distinction matters.

From Shadow AI to operational capacity

The purpose of AI should never be:

“We use AI.”

Nobody cares.

The useful question is what changes in the business.

• Does a process require less manual work?
• Can people find information faster?
• Are supplier exceptions discovered earlier?
• Can the same team handle more volume?
• Does management get better visibility?
• Do fewer questions require senior intervention?

That is where AI becomes commercially meaningful.

In one STREVIO wholesale operation, changing how supplier tracking and operational information moved through the business reduced procurement and supplier follow-up by around 60%, recovered more than three hours per day, and enabled the same team to handle roughly 35% more orders.

That is operational capacity.

Not:

“We gave everyone an AI assistant.”

But:

“The team no longer spends three hours every day doing work that did not require them.”

Shadow AI often begins with that same instinct.

Employees are trying to remove work.

The opportunity is to turn that instinct into something safe, repeatable and scalable.

The business needs guardrails, not handcuffs

Employees should know:

• Which AI tools are approved?
• Which information is restricted?
• Which outputs require checking?
• When should I ask?
• Where can I share a useful AI experiment?

That is enough to begin.

Do not create so much bureaucracy that the unofficial route becomes more attractive again.

Because then you have not eliminated Shadow AI.

You have simply trained it to hide better.

The real question is not “Are our employees using AI?”

They probably are.

Or they will.

Ask instead:

“Do we understand how AI is changing the way work gets done inside our business?”

That is where opportunity and risk exist together.

Employees can use AI to recover capacity.

They can also create undocumented processes.

They can improve decision speed.

They can also introduce errors.

They can reduce repetitive work.

They can also create new key-person dependency.

They can discover better workflows.

They can also send company information somewhere management never intended.

The job of leadership is not to pretend one side does not exist.

It is to manage both.

Shadow AI may be telling you something uncomfortable about your company

If employees repeatedly look outside approved systems to get work done, perhaps their behaviour is not the only thing worth investigating.

Perhaps the business is asking them to tolerate too much friction.

A report that takes three hours.

Information nobody can find.

Documents employees manually compare.

Data they repeatedly clean.

Customer emails they rewrite every day.

Spreadsheets they reconstruct every week.

People look for leverage.

AI gives them some.

So pay attention to where they use it.

Because that unofficial AI tool may be telling you something your process map never did:

This work is harder than it needs to be.

And that is valuable information.

Start by making the invisible visible

If you want to understand Shadow AI inside your company, do not begin with an investigation.

Ask people.

Something as simple as:

“What AI tools are you currently using to make your job easier?”

Then:

“What do you use them for?”

You may discover things that need stopping immediately.

You may also discover some of the best process-improvement ideas in the company.

Then classify the uses.

• Which ones are low risk?
• Which involve sensitive business information?
• Which influence important decisions?
• Which save meaningful time?
• Which already form part of an important workflow?
• Which reveal a process that should be redesigned?

Then provide approved options.

Set simple rules.

Define where checking matters.

And give people somewhere to ask when they are unsure.

Do not begin by trying to build perfect AI governance.

Begin by understanding reality.

Do you know where AI is already changing your operations?

Shadow AI is often discussed like something coming in the future.

For many businesses, it is already happening.

Employees are experimenting.

Learning.

Building shortcuts.

Some are excellent.

Some are risky.

Some are probably both.

The first management responsibility is visibility.

Not surveillance.

Visibility.

• What are people using?
• Why?
• What problem are they solving?
• What information is involved?
• How important has the workflow become?
• What happens if the answer is wrong?
• What happens if the tool disappears tomorrow?
• What happens if the employee leaves?

Once you can answer those questions, you can manage AI sensibly.

Without them, you are simply hoping every employee independently makes exactly the right decision.

That is not governance.

That is luck.