←
AI-Ready Foundation
Aware · M3 · lesson 3 of 5 · queued
Preview — browse every lesson free. Enroll to mark lessons complete, open partner links and save your progress. Login & enroll →
Shadow AI: How to Find and Govern the AI Your Employees Already Use
📖
now learning

Shadow AI: How to Find and Govern the AI Your Employees Already Use

15 min

Shadow AI is any AI tool, AI feature or AI agent used for work that your organization has not approved and cannot see. It is in your building right now, used by your most diligent people, and it is moving company data to third parties under terms nobody has read. In 2026 it also includes assistants and browser extensions that staff have connected to their company mail, calendar and files with one click. This lesson shows you how to detect it in one week without frightening anyone and govern it with a register, a one-page data rule and one safe tool.

This is lesson 1 of the AI-Ready Foundation program, and the first of five lessons at Level 1 (AI Aware) on a ladder that runs from L1 to L5 (AI Native). Nothing comes before it, so you need no preparation. You will finish holding two things: a Shadow AI Register and a one-page Data-to-Tool Rule. The next lesson, the AI data inventory for mapping and classifying your data, picks up the one column of the register you will struggle to fill in.

The Shadow AI Problem and What It Costs

Start with an ordinary Tuesday. A sales coordinator has a call in twenty minutes and a 40-page customer contract she has not read. She pastes it into a free chatbot on her personal account and asks for a summary. Two floors up, a developer pastes a failing function and its stack trace into the same kind of tool. Down the corridor, a manager adds a meeting bot that now joins every call in his calendar and writes the minutes. And an account manager, tired of her inbox, clicks "Allow" on a screen that lets a personal AI assistant read and send mail from her company account.

None of them thinks they did anything wrong. Each just did the job faster, and each just handed company data, or a standing key to it, to an outside firm whose terms no one in your company has opened.

That is the whole problem in four people. Now see what the simplest version did to one of the world's largest electronics makers.

What happened in the Samsung ChatGPT leak?

In spring 2023, Samsung employees reportedly pasted confidential source code and the contents of a meeting into ChatGPT, and the company responded by banning generative AI tools for work. The order of events is the lesson, so take it slowly.

  1. The door was opened. Samsung had restricted generative AI, then lifted that restriction earlier in 2023, according to The Register (2 May 2023).
  2. Helpful people walked through it. Korean media then reported that several employees had uploaded sensitive material. One pasted confidential source code while trying to fix a faulty database. Another shared an entire meeting so the tool could produce minutes, as Fortune reported (2 May 2023). The Register adds that the reported material included equipment measurement and yield data from the semiconductor unit.
  3. The door was slammed. On Monday 1 May 2023 a staff memo banned generative AI tools for work. It said "there are growing concerns about security risks presented by generative A.I." and warned that violations could bring "disciplinary action up to and including termination of employment".
  4. A second thought arrived. A revision called the ban temporary and said Samsung would build its own tools for translation and document summarization.

Be honest with your steering committee about what is known. The incident detail rests on Korean media reports, which, The Register notes, Samsung neither confirmed nor denied. There is no audited cost and no public evidence that the data ever resurfaced. The memo is solid. The rest is reporting.

Even so, three things stand out. The leak came from employees fixing a database and writing up a meeting, not from attackers. Access was opened with no data rule and no safe default, and the reported leaks followed within weeks. And the ban was an over-correction that even Samsung paired with a plan for an internal tool: the company that banned hardest still concluded people needed something to use.

Samsung was not a careless outlier either. Fortune's report notes that Amazon warned staff in January 2023 not to upload confidential information, and that Citigroup and Goldman Sachs restricted chatbot use in February 2023. This is a pattern.

Samsung's case is the 2023 shape of shadow AI: a person pastes data into a chatbot. The 2026 shape is quieter. An extension that can read every page you open, or an agent holding delegated access to a mailbox, does not wait for anyone to paste anything. Keep both in mind.

How big is the problem in numbers?

Four studies give you the scale. Quote each caveat with its number, because your CFO will ask.

What was foundPublisher and dateRead it with this in mind
The global average breach cost USD 4.99 million, a record, up 12% from USD 4.44 million in 2025. Coverage of the report says workers using unapproved AI tools featured in 43% of incidents, more than double the 2025 share; 92% of organizations hit by an AI-related incident lacked proper AI access controls; and close to seven in ten breached organizations had no AI governance policy in place.IBM Cost of a Data Breach Report 2026 (29 July 2026), 602 breached organizations, research by Ponemon Institute; AI figures as reported by Help Net Security (30 July 2026)IBM sells security and governance products. The AI percentages come from the gated report via press coverage, not from IBM's public summary.
34.8% of the corporate data employees put into AI tools was sensitive, up from 27.4% a year earlier and 10.7% two years earlier. Of the sensitive data, source code was 18.7%, R&D material 17.1%, sales and marketing data 10.7%.Cyberhaven Labs, April 2025, telemetry from 7 million workersCyberhaven sells data loss prevention, the sample is its own customers, and the risk ratings use its own rubric.
About 75% of knowledge workers used AI at work. 78% of AI users brought their own tools (BYOAI, bring your own AI). 52% were reluctant to admit using AI for important tasks. Only 39% had received training.Microsoft and LinkedIn Work Trend Index, 2024 edition (May 2024), about 31,000 people in 31 countriesSelf-reported, Microsoft sells an AI assistant, and Microsoft has published newer editions since.
More than half of employees would find and use alternative AI tools anyway if the company did not supply what they needed.BCG AI at Work survey, June 2025, more than 10,600 respondentsA survey of stated intent, not observed behavior.

Read the IBM line carefully, because older slides will contradict it. IBM's 2025 edition estimated that a high level of shadow AI added USD 670,000 to the average breach, as IBM's X-Force summary (30 July 2025) put it. The 2026 coverage gives no equivalent premium, so do not present USD 670,000 as a current figure. The current headline is the share: from one in five incidents to more than two in five in a year.

Then read the Cyberhaven line as three numbers: 10.7%, 27.4%, 34.8%. On that vendor's telemetry, the sensitive share more than tripled in two years. Put the four rows together and the conclusion is plain. The question is not whether shadow AI exists in your firm. Assume it does, and measure it.

Why do sensible organizations get this wrong?

They make one of five mistakes, and you will recognize at least two.

  • They treat it as a discipline problem. The people involved are the keen ones trying to work faster. Punish them and you teach your best staff to hide.
  • They ban. Practitioner guidance from Tines (a workflow automation vendor, so informed opinion, not a standard) is that blanket blocking pushes use onto phones and personal accounts where no log sees it.
  • They offer no sanctioned alternative, or a slow one. The same source gives the example of a three-week approval for a summarization tool. A process like that teaches people to skip the process.
  • They look only at chatbots. They miss browser extensions, meeting bots, the AI features switched on inside software they already pay for, and the agents and connectors staff have authorized against company accounts.
  • They ask managers. If 52% of AI users in the Microsoft survey were reluctant to admit it, then "does your team use AI?" returns a wrong answer, delivered with confidence.

Underneath all five sits one more cause: nobody owns it. Shadow AI governance should sit with the CISO or head of IT security. The CIO should be accountable for supplying a safe alternative. Privacy or legal should be consulted. If you cannot name those three people today, that is your first finding.

How Shadow AI Works: The Kitchen With No Kettle

Three ideas first, each with a picture you can repeat in a board meeting.

Picture an office kitchen with no kettle. People want tea, so they bring kettles from home and plug them in under their desks. Facilities finds out and is alarmed: untested appliances, a fire risk nobody signed off.

Facilities has two options. It can confiscate kettles, and next week there are travel kettles in desk drawers that it cannot see at all. Or it can count the kettles, put one tested kettle in the kitchen, and post a one-line rule about what may be plugged in where. The kettles under the desks disappear, because nobody needs them any more.

Idea 1: Is shadow AI a compliance problem or a supply problem?

It is a supply problem first. People have a job to do, a tool does part of it in seconds, and the company has not handed them one, so they supply themselves.

That tells you where to spend. A compliance problem is fixed with rules and enforcement; a supply problem is fixed by supplying something. A policy with nothing supplied is confiscated kettles and an empty kitchen.

It also changes how you read your findings. Every unapproved tool you discover is a demand signal, which is simply evidence of a job people want help with. Forty people in customer service drafting replies with a chatbot is not forty offences. It is a free requirements study for your approved tool.

Idea 2: Is the risk the tool or what the tool can reach?

The risk is what the tool can reach, and there are two ways it reaches company data. The first is what people hand it: the same chatbot is harmless when someone asks it to reword a public job advert and dangerous when someone pastes in a customer contract. The second is what it can take for itself: its permissions and its vendor's terms.

For the first route, the control is a rule about data classes, which are a small number of plain categories that say how careful you must be with a piece of information. This lesson uses four, taken from the Tines guidance: Public, Internal, Sensitive and Prohibited. (Prohibited here is a data class, not a legal term; do not confuse it with the practices the EU AI Act bans outright.) The rule then says, for each class, which tools may receive it.

For the second route, the data class does not help, because nobody pastes anything. An extension allowed to "read and change all data on all websites" sees your HR system whatever its owner meant it for. A meeting bot set to join every invite hears the disciplinary hearing as well as the stand-up. And delegated access, usually granted through OAuth (the consent screen that asks "Allow this app to read your mail?"), lets an outside app act as the employee against company systems, often from the vendor's cloud and usually until someone revokes the grant. Add vendor terms that allow training on your inputs with no opt-out, and some tools are unacceptable at any data class.

In kitchen terms: the rule is not "no kettles" but "kettles go in the kitchen socket, nothing goes in the server room socket". Some appliances fail whatever socket you choose: a kettle with a frayed cable, or one that arrives with a copy of the office keys. Agents with delegated access are the second kind, and an agent that reads untrusted email while holding those keys can be steered by text hidden in a message, which the lesson on defending against prompt injection and oversharing covers at Level 3.

Idea 3: Why does discovery have to be safe for the people discovered?

Because otherwise you only find what people are willing to show you. If staff believe the survey is the first step toward a disciplinary list, they answer "none", and you get a tidy, reassuring and false picture.

That is why the method below starts with an amnesty: a written promise from a senior leader that, for a fixed period, telling the company what you use carries no consequence. It is also why you look at totals per website and per app before you look at any named person. The aim of discovery is a count of kettles, not a list of culprits.

Count the kettles, check what each one is wired into, put a tested one in the kitchen, and post one line about what may be plugged in where. That is the whole of shadow AI governance.

The Workshop: Detect Your Shadow AI and Build the Register in One Week

This is a workshop, not a project: a 60-minute kickoff, a 5-day survey window and one 2-hour triage session. It answers the question of how to detect shadow AI by combining three sources that correct each other: what people tell you, what the network logs show, and what your identity provider says has been granted access to company accounts.

Who is in the room, and what do you bring?

People: the security lead, who owns the exercise; IT operations, with the proxy and DNS logs; the identity administrator, who exports app consent grants; an engineering lead, who lists tokens and apps connected to source control; privacy or legal, who checks the monitoring and reads vendor terms; and HR or internal communications, who sends the amnesty. Add one manager from the heaviest-using team to explain the tasks; the survey does the rest better than a room of managers.

Inputs: 30 days of proxy or DNS logs, the single sign-on application list, an export of OAuth consent grants from your identity provider, SaaS expense claims, and a browser extension inventory from device management. Three of the five is enough to start.

The seven steps

  1. Name one owner and announce an amnesty. One named person owns shadow AI from today. Then a senior leader puts the amnesty in writing: for the next 30 days, telling us what you use carries no consequence. A version you can adapt: "We know many of you use AI tools to get work done. We are not looking for anyone to blame. For the next 30 days, nothing you tell us about the tools you use will be held against you. We want to know what you need so we can give you a safe version of it."
  2. Run an anonymous five-question survey. Keep it to five so people finish it:
    • Which AI tools, extensions or assistants do you use for work?
    • On what account: personal or company? Have you connected any of them to your work mail, calendar or files?
    • For which tasks?
    • With what kind of information?
    • What would you need in order to stop using a personal tool?
    The fifth question writes your requirements for you.
  3. Pull 30 days of logs and count. Count distinct AI domains and rough user numbers for each, aggregated by domain. Before anyone looks at named individuals, check with privacy counsel and any employee representatives.
  4. Add the quiet sources. These are the ones the logs and the survey both miss: AI features inside SaaS you already pay for, browser extensions and the permissions they hold, meeting bots, and AI line items in expense claims. Then the agents and connectors staff authorized against company accounts: check OAuth consent grants in your identity provider, and coding-agent tokens and connected apps in your source control. An agent working from its vendor's cloud may never cross your proxy, so this may be your only view of it.
  5. Merge into the Shadow AI Register and triage each row. One row per tool and use. The same chatbot used by two teams for two different tasks is two rows, because the data involved is different. Each row gets one of four decisions: approve, approve with conditions, replace with the sanctioned tool, or block. Block where the data class is Prohibited and no condition fixes it, or where the tool's permissions or terms cannot be made acceptable at any data class (for example an extension that can read every page you open, or a service that trains on your inputs with no opt-out). Revoke any grant you block.
  6. Publish the Data-to-Tool Rule and supply the alternative. Use the four classes and one question. Tines calls it the Public Test: "Would I post this publicly on the internet?" This program adapts it: if yes, any tool will do; if no, check the table. On the same day, make one approved tool available to everyone you surveyed, start your approved AI tools list with it, and open a request channel with a published target measured in days, not weeks. Then show people how to use it safely, which is what the AI literacy program that meets EU AI Act Article 4 builds.
  7. Record the baseline and book the re-run. Write down today's metrics (they are listed in the next section) and put a repeat of the whole exercise in the calendar for 90 days from now.

A rhythm that works: kickoff Monday morning, survey open to Friday while IT pulls logs and grants, triage the next Monday.

A worked example: the 420-person freight forwarder

This example is invented for illustration. It is not a real company, and the numbers show what the output looks like, not what you should expect.

Picture a freight forwarder with 420 staff. The head of IT security owns the exercise; the managing director signs the amnesty on Monday. By Friday the anonymous survey has 212 responses. Of those, 131 say they use an AI tool for work, and 104 of the 131 do so on a personal account. Meanwhile the proxy logs show 23 distinct AI domains visited in 30 days, and the consent grant export shows 11 AI apps holding delegated access to company accounts.

See what the sources do for each other. The survey alone suggested three or four tools. The logs alone gave 23 domains and no idea what anyone did on them. Neither showed the 11 grants. Together they produce a register. After merging, the team has seven rows that matter:

Tool or featureWhoFound byData class or reachDecision
Public chatbot, free tier, personal accountsCustomer service, about 40 people, drafting shipment status repliesSurvey and logsSensitive (customer names, consignments)Replace with the approved tool
Coding assistant with a personal token to the code repository4 developers, fixing integration scriptsSurvey and source controlSensitive (source code, credentials in config files)Approve with conditions: company plan, secrets scan, token limited to one repository
Meeting bot set to join every calendar invite9 managers, minutesSurvey and consent grantsHears every call, including HR cases (Prohibited)Block; replace with a sanctioned recorder that joins only when invited
Personal AI agent with delegated access to a sales manager's mailbox and calendar1 sales manager, triaging mailConsent grants onlyReads and sends all mail, including customer contractsBlock and revoke the grant; offer the approved tool
Writing extension allowed to read and change all data on all websitesAbout 30 staff, rewording emailsExtension inventoryEvery page the browser opensBlock at any data class
AI writing feature inside the existing office suiteEveryone, rewording emailsSaaS admin consoleInternalApprove
Translation site, no accountCustoms desk, carrier noticesLogsPublic and InternalApprove for Public and Internal only

Row 1 is the big one: not misuse but a business process running on personal accounts. The task is legitimate and the data Sensitive, so move the task, do not stop it: replace.

Row 2 shows what "conditions" means. The developers need the tool; the danger is the personal plan, the broad token and the credentials in config files. So the conditions are a company plan (a contract the company has read), a secrets scan (an automatic check that strips passwords and keys before code leaves) and a token that reaches one repository instead of all of them. Meet them and the tool is fine.

Rows 3, 4 and 5 are blocked for what they can reach, not for anything anyone pasted. The meeting bot joins HR meetings because it joins everything. The agent's grant reaches every customer contract in that mailbox, under vendor terms nobody reviewed. And no data rule protects a page the extension can already read. The agent showed up only in the consent grants: its owner thought of it as a mail setting, not an AI tool.

Row 6 is the quiet source: already on a company account under an existing contract, handling Internal data. Approve, and write it down. Row 7 has no account, so no contract: fine for Public and Internal text such as carrier notices, and nothing above that.

The two-hour session ends with 7 decisions, 1 approved tool announced to all staff, 2 consent grants revoked, and a request channel with a target of 5 working days.

The Shadow AI Register template

This is your first artifact. A spreadsheet is fine; resist adding fields until the second run.

FieldWhat to write
Tool or featureA plain name, including features inside other software, extensions and agents
VendorWho receives the data
Discovered viaSurvey, log, consent grant, extension inventory, expense or SaaS admin
Teams and approximate usersFor example "Customer service, about 40"
TaskThe job people are getting done
Highest data class involvedPublic, Internal, Sensitive or Prohibited
Permissions requestedWhat it can reach without being handed anything: mailbox, calendar, files, repositories, every web page
Account typePersonal, company or none
Vendor terms checkedTrains on inputs? Retention stated? Data location? Data processing agreement? Indemnity? Admin controls on this tier? (Lesson 5 turns these into written checks)
Decision and conditionsApprove, approve with conditions, replace or block, plus what must stay true for an approval to hold
OwnerOne named person
Decision date and review dateWhen it was decided and when it will be looked at again

Two columns do the most work. Record the highest data class, because a tool that sees Sensitive data once a week is a Sensitive tool. "Permissions requested" catches the tools that never needed anyone to paste anything. If either is blank, the row is not finished.

The one-page Data-to-Tool Rule

This is your second artifact, and the seed of your shadow AI policy: four classes, two examples of each from your own company, and which tools may receive each class. Here is the freight forwarder's version.

ClassTwo examples from this companyWhich tools may receive it
PublicPublished tariff sheets, carrier noticesAny tool
InternalProcess notes, routine internal emailsTools marked approved in the register
SensitiveCustomer names and consignment details, source codeOnly the approved tool on a company account with checked terms
ProhibitedHR case files, passwords and credentialsNo AI tool

Above the table goes the Public Test, in large type: "Would I post this publicly on the internet?" Below it goes one line for the second route: "If an app or extension asks to read your mail, calendar, files or every website, ask the service desk first, whatever the data." Examples matter more than definitions. A customs clerk does not think "is this Sensitive?" She thinks "is this like a consignment record?" Give her the consignment record.

One page is a hard limit, or nobody will use it at the moment of pasting. Your first AI acceptable use policy, lesson 5, turns this page into section 1 of the policy and adds the written vendor checks a tool must pass before it joins the approved list.

Failure Modes, Warning Signs and the Metrics to Track

Each way this method fails has a signal you can see in your own data, usually by comparing one source with another.

How do you know the exercise is going wrong?

Failure modeHow to spot itWhat to do
Discovery read as a witch huntSurvey response rate under about a third. Answers of "none" from teams whose domains fill the logs.Reissue the amnesty from a more senior name and show what the first results were used for.
Ban with no alternativeAI domains fall on the corporate network while survey use stays flat. It moved to phones.Supply the approved tool before tightening any block.
Discovery that only watches trafficThe register has no row found by consent grant or extension inventory.Export consent grants and extension permissions and go back to step 4.
Register built once and leftReview dates in the past. New tools appear in expenses but not in the register.Keep the 90-day re-run in the calendar and give each row an owner.
Slow request channelRequests older than the target. Staff say "I stopped asking".Publish the queue and the median, and give one person authority to decide.
Rule nobody can applyStaff cannot say which class a customer email belongs to.Replace definitions with examples from their own work.

The second row deserves a pause, because it looks like success: AI traffic drops after a block and someone reports a win. Check it against the survey. If people say they use the tools just as much, the kettles have moved into the drawers. You have not reduced risk. You have reduced visibility. The third row is the same trap in 2026 form: an agent with delegated access may never appear in your traffic at all.

Which metrics should you track?

Track five, each from a source you already have, and record them at the end of the workshop as your baseline.

MetricWhere it comes fromHow often
Share of staff using unapproved toolsThe anonymous surveyEvery 90 days
Number of distinct AI domains in web logsProxy or DNS logsMonthly
Number of unapproved AI apps holding consent grants to company accountsIdentity provider exportMonthly
Share of staff with access to an approved toolLicense or single sign-on count divided by headcountMonthly
Median days to approve a tool requestTicket timestamps from the service deskMonthly

Read them together, never one alone. The healthy pattern is the first and third falling while the fourth rises: people stop using unapproved tools because they have an approved one. The second falling alone can mean the opposite of what it seems.

The freight forwarder's baseline reads: 104 of 212 respondents using a personal account for work AI, 23 distinct AI domains, 11 AI apps with consent grants (9 after the 2 revocations), and no median approval time, because there was no request channel to measure. "Not yet measurable" is an honest baseline.

What does this set up for the next lesson?

When you run the triage session, one column will cause arguments: the data class. People will disagree about whether shipment data is Internal or Sensitive, and no one will be able to settle it, because the data itself has never been listed or classified. That gap is what lesson 2, building an AI data inventory before AI touches your data, closes. It takes the four classes you introduced here and applies them to the actual systems behind one use case.

Frequently Asked Questions

What is shadow AI?

Shadow AI is any AI tool, AI feature or AI agent used for work that the organization has not approved and cannot see. It usually comes from staff trying to work faster, not bad intent. It includes public chatbots on personal accounts, browser extensions, meeting bots, AI features inside software you already own, and assistants staff have connected to company accounts.

How common is shadow AI in 2026?

Common enough to assume it is in your organization: coverage of IBM's Cost of a Data Breach Report 2026 (July 2026) says unapproved AI featured in 43% of the incidents studied, up from 20% a year earlier. Microsoft and LinkedIn's Work Trend Index (2024 edition) found 78% of AI users brought their own tools to work. Both come from vendors, so measure your own figure with the survey and the logs.

What should we do about employees using ChatGPT at work?

Supply a safe alternative before you block anything: a ban on its own tends to move use onto personal devices where you can no longer see it, and BCG (June 2025) found more than half of employees would use alternatives anyway. Offer one approved tool, a one-page data rule and a fast request channel instead. Block a specific tool where the data is Prohibited and no condition fixes it, or where its permissions or terms are unacceptable at any data class.

Do AI browser extensions and agents count as shadow AI?

Yes, and they are the hardest kind to see, because they reach company data through permissions rather than through anything an employee pastes. An agent with delegated access to a mailbox can work from its vendor's cloud and never cross your proxy. Find them in your identity provider's consent grants, your extension inventory and your source control tokens.

Can we look at employees' AI use without it becoming surveillance?

Yes, if you start with aggregated counts by domain and app, which tell you most of what you need without naming anyone. Check with privacy counsel and any employee representatives before you look at individuals, and put a written amnesty in place first. Then show staff the result was a better tool, not a disciplinary list.

Key Takeaways

  • Assume shadow AI exists in your organization and measure it. Asking managers will not work, because many users are reluctant to admit it.
  • It is a supply problem before it is a compliance problem. Every unapproved tool is a demand signal telling you what your approved tool must do.
  • The risk is what the tool can reach: the data people hand it (four data classes and the Public Test) and the permissions and terms it holds (the register).
  • Detect it from three sources that correct each other: an anonymous survey under a written amnesty, network logs read domains first, and consent grants and extension permissions.
  • Triage every row into approve, approve with conditions, replace or block. Block when the data is Prohibited and no condition fixes it, or when a tool's permissions or terms are unacceptable at any data class.
  • A ban without an alternative reduces visibility, not risk. Supply one approved tool and a request channel with a target in days.
  • Leave with the Shadow AI Register, the one-page Data-to-Tool Rule and five baseline metrics, and re-run the exercise in 90 days.