People use their own AI tools when the organisation gives them none. Here is what the 2026 data says and five steps to manage shadow AI without a ban.
TL;DR
Where an organisation has not given people an approved AI tool and clear rules, people tend to find their own. That is shadow AI, and it is rarely about bad intent.
This guide covers what shadow AI includes, what the 2026 data says, why people do it, and five steps to manage it without a ban.
IBM defines shadow AI as the unsanctioned use of any AI tool or application by employees without the formal approval or oversight of IT. Its examples include chatbots, machine learning models used for data analysis, marketing automation tools, and data visualisation tools. The most common version is the simplest: a personal account on a public chatbot, used to draft, summarise, or analyse something for work.
Shadow IT, the older problem of unapproved software, is the closest comparison. IBM notes that shadow AI raises its own concerns around data management, model outputs, and decision-making. A technology partner at the law firm Travers Smith, quoted by Bloomberg, made a sharper point: an unauthorised AI system is nearly always the consumer version, which can take in company data and use it for training. For a leader, that is the difference that matters. The data is processed by a tool whose terms nobody at your organisation has read.
Two recent sources give a sense of scale. They measure different things, so I would not compare the numbers directly.
They agree on direction. Unapproved use is widespread, and IT teams do not see all of it. The UK National Cyber Security Centre said as much in a warning quoted in the Bloomberg report: security teams should not assume they are seeing the full picture.
The same Deloitte survey points at the causes. Half of AI users said they had received no formal guidance or training, and most respondents said their company leaders had a poor or non-existent understanding of AI. The same share who had used AI at work, 63%, worried that their manager might think AI can do their job. The researcher who ran the survey described a stigma around AI use. People who fear that AI will be used to question their worth have good reason to keep quiet about using it.
IBM adds two more. Its explainer notes that people turn to unsanctioned tools when approved options seem insufficient or too slow. Its September 2026 piece on security analysts pasting scrambled commands into personal chatbots makes the point that capable people under time pressure used the best tool available because nobody gave them an approved one, which makes it a leadership problem before it is a user problem.
That gives four causes to check in your own organisation: no approved tool, an approved tool that feels weak or slow, no clear rule or training, and a culture where admitting AI use feels risky.
IBM's 2026 Cost of a Data Breach Report puts the global average cost of a breach at USD 4.99 million. Kiteworks' summary of the report says shadow AI was involved in 43% of breached organisations, up from 20% the year before. It reports that the average cost of those incidents rose from USD 4.63 million to USD 5.39 million, and that 21% led to a regulatory fine.
Read those figures with care. The sample is 602 organisations that had suffered a breach, so it shows what shadow AI costs when a breach happens, not how likely a breach is in your business. I have also relied on Kiteworks' summary rather than the full IBM report.
The regulatory side deserves attention wherever you operate. South Africa's POPIA came into force on 1 July 2021, with penalties of up to R10 million and up to 10 years in prison depending on severity, according to ITWeb. In September 2026, a POPIA executive at the Information Regulator told ITWeb that AI models often retain personal information beyond its intended purpose, which sits badly with the Act's limits on storage. Under GDPR, IBM notes that major infringements can reach EUR 20 million or 4% of worldwide revenue, whichever is higher. Take advice from your own legal counsel on how this applies to you.
The first instinct is to block the chatbot sites and publish a policy. IBM compares this to the failed bans on shadow IT a decade ago: the tools are too useful, the workarounds are easy, and a blocklist goes stale within weeks. Its sharper observation is that a ban pushes the habit somewhere worse. Someone who cannot use a monitored work connection will use an unmonitored personal one, and the organisation loses visibility. The NCSC's David Chismon makes a similar point in the Bloomberg report: organisations cannot block every AI tool, so they need open dialogue and clear guardrails.
I would not ban AI use across the board. I would ban specific data from specific places, which is a narrower and more enforceable rule.
Some data should not go into any tool you have not approved: personal information you hold under a legal duty, and client material covered by confidentiality terms the tool cannot honour. For those categories, a ban on the data is justified. Pair it with an approved route for everything else, so people have somewhere to go.
Shadow AI is usually the first visible sign of a gap in governance, adoption, or both. If you want a scored view of where your organisation stands, the AI Readiness Diagnostic covers strategy, data, people, process, and tooling, and is free to take.
Shadow AI is the use of AI tools for work without the formal approval or oversight of the organisation. A common example is using a personal chatbot account to draft or analyse work material.
Shadow IT covers any unapproved software or hardware. Shadow AI is the AI-specific version, and it raises extra concerns about where data goes, how long it is kept, and how reliable the outputs are.
Using an unapproved tool is not illegal in itself. Putting personal information into a tool without a lawful basis or adequate safeguards can breach data protection law such as POPIA or GDPR. Check your position with your legal counsel.
Start by asking staff in a way that makes honesty safe, then review what your network and software logs show, and monitor for sensitive data leaving the organisation. Assume IT sees only part of the picture.
A blanket ban tends to hide use rather than stop it. A narrower rule works better: provide an approved tool, and prohibit specific categories of data from going into any unapproved one.
Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.
Take the AI Readiness Diagnostic