AI Governance · Case Study · 2026

Bringing Shadow AI Into Managed Use

How hidden employee AI use can become visible, understood and responsibly managed without shutting down useful experimentation.

Read Case Study

Client Context

A multi-site hospitality group had two AI stories running in parallel, one sanctioned, one not.

The client is a UK hospitality group operating a portfolio of restaurants and boutique hotels across several regions. Leadership had formally rolled out two AI systems: an AI calling and booking platform handling phone reservations across sites, and an AI-assisted invoice parsing and checking tool supporting the finance team through a high volume of small supplier invoices. Both were deliberate choices, reasonably well specified, and understood by the people who had commissioned them.

What leadership hadn't accounted for was everything else. Individual reservations staff, site managers and finance assistants had, independently and often well before the formal rollout, started using consumer AI tools to draft supplier emails, summarise guest complaints and check figures. This is close to the industry norm rather than an exception: PagerDuty's 2026 Shadow AI Survey, conducted by Wakefield Research among 1,250 office professionals across Australia, Japan, the UK and the US, found that two-thirds (66%) of office professionals have used AI tools at work despite believing they were not permitted under company policy. In the UK specifically, 51% had entered work-related correspondence into public AI tools such as ChatGPT, Claude or Gemini, well above the 43% global average. The same survey found the exposure was worse at smaller organisations: 40% of workers at companies with fewer than 1,500 employees had entered customer data into public AI tools, against 27% at larger organisations, a pattern that maps closely onto multi-site hospitality groups, where individual sites often operate with a high degree of local autonomy.

INTAI's brief wasn't to introduce AI to this business, it already had, in two different ways. It was to find out what was actually running underneath and alongside the two sanctioned systems, and build a way to bring it into view without shutting down experimentation that, in a number of cases, was already solving a real operational problem.

This case study sets out what that discovery found, how INTAI worked with the group to manage it, and what changed in the two sanctioned systems themselves once they were examined properly. Client details have been anonymised and adapted to protect confidentiality; the substance reflects the engagement approach INTAI applies with multi-site hospitality operators adopting AI.

01 · Two Systems, One Blind Spot

Having a governed AI system doesn't mean you've governed AI use.

The AI calling and booking platform had been introduced to solve a specific, well-understood problem: hospitality research from Canary Technologies has found that up to 40% of calls to hotels go unanswered, roughly a third of them from guests ready to book, a direct and quantifiable revenue loss. The group's voice AI system answered every call, checked availability and took bookings around the clock. The invoice parsing tool solved an equally specific problem: a high volume of small supplier invoices arriving in different formats across many sites, previously keyed in by hand.

Both projects had an owner, a business case and a rollout plan. From leadership's perspective, that meant the organisation's "AI question" had been answered. In practice, it meant two specific, high-visibility uses of AI had been governed while dozens of smaller, informal ones had not, and the formal projects themselves had been assessed for cost and functionality rather than for the compliance obligations now attached to systems of exactly this kind.

Neither gap was unusual, and neither reflected carelessness. It reflected a common pattern: governance conversations tend to start and stop at the systems leadership commissioned, while the AI question inside a business is rarely limited to those systems alone.

02 · What the Discovery Found

Shadow AI wasn't hiding out of malice. Most of it was hiding because nobody had asked.

INTAI's discovery process across the group's sites surfaced consumer AI tools being used to draft guest correspondence, summarise complaint emails, compare supplier pricing and draft rota and shift-cover messages, none of it sanctioned, none of it visible to head office, and almost none of it malicious. The PagerDuty/Wakefield survey helps explain why: 72% of office professionals believe they understand how to use AI for their job better than whoever is responsible for managing AI at their company, and 39% said they would rather use AI without telling anyone than risk being told it wasn't allowed. Staff weren't concealing a problem. Most believed, reasonably, that they'd found a solution and had little reason to flag it.

The risk sat in what was being typed into these tools, not in the fact that they were being used at all. In hospitality specifically, that meant guest names, contact details, dietary and allergy information, and occasionally booking or payment context moving into public AI tools with no visibility and no data agreement governing what happened to it next. The same survey found more than a third of respondents (34%) had entered customer data into public AI tools generally, a figure this engagement's discovery phase found was, if anything, an understatement of what was happening at individual sites.

None of this showed up in the group's existing risk register, because the existing risk register had been built around the two systems leadership knew about.

03 · The Sanctioned Systems Had Their Own Compliance Clocks

The two "safe" AI systems turned out to be carrying real, dated compliance obligations of their own.

The AI calling and booking platform sits directly inside the EU AI Act's transparency regime. Article 50 requires that any AI system designed for direct two-way interaction with a real person, including voice-based booking and reservation agents, tell the caller, audibly and from the very first exchange, that they are speaking with an AI rather than a person. These obligations apply from 2 August 2026, with fines of up to €15 million or 3% of global annual turnover for non-compliance. At the point INTAI reviewed it, the group's voice AI system carried no such disclosure.

The same system also took card details over the phone to secure booking deposits and guarantees. PCI DSS compliance applies to any environment where payment card data is spoken, heard, recorded or stored, including AI voice agents and their call recordings, and PCI DSS 4.0.1's Targeted Risk Analysis requirements now expect a documented risk assessment of exactly how that exposure is being managed, rather than an assumption that a "pause and resume" approach is sufficient. The booking system had been assessed for whether it worked. It had not been assessed against this standard.

The invoice parsing and checking tool sat inside a different, equally live risk. The Association for Financial Professionals' 2025 Payments Fraud and Control Survey found that 79% of organisations were victims of attempted or actual payments fraud in 2024, with vendor imposter fraud, fraudulent requests to change a supplier's bank details, up 11 percentage points on the year before. Faster, AI-assisted invoice processing is a genuine efficiency gain, but speed without an added control point can just as easily accelerate fraud losses as it accelerates legitimate payments; industry data suggests automated financial workflows with proper controls can cut internal fraud risk by around 25%, but only when the controls, not just the automation, are deliberately built in.

In other words: the two systems the group had actually governed were, on inspection, only partly governed, and the gap wasn't in the AI itself, but in what hadn't yet been checked around it.

The Wider Pattern

What this hospitality group found wasn't unusual. The data shows most organisations carry the same blind spot.

Widespread unsanctioned AI use, a UK-specific spike in sensitive data entering public tools, and a fraud landscape that rewards speed without controls, together, these are the conditions this engagement was built to address.

Used AI Despite Believing It Wasn't Allowed

Of office professionals have used AI tools at work despite believing they were not permitted under company policy.

PagerDuty / Wakefield Research, 2026 Shadow AI Survey

UK Workers Entering Work Data Into Public AI

Of UK office professionals have entered work-related correspondence into public AI tools, above the 43% global average.

PagerDuty / Wakefield Research, 2026 Shadow AI Survey

Organisations Hit by Payments Fraud in 2024

Of organisations were victims of attempted or actual payments fraud in 2024, with vendor imposter fraud rising sharply.

Association for Financial Professionals, 2025 Payments Fraud and Control Survey

Together, these figures describe a workforce that has already adopted AI faster than most organisations have built the visibility or controls to manage it, in the marketing team, at the front desk, and in the systems finance relies on to pay suppliers.

04 · Bringing Shadow AI Into View Without Shutting It Down

A blanket ban was never the right first move. Visibility was.

INTAI's discovery work mapped what was actually in use across the group's sites, not what an approved-tool list said should be in use, talking directly to reservations staff, site managers and the finance team about the workarounds they'd built for themselves. Each use was then classified by two things: the sensitivity of the data involved, and how much a heavy-handed response would cost the business in lost efficiency if that use case turned out to be genuinely useful.

That distinction mattered, because the survey evidence suggests bans alone tend to backfire: 39% of employees who use AI at work said they would rather conceal that use than disclose it and risk being told to stop, and nearly half (48%) of those caught using AI outside company policy already faced formal consequences without that changing the underlying behaviour. Prohibition without an alternative doesn't remove shadow AI. It just removes visibility of it.

Instead, INTAI worked with the group to separate what needed to stop immediately, anything putting guest personal data or commercially sensitive supplier terms into public AI tools, from what needed to be brought into managed, sanctioned use because it was solving a real problem: several of the drafting and summarisation use cases staff had adopted informally were kept, moved onto an approved tool with proper data handling, and given an actual owner.

05 · What Was Built

A named owner, an approved-tool list built from real usage, and two sanctioned systems brought up to compliance standard.

A group-level owner was named for AI use overall, supported by a small number of site-level points of contact who understood what their teams actually needed, replacing the assumption that the two commissioned AI systems represented the whole picture with an ongoing view of what was really happening.

On the calling and booking platform, INTAI worked with the group to add a clear, audible AI disclosure at the start of every call ahead of the EU AI Act's Article 50 deadline of 2 August 2026, and moved card capture out of the spoken conversation into a secure keypad-entry step, taking cardholder data out of scope of the call recording and reducing the PCI DSS exposure the system had been carrying since launch.

On the invoice parsing and checking tool, a specific control was added rather than a general slowdown: any change to a supplier's bank details now triggers mandatory human verification against a known contact, regardless of how quickly the invoice itself was processed, directly targeting the vendor-imposter fraud pattern the AFP's survey data shows is rising fastest.

For the wider, previously informal AI use, a short approved-tool list was published with plain guidance on what kinds of data could and couldn't go into it, alongside a simple route for staff to request a new tool be reviewed and added, replacing concealment with a process that took less effort than hiding.

06 · What This Means for Other Hospitality Operators

The lessons from this engagement generalise well beyond one group's front desk and finance team.

Shadow AI is an information problem before it's a discipline problem. With 66% of office professionals already using AI tools they believe aren't permitted, treating every instance as a policy breach to be punished misses that most staff are trying to solve a real problem and simply weren't asked how.

A commissioned AI system carries more than one compliance clock. A voice booking agent isn't just a product decision, it's an EU AI Act Article 50 disclosure obligation from 2 August 2026 and a PCI DSS question the moment it touches a card number.

Treat a vendor bank-detail change as a control point, not a data field. With 79% of organisations hit by payments fraud in 2024 and vendor imposter fraud rising fastest, AI-assisted invoice processing needs a deliberate human checkpoint built in, not just faster throughput.

Multi-site autonomy multiplies the risk, not just the visibility gap. Smaller sites and teams were shown to be more likely to put customer data into public AI tools, exactly the operating model most hospitality groups run.

Bring it into view before deciding whether to ban it. Prohibition without an alternative pushes AI use further underground; visibility, paired with a genuinely usable approved route, is what actually changes behaviour.

Conclusion

Governing the AI you commissioned was never the whole job. Governing the AI your people actually use is.

This engagement started with two visible, sanctioned AI systems and found a much larger, invisible layer of AI use sitting alongside them, plus compliance gaps inside the visible systems that hadn't been checked at all. Neither discovery was a failure of the original projects. It reflected how AI actually spreads through an organisation: faster than the governance built around any single system can keep up with.

What closed the gap wasn't a ban. It was mapping what was really happening, keeping what worked inside a managed and visible process, fixing the specific compliance exposures, voice disclosure, card data handling, vendor verification, sitting inside the systems already in production, and giving the whole picture a named owner going forward.

For a hospitality group whose reputation depends on guest trust and financial control in roughly equal measure, that combination, visibility, proportionate control, and a standing owner, is what turns AI from a risk hiding in plain sight into infrastructure the business can actually rely on.

References & Sources

Research and regulatory context informing this case study.

This case study draws on independent workforce survey data, sector research, official fraud data and applicable AI and payments regulation. Statistics are attributed to their originating source wherever directly cited.

PagerDuty (Wakefield Research). 2026 Shadow AI Survey. June 2026.

Referenced for findings on unsanctioned AI use, data entered into public AI tools, and employee attitudes toward disclosure across the UK, US, Australia and Japan.

View source

Association for Financial Professionals. 2025 AFP Payments Fraud and Control Survey Report.

Referenced for findings on the prevalence of payments fraud and the rise of vendor imposter fraud in 2024.

View source

European Union. EU AI Act, Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems.

Referenced for the disclosure requirements applying to AI voice and chat systems from 2 August 2026.

View source

Canary Technologies, via Hospitality Net. AI in Hospitality: Voice, Messaging, and Revenue Growth.

Referenced for research on the proportion of hotel calls that go unanswered and the resulting booking impact.

View source

PCI DSS Guide. What You Should Know About PCI-Compliant Call Recording.

Referenced for the scope of PCI DSS obligations where payment card data is spoken, heard, recorded or stored, including within AI voice systems.

View source

Quadient. Accounts Payable Automation Trends for 2026: AI, Touchless AP, and Fraud Prevention.

Referenced for industry data on AI-assisted invoice processing, touchless AP rates and fraud-prevention controls.

View source

This case study reflects INTAI's engagement approach with multi-site hospitality operators managing both sanctioned and informal AI use. Client details, including business size, sector focus and specific circumstances, have been anonymised and adapted to protect confidentiality, and any figures describing this specific engagement are illustrative rather than reported client metrics. External statistics, survey findings and regulatory publications are attributed to their original source and reflect the reporting periods of those source publications. This case study is provided for general informational purposes and does not constitute regulatory or legal advice.

Scroll