Last reviewed:
How do you audit your organisation's AI maturity and shadow AI?
Auditing AI maturity means measuring what is actually used, what data flows through it and what framework exists, then placing the organisation between denial, prohibition, framing and governance. Without an inventory of usage, an organisation cannot claim it controls its AI risks.
Auditing AI maturity means measuring the gap between actual usage, exposed data and the written framework. You detect hidden usage, map exposure, then place the organisation on four levels: denial, prohibition, framing, governance.
01The answer in plain terms
Auditing an organisation's AI maturity means measuring the gap between three realities. What employees actually do with AI. The data that flows through those tools. The written framework that management has put in place. The audit does not look for someone to blame. It produces a dated record on which the executive committee can decide.
The method has three steps. First, detect usage, including usage that nobody has declared. Second, map exposure by crossing the tool, the account and the data. Third, place the organisation on four levels: denial, prohibition, framing, governance. The following sections give the grid for each step.
Shadow AI means employees using AI tools without the organisation's approval. It is the first object of the audit, because it is invisible by design. An audit that only looks at tools bought by the company misses part of the usage.
02Why this is an executive committee issue
AI use at work is massive. In France, 73% of private-sector employees use AI at work (+10 points in a year) and 55% do so regularly or daily. Only 37% systematically use tools approved by their company. These figures come from the Greenworking × Ipsos bva study published on 23 September 2026 (1,100 private-sector employees excluding blue-collar workers, including 474 managers, fieldwork from 12 to 25 May 2026). In the United States, 45.2% of working adults aged 18 to 64 use generative AI for their work in the second quarter of 2026, up from 37.4% in August 2025 (FRED, Federal Reserve Bank of St. Louis).
Part of this usage falls outside any framework. In the Wakefield Research survey for PagerDuty (1,250 office workers outside IT, companies with $500M in revenue or more, United States, United Kingdom, Australia and Japan, published on 11 June 2026), 66% of office workers who use AI for their job have already done so with a tool they believed was not authorised. 88% have already shared professional information with a public AI. 38% have handed in AI-assisted work without disclosing it.
Management knows this perimeter poorly. In the EQS Privacy Barometer 2026 (206 professionals, published on 12 February 2026), 80% of French organisations have no clear view of AI-related risks and 48% have not identified the AI systems used in their organisation.
A lack of visibility has a cost. According to IBM (Cost of a Data Breach 2025, Ponemon Institute, 600 organisations, 30 July 2025), 1 in 5 organisations that suffered a data breach did so because of shadow AI. High shadow AI usage is associated with an average additional cost of $670,000.
These figures come from different surveys, on different populations and in different countries. They do not add up and cannot be compared line by line. For other public benchmarks, see the Shadow AI Barometer 2026 from WYP.
These surveys converge on one point. Actual usage is wider than the usage management knows about. An audit replaces an impression with a record.
03Step 1: detect hidden usage
A single detection channel gives a false picture. Technical signals see connections but not intent. A survey sees intent but not everything. Cross five sources and keep the gaps between them. Those gaps are the information.
Technical signals
Firewall or proxy logs, sign-ins through the identity directory, inventory of devices and browser extensions. They show which AI services are reached, from which functions and how often. Aggregate by service or by function. The aim is not to track individuals.
Financial flows
Expense reports, company cards, reimbursed subscriptions, departmental purchases. A reimbursed individual subscription is professional usage that accounting has already recorded.
A short anonymous survey
Four questions are enough: which tools, for which tasks, with which types of content, with which account. Guarantee in writing that no sanction will follow. In the PagerDuty survey, 38% of employees handed in AI-assisted work without disclosing it. Without that guarantee, answers will understate usage.
Interviews by function
A few short interviews with managers and employees in each function: sales, finance, human resources, legal, support. Ask people to show a recent use rather than describe a rule. Managers see their team's practices up close. Start with them.
AI built into tools already in place
Email, office suite, CRM, video conferencing with automatic minutes. These features are often switched on by an update, with no purchase or approval. List them tool by tool.
Finally, distinguish the type of account used. In the Netskope Threat Labs Report: Europe 2026 (data from 1 March 2025 to 31 March 2026), 43% of AI users go through personal accounts, down from 79% a year earlier. Accounts managed by the company go from 28% to 72%. The share of personal accounts is a simple indicator to follow from one audit to the next.
04Step 2: map exposure
A list of tools is not a map. Exposure comes from crossing three variables: the tool, the account and the data. The same tool can carry little risk with a managed account and public content. It can carry a lot with a personal account and a client file.
The tool and its terms
For each detected tool, record the vendor, the function, the plan subscribed and the conditions for reusing data. These conditions often differ between the consumer plan and the business plan of the same vendor. Read them rather than assume.
The account
Company-managed account or personal account. A managed account lets you remove access, keep traces and sign contracts. A personal account allows none of this.
The data
Classify the content that goes into the tool. In the PagerDuty survey, the items shared with a public AI include emails and correspondence (43%), meeting minutes (40%), client data (34%), financial information (31%), confidential documents or strategies (31%) and code or specifications (23%). Use these categories as the basis for your classification.
The legal framework
As soon as personal data is involved, the question falls under the GDPR. Note it in the map and involve the data protection officer.
Two benchmarks help weigh the findings. According to Netskope, 59% of data policy violations linked to AI and personal applications involve regulated data. In the IBM study, 13% of organisations suffered a breach of their AI models or applications, 97% of which lacked proper access controls. So also check who can access which AI tools and which connected data.
The deliverable of this step is an exposure table. One line per tool and function pair, with the account, the data categories, an exposure level (low, medium, high) and a proposed decision.
05Step 3: place maturity on four levels
Four levels are enough to place an organisation. They are not a ranking of worth. They show the next decision to take.
Denial
No inventory, no rule, no measurement. Management believes AI is little used or concerns other teams. This starting point is not rare: in the EQS Privacy Barometer 2026, 48% of organisations have not identified the AI systems in use. Next decision: measure. Launch detection before writing anything.
Prohibition
A memo bans or strictly limits AI. Usage continues off the radar, often with personal accounts. A ban alone says nothing about actual usage. In the PagerDuty survey, 66% of office workers who use AI for their job have already done so with a tool they believed was not authorised. Next decision: offer an approved alternative and measure whether the memo is applied.
Framing
A written policy exists, at least one tool is approved, prohibited data is listed and teams are informed. Application is not yet measured. According to IBM, 63% of organisations that suffered a breach have no AI governance policy or are still writing one. Reaching this level is therefore already progress. But only 37% of French employees systematically use approved tools (Ipsos bva). Next decision: appoint an owner and measure.
Governance
An owner is appointed. The inventory is kept up to date. Access is controlled. Employees are trained. Each new use case goes through a decision process. An indicator is followed at the executive committee and the audit is repeated at a fixed interval. Next decision: maintain. This level is lost if nobody rereads it.
The prohibition level is often mistaken for control. It is an impression, not a record. Only detection says whether the memo is applied.
06The grid: six axes, one piece of evidence per axis
Score each axis from zero to three, with supporting evidence. Zero: nothing exists. One: it exists but is not verified. Two: it is in place and verified on a sample. Three: it is tracked by an indicator and reviewed on a fixed date. A verbal statement is not evidence. Ask for a document, a log, a configuration extract or committee minutes.
Inventory
Do we know which AI tools and features are used, and by which functions of the company? Evidence: the exposure table, dated, with its owner.
Data
Are the prohibited or authorised data categories defined and known to the teams? Evidence: a written list, reviewed by legal and security, with concrete examples.
Accounts and access
What share of usage goes through managed accounts? Who can access which tools? Evidence: a statement of managed and personal accounts, and the list of access rights.
Written framework
Do an AI policy, a list of approved tools and an exception process exist? Evidence: the document signed by management, its date and its distribution.
Skills
Do the people who use AI know what they may put into it and how to check what comes out? Evidence: a training course completed, with its content and the people concerned.
Steering
Is an owner appointed, an indicator followed at committee, an incident already handled? Evidence: committee minutes and the incident register.
To move from scores to a level, apply the weakest-link rule. A written policy without an inventory remains framing on paper.
- Denial: the inventory is at zero and no written framework exists.
- Prohibition: a ban exists on its own. The inventory and accounts stay at zero or one.
- Framing: all axes are at one or more and the written framework is at two.
- Governance: all axes are at two or more and steering is at three.
- An in-between case is placed at the lower level.
Each detected tool then calls for one of three decisions: stop it, approve it with a managed account, or tolerate it for a limited time with a deadline. Record the decision, its owner and the review date.
07Frequent mistakes
Auditing the tool instead of the usage
Banning or approving a tool says nothing about what goes into it. The same application can be used to reword a public text or to summarise a client file. Audit usage and data.
Running a punitive audit
If the audit is used to sanction, teams will stop declaring. In the PagerDuty survey, 38% of employees have already handed in AI-assisted work without disclosing it. An audit looks at usage and functions, not at people.
Ignoring where habits come from
In the PagerDuty survey, 89% of users first adopted AI in their personal life and 79% now use it more at work than outside work. The reflexes come from the private sphere. A rule that offers no professional alternative risks being bypassed.
Doing a one-off audit
Usage moves fast, as the FRED series cited above illustrates. An old record no longer describes the situation. Set a renewal frequency, at least yearly.
A useful audit fits in a few pages. An exposure table, a maturity level, three to five dated decisions. Beyond that, it becomes a report that nobody rereads.
08What WYP does, and where to start
WYP is a strategy consultancy for executives, executive committees and boards of directors. It does not publish a detection tool and does not develop software. The work covers usage, data and the committee's decisions.
The usage audit is part of Cadrage IA, WYP's entry offer. It is a fixed-scope package at €2,900 excluding VAT, delivered in 10 working days. It includes the usage audit, a written AI policy, an approved tool and a presentation to the executive committee. For the analysis, WYP cross-checks several AI models (Claude, ChatGPT, Gemini, Perplexity, Mistral, Grok), compares their answers and identifies blind spots and contradictions. A person then reviews and signs. AI sharpens the decision, it does not make it.
To start, take the AI maturity diagnostic, free, in 90 seconds. It places your organisation in broad terms. Then browse the AI glossary for the notions used here, and the Shadow AI Barometer 2026 from WYP for external benchmarks. If the record shows a gap with what you believed, Cadrage IA turns it into a written policy.
FAQ
How often should AI maturity be audited?
At least once a year, and after any major change of tool or regulation. Usage changes fast. In the United States, the share of working adults using generative AI for their work rose from 37.4% in August 2025 to 45.2% in the second quarter of 2026 (FRED, Federal Reserve Bank of St. Louis).
How do you detect shadow AI without monitoring employees?
Cross signals aggregated by function (network logs, account inventory, expense reports) with an anonymous survey. Guarantee that no sanction will follow. In the PagerDuty survey, 38% of employees handed in AI-assisted work without disclosing it. The audit looks at usage, not at people.
What is the difference between an AI maturity audit and a compliance audit?
A maturity audit measures actual practice and the level of framing. A compliance audit checks that specific obligations are met, for example the GDPR. The first provides the inventory the second needs. According to the EQS Privacy Barometer 2026, 48% of organisations have not identified the AI systems used in their organisation.
Who should own the audit in the organisation?
A member of the executive committee, who makes it a decision topic. The security lead and the data protection officer provide the technical and legal evidence. Managers describe their team's usage. Giving the audit to IT alone leaves out the usage that does not go through IT.
Should unapproved AI tools be banned?
A ban alone measures nothing. In the PagerDuty survey, 66% of office workers who use AI for their job have already done so with a tool they believed was not authorised. Offer an approved tool with a managed account, then follow the share of personal accounts. In its Europe 2026 report, Netskope measures 43% of AI users on personal accounts, down from 79% a year earlier.
What does a shadow AI incident cost?
According to IBM (Cost of a Data Breach 2025), high shadow AI usage is associated with an average additional cost of $670,000, and 1 in 5 organisations that suffered a data breach did so because of shadow AI. The real cost for an organisation depends on its data and its controls.
See also
Further reading
Sources
- PagerDuty Shadow AI Survey, Wakefield Research, fieldwork 9 to 20 April 2026 (1,250 office workers outside IT, companies with $500M in revenue or more, United States, United Kingdom, Australia, Japan, margin of error ±2.8 points), published 11 June 2026. https://www.pagerduty.com/newsroom/shadow-ai-workplace-survey-2026/
- Generative AI, Adoption Rate for Work: Employed Adults (series RPSGENAIUSAGESHAREWORK, Real-Time Population Survey, working adults aged 18 to 64), FRED, Federal Reserve Bank of St. Louis, Q2 2026 value published 4 August 2026. https://fred.stlouisfed.org/series/RPSGENAIUSAGESHAREWORK
- Greenworking × Ipsos bva, « IA au travail : derrière l'adoption massive, des usages fantômes et des managers en première ligne » (in French), 23 September 2026 (fieldwork 12 to 25 May 2026, 1,100 private-sector employees excluding blue-collar workers, including 474 managers). https://www.ipsos.com/fr-fr/ia-au-travail-derriere-ladoption-massive-des-usages-fantomes-et-des-managers-en-premiere-ligne
- Netskope Threat Labs Report: Europe 2026, 25 May 2026 (data from 1 March 2025 to 31 March 2026). https://www.netskope.com/resources/threat-labs-reports/threat-labs-report-europe-2026
- IBM, Cost of a Data Breach 2025 (Ponemon Institute, 600 organisations), 30 July 2025. https://newsroom.ibm.com/2025-07-30-ibm-report-13-of-organizations-reported-breaches-of-ai-models-or-applications,-97-of-which-reported-lacking-proper-ai-access-controls
- EQS Group, Privacy Barometer 2026 (7th edition), 206 professionals, in collaboration with Grant Thornton and the AFJE, published 12 February 2026. https://www.eqs.com/fr/ressources-compliance/livres-blancs/barometre-privacy-2026/