What is shadow AI?
It covers any use of an AI tool that escapes the organization: a personal ChatGPT account used to summarize a contract, a browser extension that rewrites emails, an AI feature switched on in software that nobody assessed. The problem is not the tool but the lack of visibility: you cannot protect or encrypt what you cannot see.
Three notions come up in any serious discussion of the topic:
- Shadow AI
- Use of AI tools that the organization has not approved or governed, often from personal accounts.
- Token
- The unit AI providers bill for: a piece of a word. Tokens sent (input) and produced (output) are billed separately, and output costs more.
- PIA
- Privacy impact assessment, required by Quebec’s Law 25 in some situations, including before communicating personal information outside Quebec.
How many employees use AI without oversight?
Many, and often without the employer knowing. In Canada, fewer than 2 in 10 workers say their employer has a comprehensive generative AI policy, and 37% are not aware of any controls [1].
Worldwide, 75% of knowledge workers use AI, and 78% of those users bring their own tools to work, 80% at small and medium businesses [2]. 82% of data pasted into AI tools goes through personal accounts, outside any organizational control [3].
What are the risks for data and personal information?
Whatever is pasted into an assistant leaves the computer for the provider’s servers. The share of sensitive data in what is sent to chatbots rose from 10.7% to 27.4% in one year [5]. In 2023, Samsung banned generative AI for its staff after internal source code was uploaded to ChatGPT [13].
The cost of an incident follows: one organization in five attributes a data breach to shadow AI, and high levels of shadow AI add USD 670,000 on average to the cost of a breach [4]. Gartner predicts that by 2030 more than 40% of organizations will suffer a security or compliance incident tied to unauthorized AI tools [14].
What does generative AI really cost, token by token?
For the same task, the price varies enormously with the model. Within a single provider, a million output tokens costs from USD 0.50 to USD 50 at Anthropic [6], and from USD 0.40 to USD 180 at OpenAI [7]: gaps of 100 and 450 times. Public prices change often; these were collected on October 30, 2026.
Without measurement, spend scatters across per-seat subscriptions, API invoices and forgotten trials. Gartner notes that AI cost estimates can be off by 500 to 1,000% [9].
| Model (provider) | Public output price, per million tokens | Gap |
|---|---|---|
| Haiku 5.5 (Anthropic) | USD 0.50 | baseline |
| Sonnet 5.5 (Anthropic) | USD 10 | 20 times |
| Fable 5.1 (Anthropic) | USD 50 | 100 times |
| gpt-5-nano (OpenAI) | USD 0.40 | baseline |
| gpt-5.5-pro (OpenAI) | USD 180 | 450 times |
How can you cut your AI bill?
The levers are documented by the providers themselves. Batch processing costs 50% less at Anthropic [6]. Prompt caching lowers the price of tokens already sent: a cache read costs 0.1 times the input price at Anthropic [6], and OpenAI advertises up to a 95% discount on cached tokens [8].
The simplest lever remains model choice: keep the most expensive models for the tasks that need them. That requires knowing which team uses which model, which means tying every subscription and API invoice to a team.
What does Quebec’s Law 25 say when an employee uses ChatGPT?
Law 25 does not target AI as such, but personal information. According to law firm analyses, communicating personal information outside Quebec, including to a provider, first requires a privacy impact assessment [11]. An employee who pastes a client’s personal information into an assistant hosted outside Quebec could therefore fall under that obligation.
A decision based exclusively on automated processing also requires informing the person concerned [12]. Administrative penalties can reach CAD 10 million or 2% of worldwide turnover [10]. These benchmarks are not legal advice: have your situation reviewed by your legal counsel.
How do you govern AI use in five steps?
1. Measure actual use: which assistants, which accounts, which teams, which data. A written policy without this measurement governs imaginary use.
2. Offer an approved path: company accounts for legitimate uses. Banning without an alternative pushes use to personal accounts.
3. Name the data that must never leave, with examples: client personal information, contracts, source code.
4. Write a short AI acceptable use policy, share it with everyone and train the teams.
5. Manage spend: tie every subscription and API invoice to a team, set caps, review model choices.