The organization that has decided not to decide has still decided. It has chosen the version of AI with no account, no policy and no record.
Waiting is the most defensible-sounding option in the room and the only one nobody writes a business case for. This page is not an attempt to frighten you into a purchase. It is the sourced picture of what happens in the meantime, who carries it, and the smallest governed step that changes the situation.
Policies against consumer AI are widespread and mostly sincere. The research on whether they change behaviour is less comfortable, and it comes with caveats worth reading: several of these studies are vendor-commissioned, the populations differ, and the percentages must not be averaged into one number. Cited individually, with who was asked, they still describe the same situation.
"While only 40% of companies say they purchased an official LLM subscription, workers from over 90% of the companies we surveyed reported regular use of personal AI tools for work tasks." MIT's Project NANDA, "The GenAI Divide: State of AI in Business 2025", July 2025. The report calls itself preliminary findings and rests on 52 organizational interviews and 153 survey responses, so read it as a shape rather than a measurement.
Among more than 2,600 security and privacy professionals across 12 countries, 46% said they had entered employee names or information into generative AI applications, 42% non-public information about their company, and 31% customer names or information. Cisco, "2025 Data Privacy Benchmark Study", survey fielded in late 2024. These are the respondents most aware of the risk, describing their own behaviour.
41% of employees said they find ways around application blocking, in a survey of 1,020 US and UK employees fielded between 30 July and 11 August 2025. UpGuard, published 10 November 2025. Separately, Gartner reported that 69% of organizations suspect or have evidence that employees are using prohibited public generative AI tools, from a survey of 302 cybersecurity leaders conducted March to May 2025 (press release, 19 November 2025).
"One in five studied organizations (20%) experienced breaches linked to shadow AI", and those incidents added as much as USD 670K to the average breach cost. Among organizations reporting AI-related breaches, 97% said they lacked proper access controls, and 63% of the breached organizations studied lacked AI governance policies. IBM, "Cost of a Data Breach Report 2025", summarised 12 November 2025. A vendor report built on breach research, and the 63% is of breached organizations, not of all organizations.
A joint investigation by the Office of the Privacy Commissioner of Canada and the privacy regulators of Quebec, British Columbia and Alberta into a public AI tool found that the way personal information had been collected to train the models was "overbroad and therefore inappropriate", and that the company "failed to obtain valid consent". The company told investigators that "untraining" a model so it no longer generates specific personal information "is not currently feasible" (PIPEDA Findings #2026-002, 6 May 2026). That is the part a policy cannot undo after the fact.
Inside almost every organization that has stalled, the stall has a shape. One person has said no and owns the consequences of being wrong about that. Another watches their team lose ground to competitors whose people are allowed to work differently. Neither is being unreasonable, and nothing moves until both can say yes to the same controlled capability.
The security owner carries the risk if they say yes. The operating owner carries the cost of saying no. That standoff is the opportunity.Why this page exists at all
It has a name attached to it, it arrives in one bad afternoon, and it is the kind of event that gets written up. Anyone who has ever approved a tool understands exactly why the cautious answer is attractive, and why it is rarely revisited once given. Written for that reader here.
It arrives as slower turnaround, work that stays in one person's head, and a slow drift in what a team of that size can absorb. No single quarter shows it. Nobody signs it. It is carried by the operating owner and it is real. Written for that reader here.
While the two of them disagree, the work continues in browser tabs on personal accounts. The ban did not stop the adoption; it stopped the governance. That is the outcome neither person would have chosen if it were put to them as a choice.
The breach numbers are the visible cost and they are not the important one. The deeper cost is adaptive capacity: while an organization debates tools, its peers are learning how to turn their own processes into capabilities, repeatedly, and that skill compounds. The organization that starts later does not start from where it is now. It starts from where it is then, without the practice.
| Question | Under a ban that is routed around | Under one governed capability |
|---|---|---|
| Where does the data go | ✕Wherever a personal account sends it | ✓An account you own, with a policy enforced on every call |
| What is recorded | ✕Nothing you can read | △Model, caller, latency and token counts. Deliberately no prompt or answer text |
| Who is accountable | ✕Whoever pasted it | ✓A named administrator, with grants and quotas |
| What the organization learns | △Individual habits, unshared | ✓A practice: what works becomes a module everyone gets |
| What happens at a review | ✕There is nothing to scope | ✓An architecture, a posture and a controls matrix with its exceptions |
The middle column is not a caricature. It is what the surveys above describe, and it is the state most organizations are actually in while they consider their options. The right column is what one deployment produces; it is not a claim that everything in the organization becomes governed at once.
Nothing here requires an enterprise programme, and the first move is deliberately small enough to reverse. Each step ends in a decision somebody makes on the record, which is the point: the alternative is a situation nobody ever decided.
One list, written by the people who know: the categories of material that must stay inside an account you control. Most organizations have never written it down, and the writing is more clarifying than any vendor conversation.
A network topology and data-flow map of what is actually happening, including the personal accounts. This is the deliverable of the $2,500 strategic assessment, and you keep it whether or not you proceed with anything.
One deployment, one posture tier, one group of people, with a content policy enforced by cloud identity policy rather than by an application setting, per-person quotas and an administrator who can change grants without a redeploy. Baseline suits a first evaluation; a security review will usually want Standard.
A named bottleneck, with an owner, measured against a baseline you set yourself. Most of what gets built at this stage is governed business software rather than anything that calls a model, which is a feature of the approach rather than an apology for it.
We are not saying your organization has been breached, that these survey percentages apply to you, that AI is mandatory, or that delay is irrational. Several of the studies above are vendor-commissioned, one of the Gartner figures is a forecast about 2030 rather than a measurement, and the populations differ enough that averaging them would be dishonest. The argument is narrower: the realistic alternative to a governed capability is usually not zero use, so the comparison should be drawn against what is actually happening.
Then you are in a better position than the surveys describe and the first step costs you very little: check. If use is genuinely near zero, the question becomes the operating one rather than the security one, which is a much more pleasant conversation to have. The posture questionnaire is a reasonable structure for asking.
Regulation is moving, and nothing on this site tells you what your obligations will be. What we would note is that the current bill before Parliament is a bill, that the existing federal privacy law has applied throughout, and that accountability for personal information travels with the data regardless of where it is processed. A governed deployment is easier to adapt to a new rule than an unrecorded habit is.
$2,500 for the assessment, which is bounded and whose deliverable you keep. Deployment starts at $2,500 and typically lands around $5,000 once the work is scoped. Managed operations are $500 a month, cancel anytime. Compute is billed to you directly by your cloud provider at list price. The whole ladder is on the pricing page with nothing underneath it.
Often because the pilot had no owner, no boundary and no route to production, not because the technology failed. The engagement path puts a gate at each of those points, and the gates are yours to fail us on. If the answer at gate one is that governed AI does not create enough value in your organization, that is a legitimate result and a cheap one.
Start with the map, not the platform. A bounded assessment that shows where your data goes today, and what one governed route would change.