The Best AI Governance Is Invisible

Between November 2024 and January 2025, researchers at the University of Melbourne and KPMG surveyed 48,340 people in 47 countries about how they use AI. Among employees who use it at work, 56 percent said they had used AI tools without knowing whether it was allowed. Almost half said they had uploaded company information, such as financial, sales or customer data, into public AI tools.

Most governance teams treat numbers like these as a communication problem. The answer is another policy, another training module, another all-staff email. I think they describe a design problem. The best AI governance is invisible to the people building with AI and fully visible to the people accountable for it. When a team routes around a control, the control usually has a design fault, and the evidence on bans suggests that tighter enforcement does not mean less risky use.

Bans did not stop the uploads

The same survey has a finding that should give pause to anyone drafting an AI ban. Uploading sensitive company information or copyrighted material to public AI tools was most common among employees whose organisation had banned generative AI (67 percent) or had a policy guiding its use (56 percent). In organisations with no such policy the figure was 33 percent. The researchers conclude that "outright bans may be ineffective, and that simply having policies does not guarantee compliance."

This is a correlation in a self-reported survey, and organisations that ban AI may differ from the rest in other ways. It does rule out the comfortable reading: in this sample, a ban did not go with less uploading. Nor is the 33 percent a case for having no policy. Those uploads happen out of sight too, with no rule to measure them against.

Other surveys find the same habit. In Microsoft and LinkedIn's 2024 Work Trend Index, a survey of 31,000 knowledge workers in 31 countries, 78 percent of AI users said they were bringing their own AI tools to work. IBM's 2025 Cost of a Data Breach study, run by the Ponemon Institute across 600 breached organisations, suggests what it costs. One in five reported a breach due to shadow AI, and those with high levels of shadow AI saw breach costs about $670,000 higher than those with little or none. Sixty-three percent either had no AI governance policy or were still writing one. Of those with a policy, only 34 percent regularly audited for unsanctioned AI.

The default decides

Behavioural science has a plain lesson for what happens next: people take the default. In 2003 Eric Johnson and Daniel Goldstein compared organ donor consent across European countries in Science. Where citizens had to opt in, effective consent ran from 4.25 percent in Denmark to 27.5 percent in the Netherlands, with Germany at 12 percent. Where citizens were donors unless they opted out, it ran from 85.9 percent in Sweden to 99.98 percent in Austria. The two groups did not overlap.

The detail I find most useful for governance is the Dutch one. When the Netherlands created a national donor registry in 1998, it sent more than 12 million letters, in a country of 15.8 million people, asking citizens to register. The effective consent rate did not change. An awareness campaign could not beat a default.

Most AI governance is built as an opt-in system. The compliant path is the one with extra steps: a separate portal, a risk form filled in from memory, a review queue with no published turnaround. The fastest path is a public chatbot in another browser tab. Training sessions and reminder emails are the 12 million letters.

What a paved road looks like

Netflix's security team met this problem in 2017, as Netflix Studios began building applications for every part of the business, "from pitch to play". In the team's account, published in 2021, there were "too many security things that each software team needed to think about", from TLS certificates to authentication, headers, logging and rate limiting. "There were security checklists for developers, but they were lengthy and mostly manual, neither of which contributed to the goal of accelerating development."

So they built what Netflix engineers call a paved road: a gateway product, Wall-E, that handled authentication for every application behind it. The web application firewall, DDoS protection, security header checks and logging were added to it as required filters, and the checklist shrank item by item. The review question changed from "Tell me how your app does this important security thing?" to "Are you using this paved road product that handles that?" A typical application could go from "git init" to production, authenticated and on the internet, in a little under ten minutes. In the authors' words, developers "didn't necessarily care that the original motivating factor was about security: what they saw in practice was that apps using Wall-E could get in front of users sooner, and iterate faster."

The part governance teams should study is the early stage. Setting up Wall-E first took too much time and effort, and early teams said they had product goals to meet and would talk again next quarter. "Security improvements alone were not enough to drive organic adoption." Adoption came once the team automated the onboarding. Spotify reached the same place with its Golden Path, "the opinionated and supported path" to build something. Engineers can leave it, but then they "will not have the same support."

For AI, a paved road is an approved model gateway that is as quick and as capable as the public tool, with logging, data controls and the system inventory built in. It is a risk tier filled in from the system's registry entry, a review that starts on its own when a system crosses a threshold, and test results that arrive where the reviewer already works. The builder sees a faster path. The risk function sees every call, under logging rules the people using it have been told about, kept to what oversight needs. How much speed the governance keeps is a choice, as I argue in The Hidden Tax on AI Speed.

Rams' ten principles as a governance test

Dieter Rams set out ten principles for good design in the late 1970s, and Vitsoe publishes them in full. They make a useful test for a governance programme because they judge a thing by how it is used. The table pairs each principle with the question I would put to every governance step.

Rams meant the ninth principle literally, about the environment. For governance I read it as the attention a process consumes. Three of the ten do most of the work. Unobtrusive is the paved road. Honest is the line that says where governance stops, which most programmes leave out because it looks like weakness, and its absence is how governance theatre starts. As little design as possible, Rams' "less, but better", is the hardest, because in most programmes an incident adds a control and nothing ever removes one. The UK government's design principles put the cost plainly: "Making something look simple is easy. Making something simple to use is much harder."

Invisible to builders, visible to the board

The strongest objection is that governance nobody notices is governance nobody can check. Automated checks drift. A gateway can log everything and alert no one. Boards and regulators want evidence that a person looked.

The risk is real. The answer is to put the visibility where it belongs. Invisibility is for the person building, while the person accountable needs more visibility than a manual process gives. A paved road can supply it: every request through the gateway leaves a record, which a form filled in from memory never did. Netflix's yes-or-no question is also an audit question: which systems are on the road, and why is each of the others off it? The share of AI work that runs off the road is the measure I would put in front of a board, ahead of the number of policies signed.

Some decisions should stay visible to the builder as well. A system that will decide about people's jobs, credit or health deserves a review that someone has to sit through. Paving the routine work is what frees reviewers' time for those cases.

Three moves for this quarter

1. Find where people go off the road. Ask, anonymously, which AI tools people use for work and compare the answers with the approved list. The gap is your map.
2. Pave the busiest route first. Take the most common off-road use and give people an approved tool that is at least as fast, with logging and data controls built in.
3. Delete one control. Pick a governance step that has not changed a decision in the past year and does not guard a rare, severe event, and remove it. If you cannot tell which steps have changed decisions, start recording that.

An organisation starting from nothing can begin with minimum viable governance: an inventory, a risk register and a named owner. Build each of the three as part of the road from the first day.

More than half of the people in that survey could not say whether what they were doing was allowed. What reaches them is a road they would choose anyway, built so that the people accountable can see everyone on it.

Sources

• Nicole Gillespie, Steve Lockey and others, "Trust, attitudes and use of artificial intelligence: A global study 2025", University of Melbourne and KPMG, April 2025 (48,340 respondents, 47 countries, November 2024 to January 2025), pp. 75 to 76.
• Microsoft and LinkedIn, "AI at Work Is Here. Now Comes the Hard Part", 2024 Work Trend Index, 8 May 2024 (31,000 knowledge workers, 31 countries).
• IBM, Cost of a Data Breach Report 2025, press release, 30 July 2025 (Ponemon Institute, 600 organisations, breaches from March 2024 to February 2025).
• Eric J. Johnson and Daniel Goldstein, "Do Defaults Save Lives?", Science 302, 1338 to 1339, 21 November 2003.
• Jose Fernandez, Arthur Gonigberg, Julia Knecht and Patrick Thomas, "The Show Must Go On: Securing Netflix Studios At Scale", Netflix Technology Blog, 13 September 2021.
• Gary Niemen, "How We Use Golden Paths to Solve Fragmentation in Our Software Ecosystem", Spotify Engineering, 17 August 2020.
• Dieter Rams, "Ten principles for good design", Vitsoe.
• Government Digital Service, "Government design principles", GOV.UK, published 3 April 2012, updated 2 April 2025.

All sources checked on 6 October 2026.

Changes on 6 October 2026. Cut from about 4,800 words to an argument. Removed: the claim that teams spend 56 percent of their time on manual governance, which traces to a vendor blog citing unnamed "benchmarking data" (the closest published figure, in ModelOp's 2025 benchmark of 100 AI and data leaders, is that 56 percent said a generative AI project takes six to 18 months from intake to production); statistics from vendor and consultancy blogs; a benchmark chart with no source; the Anthropic, Estonia and Singapore cases, cut for length; and a download box pointing at a missing file.