[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-what-gartners-first-ai-governance-mq-gets-right":3},"\nIn June 2026, Gartner published its first Magic Quadrant for AI Governance\nPlatforms. Whatever one thinks of quadrants, the event itself carries\ninformation: a category earns an MQ when buyers are spending real budget\non it, and finance signs off on named line items. AI governance has gone\nfrom conference topic to procurement category.\n\nThe timing is not mysterious. The EU AI Act's obligations are phasing in\nthrough 2026 with penalties reaching into the tens of millions of euros or\npercentage points of global turnover, and surveys keep finding the same\nuncomfortable spread: AI deployed nearly everywhere, governed nearly\nnowhere. In supply chain alone, an [IDC InfoBrief commissioned by\nKinaxis](https://www.techtimes.com/articles/324392/20260813/supply-chain-ai-deployed-88-governed-12-idc-finds-trust-real-barrier.htm)\nfound 88% of organisations had deployed AI while only 12% had governance\nfully embedded. It is a vendor-sponsored study, so read it as directional.\nThe direction, though, matches everything else being published. A budget\nline for closing that gap is rational, and overdue.\n\n## The bolt-on temptation\n\nA procurement category wants a product to buy, and the natural product\nshape is a governance platform: a separate system that inventories your\nmodels, monitors your agents, and reports on your AI estate. A watcher\nbeside the watched.\n\nInventories, monitoring and estate-wide reporting are genuinely useful,\nand for many organisations a governance platform will be the first time\nanyone could even list their models. But a watcher has structural\nlimits worth naming before the budget is committed: it knows what it is\nshown, so shadow AI stays invisible to it by definition, and it is\nupdated on the compliance calendar while the estate changes daily, so\nit drifts. A governance platform is necessary furniture. It is not, by\nitself, governance, in the same way a controls register is not a\ncontrol.\n\n## Governance as a property, not a platform\n\nThe alternative shape puts the governance inside the thing being\ngoverned. Concretely, that looks like:\n\n- **Provenance carried by the model itself.** Version, owner, frozen\n  training slice, backtest with the misses included, approval chain,\n  append-only. Not an inventory entry about the model. A passport the\n  model cannot travel without.\n- **Permissions enforced where the data lives.** An answer, a search\n  result or an aggregate that respects the viewer's entitlements at the\n  data layer, so there is no unguarded path for a governance layer to\n  miss.\n- **Agents that inherit their constraints.** When agents act, they carry\n  the user's security scope, spending budgets and audit logging by\n  architecture, not by a policy document asking nicely.\n- **Determinism where decisions are made.** Same inputs, same number,\n  every time, because a figure that changes on refresh cannot be governed\n  at all.\n\n> A control that lives beside the system can be skipped. A control that\n> is a property of the system cannot.\n\nThe test worth applying to any AI governance approach is simple: if the\ngovernance component were switched off, would the governed behaviour\ncontinue? If yes, the governance was architectural. If no, it was a\nreport.\n\nProphesee was built on the architectural answer. Passports on every\nmodel, permissions at the data layer, inherited constraints for anything\nthat acts. See what governed-by-construction looks like on your own\ndata. [Start here](/contact).\n",1786786820756]