Your policy now excludes the thing you automated
AI exclusions arrive as renewal endorsements, and a document extraction error falls between professional indemnity, cyber and product liability. What the wording does, why a human-review answer on an application behaves like a condition of cover, and the four artifacts that carry the argument when a loss happens.
15 minThe dossier cannot be written afterwards
High-risk obligations became enforceable on August 2, 2026, and documentation moved ahead of go-live rather than after it. What the technical dossier has to establish for a document workflow, why four of its eight sections need measurements that cannot be produced retroactively, and what to demand from a vendor at the proof of concept.
14 minThe agent became a regulated entity
Regulation used to attach to the model — what a system was, how it was trained, what it disclosed. That framing breaks the moment software stops recommending and starts acting: the agent files the invoice, sends the instruction, updates the record. The object of concern moved from the output to the action, and authority is the thing legal systems have always regulated — who granted it, how far it extends, who answers when it is exceeded. So jurisdictions began writing agent-specific rules rather than stretching model-era ones, some with independent security assessment attached. This is the governance regime, not the controls it gets confused with: a named business owner, technical steward and risk sponsor per agent; an inventory carrying purpose, scope and review date; a decision and action trail that stands up as regulatory evidence; and an intervention path you have actually tested. Plus the four failure shapes and the honest cost of the regime.
13 minEncrypted today, readable tomorrow
Post-quantum migration became a board project in 2026, and not because of a quantum computer that exists — because of one that will. NIST's standards are final, and "harvest now, decrypt later" means an adversary captures your encrypted long-retention document today and decrypts it when the machine matures: the retention period is the exposure. The threat is double — the cipher that keeps a contract secret and the signature that makes it non-repudiable both rest on algorithms with a known expiry, so an ECC signature or timestamp quietly loses its future validity and the audit trail built on it stops being evidence. Distinct from C2PA content provenance (which attests origin), this is migrating the encryption and signature algorithms themselves. The moves forming — a cryptographic inventory, crypto-agility, hybrid signatures through the transition, re-sealing the archive, and prioritising by sensitivity times retention — the four ways it goes wrong, and the trade-off stated honestly.
13 minThe agent nobody decommissioned
Standing up an agent became a menu click in 2026; aging one out never got the same treatment — so the fleet grows in one direction and the organisation proud of two hundred agents in production discovers, at audit, that a large share are orphaned and still hold access to systems that matter. The dormant agent keeps its API key, cached token, memory, embeddings and integrations: a live identity with standing privilege nobody is watching. Agent lifecycle governance is distinct from machine-identity security (NHI sprawl) and observability (runtime tracing): every agent needs an owner, a review date and a decommissioning gate. A registry because you cannot govern what you have not inventoried, a sponsor and review cadence, a gate that revokes everything at once, and an agent-to-privilege-to-data map. The four ways it goes wrong, and the trade-off stated honestly.
13 minWhen the agent acts, who answers?
Insurers began excluding AI from general liability by default in 2026 — CGL endorsements that surface quietly at renewal — opening a coverage gap exactly when the agent approves a payment, decides, and writes to a system of record, while standalone AI-liability policies emerged and governance became an underwriting requirement. This is risk transfer, distinct from agent identity (which agent acted) and observability (what it did): who bears the loss, under which clause, backed by which policy. Allocate liability explicitly in the MSA, treat governance evidence as what unlocks insurability and a higher limit, make the signed audit trail a claim prerequisite, and design authority ceilings and human review as premium reducers. The four ways it goes wrong, and the trade-off stated honestly.
13 minDeleting the file isn't forgetting
Erasing the source document does not remove the influence that data left behind: once information has passed through a system it lives in fine-tuned weights, an indexed embedding, a cache and a log, and the right of erasure has stopped accepting "we deleted the record" as proof. Machine unlearning is the discipline of removing a specific data point's influence without retraining from scratch — distinct from privacy-preserving AI (which limits what the model absorbs) and from a governed vector store (which deletes the stored copy). Inventory every layer the data reached, an erasure pipeline that fans out to all of them, a verifiable unlearning certificate instead of a declaration, and per-tenant isolation that keeps forgetting tractable. The four ways the forget button goes wrong, and the trade-off stated honestly.
13 minTransparency is a feature, not a footnote
The AI-transparency obligation stopped being a clause in the terms of service in 2026 and became a shippable feature: the newest rules require the product to disclose an AI interaction on the surface and mark synthetic output machine-readably, at the point of contact. Distinct from cryptographic provenance (a signing primitive) and from deepfake detection (the inbound problem), this is the outbound obligation to label your own output — embed the mark in the generation path, make it survive export, distinguish AI-assisted from AI-generated, and record that the disclosure happened. The architecture that makes the mark hold, the four ways it goes wrong, and why the platform that builds it in ships a compliance-ready default the disclaimer never can.
13 minProve your controls every single day
The governance shift that hardened in 2026 is quiet but total: the regulator and the enterprise buyer stopped accepting the snapshot — the annual audit certifies one instant and says nothing about the other 364 days. Continuous controls monitoring (CCM) is the architectural answer: instrument each control, pull the evidence automatically from infrastructure, identity, HR systems and the pipeline, and keep a versioned record so the posture is current and provable every day. Why the point-in-time audit broke, how CCM differs from a GRC checklist and a live dashboard, the architecture that makes continuous evidence hold, the four ways it goes wrong, and the trade-off stated honestly.
13 minKeep the sensitive document on the device
For the most sensitive documents, region-bound cloud stopped being enough in 2026 — the answer became on-device and edge inference, where the data never leaves the perimeter. Compact quantized small and vision models running on AI PCs and edge servers now handle the bulk of classify-and-extract traffic locally, for three reasons: residency at its limit (zero egress), latency without a cloud round-trip, and lower unit cost on predictable load. How the tier differs from sovereign-cloud and air-gapped, the hybrid pipeline that keeps a frontier model for the exception that may travel and a human for the one that may not, the model fleet and local quality measurement it demands, the four ways it goes wrong, and the trade-off stated honestly.
13 minNow the RFP asks if you're certified
ISO/IEC 42001 — the first international standard for an AI management system, a Plan-Do-Check-Act discipline for organizations that build, deploy or use AI — went from a governance nicety to an RFP line in 2026, appearing in roughly 40% of enterprise AI vendor questionnaires in Europe and near 25% in North America. The certificate answers a question self-attestation cannot: how do you govern the AI lifecycle, and who checked? What a 42001 certificate actually claims, why scope and operating evidence separate a real certification from a wall decoration, the framework that clears the procurement gate, the four ways the certificate goes wrong, and how it maps onto SOC 2, ISO 27001, NIST AI RMF and the EU AI Act as a reusable compliance shortcut.
13 minAI Bill of Materials: the new procurement bar
The AI-BOM went from optional security artifact to RFP line in 2026. What it inventories across six areas, why provenance on data and third-party models is the part that survives an audit, how CycloneDX and SPDX map to the EU AI Act, and how to generate one that stays current.
13 minThe US AI-law mosaic: fifty conflicting state clocks
The EU gave vendors one clock; the US became a mosaic you comply with jurisdiction by jurisdiction. Why exposure follows the person the decision is about, what Colorado and Texas turn on, and the architecture that survives more than one regulator.
13 minThe EU AI Act high-risk deadline: what August 2, 2026 requires
On August 2, 2026 the Act's high-risk obligations become enforceable. What Annex III triggers, why "we're not European" is no defence, the framework that survives a conformity assessment, and why the deadline is a product decision, not a legal task.
13 minRegulator-in-the-loop: pre-launch model access as the 2026 baseline
Pre-release access to frontier models became a baseline practice in 2026. The five RFP questions enterprise procurement now asks, the model-card fields buyers check, and the rollback architecture that holds when a version gets pulled.
13 minSovereign AI: residency as a contract clause, not a policy slide
Sovereign AI consolidated as an enterprise category in 2026, with a 30–60% premium and six clauses regulated buyers now want by name — reserved compute, jurisdiction-bound models, BYOK/BYOM, signed audit trail, kill switch and bordered sub-processors.
12 minAI governance for document AI: ISO 42001, EU AI Act and audit evidence in 2026
Gartner expects 70%+ of enterprises to adopt a formal AI governance standard by end of 2026. The five frameworks, three audit shapes and the evidence envelope that holds.
12 minGDPR for document AI: a practical guide for operators
Article 28 obligations, lawful basis, sub-processor governance, data subject rights, and what your DPA actually needs to say.
9 minSub-processors in AI: what your DPA needs in 2026
AI products invoke 4–7 sub-processors per request. What your DPA needs to say about LLM providers, observability vendors, and zero-retention APIs in 2026.
9 min