Blog

01 Annotation

Agents are moving into production faster than results are reaching the P&L: the share reporting EBIT impact has not moved in a year while agent use and token consumption climb. That is the gap our whitepaper The AI ROI Mirage describes, now with a running cost attached, and the report's own advice is to treat operating costs as a design constraint rather than an afterthought.

Maps to: The AI ROI Mirage (Pilot Paradox) and The Agentic Blast Radius (Excessive Agency)

01 Annotation

Two findings matter most for AI programmes. Attackers use AI chiefly to impersonate people, which defeats approval processes rather than firewalls, and AI applications are now a target in their own right. Both call for controls on what agents and staff can be talked into, the subject of our whitepaper The Agentic Blast Radius.

Maps to: The Agentic Blast Radius (Excessive Agency) and The Five Silent Pitfalls (Governance Gap)

01 Annotation

Wharton's April 2026 analysis of the McKinsey Lilli exposure is the clearest published case study to date of the failure pattern we describe in our whitepaper The Agentic Blast Radius. Three of its five control gaps are visible in the timeline. No human insider was needed. An autonomous agent ran reconnaissance, discovered 22 unauthenticated API endpoints, found the SQL injection path and reached full read-write access to the production database in under two hours. It was a researcher's test, and McKinsey reported no evidence of client data access beyond it; the point is what the same path would have given an attacker. That is the Gap 5 scenario made concrete: access that propagates faster than any conventional SOC escalation path can respond to it. The 46.5 million chat messages were not primary data. They were the memory layer of a GenAI platform: the accumulated conversational residue of tens of thousands of consultants working on strategy, M&A, and client engagements. Most organisations classify their CRM, their document management system, and their email archive as sensitive. Very few classify their agent memory, their vector stores, or their conversation logs with the same rigour. The unauthenticated API endpoints are the Gap 2 scenario: permission architecture scoped for the broadest possible task, not the narrowest current one. And the absence of a behavioural kill switch that could have terminated anomalous query patterns against the production database is Gap 5 in its operational form. The lesson is not that McKinsey got it uniquely wrong. The lesson is that the failure pattern (over-privileged agents, ungoverned memory, no machine-speed response capability) is the default pattern in most enterprise agentic deployments right now.

Maps to: Gaps 2, 4 & 5: Permission creep by design, Memory as a data leak, No kill switch

01 Annotation

EY's March 2026 Technology Pulse Poll quantifies what we describe in our whitepaper as the Governance Gap: over half of department-level AI initiatives are running without formal oversight, and 78% of leaders admit adoption has outpaced risk management. Most telling: 45% report confirmed or suspected data leaks from unauthorised AI tools in the past year alone. The "velocity paradox" EY identifies (the tension between speed-to-market and accountability) is precisely why we argue governance must precede deployment, not follow it.

Maps to: Traps 1, 2 & 4: Shadow AI, Sanctioned Leakage, Governance Gap

01 Annotation

Speaking at the Gartner Data & Analytics Summit this month, Deloitte's AI Institute head Beena Ammanath put it plainly: enterprises are caught between pressure from leadership to show AI value and foundations that are not ready. Her two blockers (data readiness and governance) mirror what we identify as the Governance Gap and the Skills Illusion. As she notes: without the right governance model, trust never forms, and adoption stalls. That is the trap. Not the technology; the organisational discipline around it.

Maps to: Traps 4 & 5: Governance Gap, Skills Illusion