AI Operating System: The Enterprise's Second Brain

Article by Asutosh Padhi, Brian Elliott, Daniel Swan, Kim Baroudy and Vik Sohoni (senior partners, McKinsey), presented as expressing the views of QuantumBlack, AI by McKinsey, published in October 2026 (exact day unspecified, dated the 1st by convention); 12 pages, 2 exhibits, a table, an architecture sidebar and two cases.

By Asutosh Padhi / October 1, 2026 / Source mckinsey.com ↗ / Reading 2 min / .md / / Auto-verified translation

#AI operating system#AI OS#second brain#second brain#AI operating system#semantic layer#enterprise ontology#knowledge graph

The article opens with a scenario. The CEO of an international insurer learns that regulators want to cap margins on certain products. Within minutes, the company's AI OS reads the draft rules, cross-checks them against the catalog, and identifies roughly 140 affected products across seven countries, a third of them at risk, while flagging that each regulator defines margin differently. Country heads sign off; one of them challenges a cost-allocation assumption and brings the number of non-compliant products down to nine. The next day, the system models the options (repricing, redesign, commissions, withdrawal), the committee settles on a plan that the system logs as guardrails, then agents generate designs and prices submitted to actuaries, compliance, and line managers. After launch, the system monitors each product; the next regulation will be handled faster.

The authors start from a gap: according to McKinsey's State of AI 2026 survey, 40% of large enterprises have deployed agents at scale, but only 6% attribute at least 5% of their EBIT to AI. The challenge is no longer adoption but how the enterprise operates. The AI OS, the enterprise's second brain, maintains context, reasons across functions, recommends or executes decisions, and learns from outcomes. It builds on what already exists (infrastructure, data already in place, models from multiple providers) and adds a semantic layer made up of ontologies and knowledge graphs, deterministic analytical models called by agents, reusable agent orchestration, and an experience layer.

Six priorities guide the build: own the context and rent or replace the technology; limit the number of AI OSs to one, two, or three; redesign processes end to end; design for reuse so the advantage compounds; factor in AI economics, since token cost has fallen by more than 95% while spend has tripled; start from technologies already paid for.

Management must change: act on signal rather than on a calendar, steer continuously, devote leadership time to judgment, and work through direct connections rather than layers, to the point of considering 20 to 40 direct reports per manager. A global OEM and a US industrial company illustrate the approach. The authors recommend targeting three to five domains, setting semantic governance, funding the shared AI OS, and redefining decision rights.

Key takeaways

Paradox. 40% of large enterprises have agents at scale, 6% see ≥5% of EBIT attributable to AI (State of AI 2026, flat year over year).

Definition. AI OS = continuous management layer — context, cross-functional reasoning, recommendation/execution, learning; a real-time cockpit, not a retrospective dashboard.

Scenario. insurer, ~140 products across 7 countries, divergent margin definitions; 9 products non-compliant after an assumption is challenged; committee decisions logged as guardrails; post-launch monitoring.

Architecture. existing stack left intact (proprietary cloud or on-prem as the target, data already in place, multi-provider LLMs including open-weight, no training on the data) + semantic layer, deterministic ML models as tools, reusable agents with model routing, experience layer.

Rule. Own the context; rent or replace the technology; 1 to 3 AI OS max; end-to-end redesign (industrial company: $140-180M/year on sales and operations).

Economics. token −95%, spend ×3; the cheapest model costs more if it produces costly errors.

Management. signal > calendar; always-on plan, bottleneck shifted to the field and control functions; judgment > relay; 10-12 → 20-40 direct reports (open question).

Actions. 3 to 5 domains, semantic governance, fund the shared AI OS rather than individual use cases, redefine human/agent decision rights.

⚠️ Scope: gains (×10, 12 weeks, −50% time-to-ROI, −90% cycle times) without method or scope disclosed; anonymous cases.

AI Operating System: The Enterprise's Second Brain

Key takeaways

Key figures

Attributed claims

AI Operating System: The Enterprise's Second Brain

Summary

Key takeaways

Key figures

Attributed claims