← /editions
#82

The orchestration layer eats the app

// Metrigy data shows enterprises are already retiring applications in favour of AI orchestration layers — and unbundling their audit trails in the process. Google counters with packaged vertical agents, while Snowflake ships scheduled autonomy on a default-open privilege.

$ today_in_10_seconds

One in ten organizations has already retired a major enterprise application because an AI orchestration layer took over its job, and half plan to by the end of next year. The part nobody's pricing: state management and audit logging were bundled into those applications for free. Pull the app out and they become procurement items.

$ lead_story

The orchestration layer is quietly becoming the application

Metrigy's Q2 2026 research across 759 companies globally found that 63.0% of CX and IT leaders expect an AI orchestration layer to eventually replace traditional enterprise applications like ticketing or CRM platforms — rising to 73.2% among the study's success group, defined by above-average improvements in key business metrics. Roughly 10.7% have already replaced at least one major enterprise application because of their orchestration layer, with 38.8% planning to by the end of 2026 and another 37.9% by the end of 2027. The mechanism is straightforwardly agentic: instead of a CRM acting as the monolithic host for context, coordinated agents read and write across connected tools simultaneously, pulling real-time context and executing end-to-end without a central database hub. Then comes the line that should stop you cold. Capabilities that traditional applications bundle natively — state management and audit logging among them — become separately priced infrastructure requirements under an orchestration layer, alongside a shift from predictable per-seat licensing to consumption pricing per API call or token. Roughly 51% of companies report that between 11% and 30% of their AI pilots get discontinued, which makes rebuilding core software functionality a high-cost bet with theoretical ROI.

// why_it_matters: You are not replacing an application — you are unbundling one, and the audit trail was in the bundle.

read more →

$ also_worth_attention

  • Google Cloud ships vertical agent packages for finance and legal

    Google Cloud launched Gemini Enterprise for Financial Services today — a Google-managed Financial Research agent, more than 50 new skills with specialized agentic instructions for financial roles, enterprise data connectors and a third-party agent ecosystem, in preview for capital markets and corporate banking, with CME Group and Deutsche Bank already using it. A parallel legal edition launched the same day with Cleary, Freshfields, Weil, and Williams & Connolly as launch customers, including MCP integrations into legal industry systems and skills covering brief drafting, citation verification, contract lifecycle management, regulatory horizon scanning, and DSAR fulfilment. This is the direct counter-bet to the lead story: buy the packaged vertical stack rather than assemble your own orchestration layer. Two very different risk profiles, and the packaged one keeps governance inside a vendor's boundary.

    link →
  • Snowflake's unattended agents inherit every role you hold

    Snowflake put CoCo automations into preview on Aug 21 — scheduled, recurring, unattended agent runs executing in a Snowflake-managed sandbox, each producing an inspectable Cortex thread. Per analysis of Snowflake's own documentation, each run executes as the creator's default role plus every default secondary role, and the EXECUTE AGENT TASK privilege is granted to PUBLIC by default — meaning any user in the account can schedule one unless an administrator changes the grant. Snowflake's docs are admirably blunt about the other half of the problem: task success means the run completed, not that every external side effect matched the intended outcome. Scheduled autonomy shipped before the permission model caught up. Check your grants today.

    link →
  • Info-Tech: the stack you built for pilots won't survive production

    Info-Tech Research Group published its Discover the Enterprise Agentic AI Technology Stack blueprint on Aug 19, arguing that piecemeal, pilot-era stacks assembled for quick wins expose organizations to integration brittleness, runaway costs, stale or untrusted data, and governance gaps as adoption scales. The six layers it maps: Application, Data & AI Lifecycle Management Tools, Foundational Models, Agentic Execution & Orchestration Engine, Data Platform, and Infrastructure. Read it next to the lead story and it stops being a framework and starts being an invoice — each layer is a thing you now buy separately because it used to come free inside the app you deleted.

    link →

$ bigger_picture

The trend behind the headlines

The interesting fight this week isn't capability, it's packaging. Metrigy's numbers say enterprises are genuinely dissolving applications into orchestration layers. Google's answer, shipped today, is the opposite bet: don't dissolve anything, buy a pre-governed vertical package with the connectors and skills already wired. Both routes are defensible. What neither route makes visible on the invoice is that the monolithic application was also a control plane — it held your state, your audit log, your role boundaries, and your definition of who was allowed to do what. Snowflake's default grant is the clean illustration: the execution layer arrived before the permission model, and the gap was filled by inheritance rather than design. Unbundle the app and you unbundle the controls. That bill lands a quarter or two after the licence savings.

$ tool_of_the_day

Temporal

Open-source durable execution. Workflows survive process crashes, restarts and redeploys because state is persisted and replayed rather than held in memory, and every step is recorded in a full event history. That history is the thing worth your attention this week: it doubles as an audit log of exactly what an agent did, in what order, with what inputs. If you're standing up long-running or scheduled agent runs and just realised nothing in your stack answers "what did it do and why," this is the boring layer to put underneath it before you need to explain yourself to an auditor.

explore Temporal →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime