Deloitte: 40% of Enterprise AI Projects Moving to Production Within Six Months

Production deployment is replacing pilot fatigue

Worker access to enterprise AI increased 50% during 2025, and the number of companies with at least 40% of their AI projects in production is expected to double within six months, according to Deloitte's 2026 State of AI in the Enterprise report. The forecast suggests budgets are shifting from experimentation toward deployment operations: identity controls, data integration, model monitoring, and workflow redesign.

The Deloitte figure is a survey-based expectation, not a measured production count, but it aligns with other evidence that the pilot-to-production gap is closing. The implication for buyers is straightforward: vendors that cannot support production scale—access controls, monitoring, security, and workflow integration—are no longer worth evaluating. Proof of concept is table stakes. The question is whether a platform can run 40% of your AI workload in production without breaking your identity layer or compliance posture.

This favors platforms with enterprise-grade orchestration and governance: Microsoft Copilot, Salesforce Agentforce, Google Cloud Vertex AI, AWS Bedrock, and ServiceNow. It also creates demand for specialist vendors that sit between the model layer and enterprise systems—orchestration, evaluation, observability, and security.

Agent governance is now a funded category

Cymphony, an AI-agent security and governance company, launched with $30 million in funding co-led by Sequoia Capital and SMBC Fin Atlas Beyond Fund. The company's platform is designed to show how employees and AI agents access sensitive data, interact with critical systems, and create organizational risk.

The financing is evidence that agent governance is becoming a dedicated budget category, not a feature request tacked onto existing identity or data-loss-prevention systems. Enterprises deploying agents into Salesforce, ServiceNow, SAP, or internal databases need controls for delegated permissions, data exfiltration, tool use, and audit trails. Cymphony competes with Microsoft Purview and Entra, Palo Alto Networks, CrowdStrike, Wiz, CyberArk, and Netskope, but its differentiation is the combination of employee activity and autonomous-agent activity in one control layer.

Buyers should treat Cymphony as an emerging vendor rather than a proven alternative to incumbent platforms. The company has not disclosed customer counts, pricing, or benchmark results. The $30 million round tells you that investors believe agent security will scale; it does not tell you whether Cymphony's product works in production.

Open-weight models take traffic but not revenue

Chinese-built open-weight models processed most developer traffic on the largest model-routing marketplace in August 2026, while capturing only 4% of model-layer revenue during the tracked period, according to the Mozilla Foundation's State of Open Source AI v1.1 report published September 15.

The split between usage share and monetization matters for buyers evaluating whether to host models privately or rely on closed APIs from OpenAI, Anthropic, and Google. Open-weight models may reduce API dependence and improve data-control options, but token traffic does not prove enterprise production adoption. Buyers must separately evaluate model quality, support, indemnification, security updates, hosting costs, and compliance—areas where closed vendors may still have an advantage.

The Mozilla data also highlights a second risk: open-weight models with high traffic but low revenue may not sustain the level of post-release support and security patching that enterprise buyers require. If a model is free to run but expensive to maintain, the cost shifts from the API bill to your infrastructure and security teams.

Regulated-industry vendors are building sovereign deployment options

OpenText and Cohere announced a partnership on September 16 to combine OpenText's enterprise data and context systems with Cohere North and Cohere models. The companies said integrated offerings for governments and regulated industries are expected in early 2027, with deployment options across on-premises, private, public, and sovereign clouds.

The partnership competes with Microsoft's Azure OpenAI and Copilot stack, AWS Bedrock, Google Cloud Vertex AI, IBM watsonx, and Palantir's AI platform. Its main competitive angle is deployment flexibility for organizations that cannot place sensitive data in a standard public-cloud environment. Government, financial-services, healthcare, and defense buyers may view sovereign or on-premises deployment as a procurement requirement rather than an optional feature.

The announcement does not provide pricing, named customers, performance benchmarks, or production availability, so it is a strategic procurement signal, not yet evidence of deployed adoption. Buyers in regulated industries should track whether the early-2027 target holds and whether OpenText and Cohere can demonstrate production deployments with verifiable compliance outcomes.

What to watch

The Deloitte forecast creates a near-term test: if 40% production adoption doubles within six months, we should see corresponding increases in spending on orchestration, governance, and security vendors by Q2 2027. If that spending does not materialize, the survey expectation was aspirational rather than predictive.

For buyers, the immediate priority is ensuring your platform vendor can support production scale before your pilot count outgrows your governance layer. The difference between 10% and 40% of projects in production is not incremental—it is the difference between a controlled experiment and an operational dependency.