Every enterprise has pilots. Few have production systems. Deloitte's 2026 State of AI in the Enterprise report quantified what practitioners already knew: 86% of companies are increasing AI budgets, but only 11% have agentic AI running in production. The money is flowing. The outcomes are stuck in staging environments.
Executive Summary
- Budgets are surging, production is lagging. 86% of enterprises are increasing AI budgets in 2026. But agentic AI, the category with the highest potential return, sits at just 11% production deployment. The rest is pilots, proofs of concept, and "strategic exploration."
- Agentic AI is accelerating faster than any prior wave. Deployments nearly doubled in four months, from 7.2% to 13.2% between August and December 2025. Banking KYC/AML workflows are seeing 200-2,000% productivity gains where agentic systems are live.
- Skills, not technology, are the bottleneck. 46% of tech leaders cite AI skill gaps as their biggest obstacle. 95% of developers use AI coding tools weekly, but the gap between using Copilot and architecting production AI systems is vast.
86%
of companies increasing AI budgets in 2026
Source: Deloitte, State of AI in the Enterprise, 2026
11%
of enterprises with agentic AI in production
Source: Deloitte, State of AI in the Enterprise, 2026
171%
average ROI from enterprise AI deployments that reach production
Source: Deloitte, State of AI in the Enterprise, 2026
46%
of tech leaders citing AI skill gaps as their biggest obstacle
Source: Deloitte, State of AI in the Enterprise, 2026
95%
of developers using AI coding tools weekly
Source: GitHub, Developer Survey, 2025
2,000%
peak productivity gains in banking KYC/AML with agentic AI
Source: Deloitte, Financial Services AI Report, 2026
Why Pilots Stall
The pattern is consistent across industries and geographies. A team builds a compelling pilot. The demo goes well. Leadership approves a budget. Then nothing ships.
Three forces kill the transition from pilot to production.
Integration debt. Pilots run on clean data in isolated environments. Production means connecting to legacy systems, handling edge cases, managing authentication, logging audit trails, and surviving at scale. The pilot took six weeks. The integration work takes six months.
Governance gaps. Pilots skip governance because "we'll figure that out later." Later arrives, and the legal team, compliance team, and security team each have requirements that reshape the architecture. In regulated industries like banking and insurance, governance requirements can double the delivery timeline.
Ownership vacuum. Pilots are owned by innovation teams. Production systems need operational owners: people who monitor performance, retrain models, handle failures, and manage the AI through its full lifecycle. Most organisations have not created these roles.
The Agentic Acceleration
The agentic AI numbers tell a different story from the broader stall. Deployments nearly doubled in four months. Banking is leading, with KYC and AML workflows showing the most dramatic productivity gains on record for enterprise AI.
Enterprise AI: Budget vs Production Reality
Source: Deloitte, State of AI in the Enterprise, 2026
Why agentic AI is moving faster: it solves complete workflows, not fragments. A traditional AI tool summarises a document. An agentic system reads the document, extracts the relevant data, checks it against three databases, flags anomalies, drafts a report, and routes it for review. The business case is obvious. The productivity gain is measurable in hours saved per case, not seconds saved per task.
The production gap closes when organisations stop treating AI as a feature and start treating it as infrastructure. Infrastructure gets built, maintained, and scaled.
Isaac Rolfe
Managing Director
The Skills Constraint
46% of tech leaders name skills as their primary obstacle. Not budget. Not data. Not executive buy-in. Skills.
The irony: 95% of developers already use AI tools weekly. The gap is not in AI awareness. It is in the specific capability to architect, deploy, and operate production AI systems. Using Copilot to write code is different from designing an agentic workflow that processes 10,000 insurance claims per day with auditability, fallback handling, and continuous monitoring.
This skills gap hits NZ and AU enterprises harder than larger markets. The talent pool is smaller. Competition for experienced AI delivery teams is fierce. Organisations that wait to build internal capability will find themselves competing for the same small group of experienced practitioners.
What Production Looks Like
Enterprises that have crossed the gap share common traits:
- They invested in AI infrastructure before AI features. Shared model management, prompt libraries, evaluation frameworks, and monitoring dashboards. The second capability ships in weeks, not months.
- They defined operational ownership from day one. Not "the data science team owns it." A named person owns the system's performance, reliability, and continuous improvement.
- They treated governance as an accelerator. Risk frameworks, data classification, and audit logging were built into the platform, not bolted on after the pilot.
- They measured business outcomes, not model accuracy. Claims processing time. Customer resolution rate. Compliance review throughput. Metrics that connect to revenue.
The 171% average ROI that Deloitte reports comes from organisations that did all four. The ones still stuck in pilot purgatory typically skipped at least two.