Article
→ Only 11% of organisations actively run agentic AI in production despite near-universal experimentation — the gap between AI enthusiasm and deployed value has never been wider. (Deloitte Insights, 2026)
→ Contextual AI's January 2026 launch of Agent Composer represents a category-defining bet: that enterprise RAG to AI agents is not a model-quality problem but an orchestration and context infrastructure problem.
→ The governance multiplier is real — organisations using AI governance tooling achieve over 12x more projects in production compared to those without, making infrastructure investment the highest-ROI decision in any agentic AI programme. (Databricks, 2026)
→ Early customers report root-cause analysis time reduced from 8 hours to 20 minutes and 60x faster issue resolution — not from better models, but from connecting agents to the right proprietary context. (PRNewswire, 2026)
Why This Matters Now
Four years after the ChatGPT inflection point, enterprise AI finds itself in a peculiar paradox. Investment is accelerating — 88% of senior executives say their organisations plan to increase AI-related budgets in the next 12 months (PwC AI Agent Survey, May 2025). Executive conviction is near-unanimous — 75% of leaders agree that AI agents will reshape the workplace more profoundly than the internet (PwC, 2025). And yet, by the most rigorous measure available, actual production deployment of agentic AI remains almost negligible: just 11% of organisations are actively running agentic AI systems in production, and only 14% have solutions ready to deploy (Deloitte Insights, 2026).
This is not a technology adoption lag in the conventional sense. It is something more specific and more solvable: a failure of the infrastructure layer that sits between powerful language models and the proprietary, often messy, operational data that makes those models useful in enterprise contexts.
The market inflection point is now visible. Gartner predicts that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% today — and that 15% of day-to-day work decisions will be made autonomously through agentic AI within the same timeframe (Gartner, cited by Deloitte, 2026). Multi-agent systems are already growing at 327% in under four months (Databricks, 2026). The RAG market alone is projected to grow from $1.96 billion in 2025 to $40.34 billion by 2035 (Roots Analysis, cited by Redis, 2025).
The organisations that close the production gap first will not necessarily have access to better models. They will have solved the harder problem: giving those models reliable, governed, action-capable access to institutional knowledge. That is the problem Contextual AI's Agent Composer is built to address — and understanding it requires rethinking what "enterprise RAG to AI agents" actually means.
What the Data Shows
The Production Gap: A Measurement Problem and a Real Problem
The first challenge in assessing enterprise AI readiness is the striking divergence between survey data sources. PwC reports that 79% of US companies say AI agents are already being adopted, with 66% of those reporting measurable productivity gains (PwC AI Agent Survey, 2025). Google Cloud's 2025 ROI Report, cited by Redis, puts 52% of enterprises using GenAI as now running AI agents in production, with 88% reporting positive ROI.
Deloitte's concurrent findings paint a radically different picture: 11% in production, 14% ready to deploy, 42% still developing a strategy roadmap, and 35% with no formal strategy whatsoever (Deloitte, 2026).
📘 Note
The divergence between PwC's 79% "adoption" figure and Deloitte's 11% "in production" figure is not a contradiction — it is a measurement problem. Survey respondents conflate using embedded AI features in commercial software (a Copilot summary, a chatbot widget) with genuinely deploying agentic workflows that take autonomous action across enterprise systems. The 11% figure is the more operationally meaningful benchmark for leaders making infrastructure decisions.
The pattern across all major research sources reveals a consistent structure:
| Metric | Data Point | Source |
|---|---|---|
| Organisations with agentic AI in production | 11% | Deloitte Insights, 2026 |
| Organisations with solutions ready to deploy | 14% | Deloitte Insights, 2026 |
| Organisations still developing agentic strategy | 42% | Deloitte Insights, 2026 |
| Organisations with no formal strategy | 35% | Deloitte Insights, 2026 |
| Organisations with mature AI governance models | 20% (1 in 5) | Deloitte State of AI, 2026 |
| Organisations truly reimagining (vs. optimising) with AI | 34% | Deloitte State of AI, 2026 |
| Predicted agentic AI project failure rate by 2027 | >40% | Gartner, cited by Deloitte, 2026 |
The Governance Multiplier
The Databricks 2026 State of AI Agents Report, drawing on data from over 20,000 organisations including more than 60% of the Fortune 500, provides the most quantitatively compelling argument for infrastructure-first AI strategy. Companies using evaluation tools get nearly 6x more AI projects into production. Companies using governance tools get over 12x more projects into production (Databricks, 2026).
🔴 Important
The 12x governance multiplier is arguably the single most significant finding in recent enterprise AI research. It means that for every project a governance-light organisation gets to production, a governance-mature peer gets twelve. No model quality differential — not GPT-4 versus a smaller competitor, not fine-tuning, not prompt engineering — produces an effect of this magnitude. Infrastructure and accountability frameworks are the true differentiators.
Multi-agent systems, which were negligible in enterprise deployments as recently as mid-2025, grew 327% in under four months per the same Databricks report. The category is not emerging — it is already in rapid expansion among the cohort of organisations that have solved the infrastructure problem.
The Business Case
For organisations that have successfully aligned AI, platforms, and business strategy, the financial returns are substantial. Accenture's research finds that such organisations achieve on average 2.2x revenue growth and a 37% EBITDA lift compared to peers (Accenture, 2025). Worker access to AI rose 50% in 2025 alone, and the number of companies with 40% or more of AI projects in production is set to double within six months (Deloitte State of AI, 2026).
These returns are not evenly distributed. They accrue disproportionately to organisations that have made the transition from passive retrieval-augmented generation (RAG) — systems that answer questions — to agentic RAG systems that take action.
How Leading Organisations Are Responding
Contextual AI: The Context-Centric Bet
Founded in June 2023 by Douwe Kiela — formerly head of research at Hugging Face and research scientist at Facebook AI Research, and now an adjunct professor in symbolic systems at Stanford — Contextual AI has built its entire product thesis around a contrarian but increasingly validated claim. As Kiela stated at the January 2026 launch: "The model is almost commoditized at this point. The bottleneck is context — can the AI actually access your proprietary docs, specs, and institutional knowledge? That's the problem we solve." (VentureBeat, 2026)
Agent Composer, launched on January 27, 2026 and backed by Bezos Expeditions and Bain Capital Ventures, operationalises this thesis as a production-grade orchestration platform. It is specifically positioned for industries where general-purpose agents fail: aerospace, semiconductor manufacturing, and advanced manufacturing — sectors characterised by complex engineering documentation, multi-system IT environments, and zero tolerance for hallucination.
The platform offers three agent-creation pathways:
| Creation Mode | Description | Best For |
|---|---|---|
| Pre-built Templates | Industry-specific workflow templates for common technical tasks | Teams needing fast deployment on known use cases |
| Natural-Language Generation | Describe desired agent behaviour; system generates architecture | Teams with clear requirements but limited ML engineering capacity |
| Custom Blank-Canvas Builder | Full control over agent logic, retrieval pipelines, and toolchains | Teams with specialised requirements or existing infrastructure |
Contextual AI describes Agent Composer as "an infrastructure and orchestration layer that manages context, enforces guardrails, and maintains agent reliability throughout multi-step engineering workflows" (PRNewswire, 2026). The emphasis on guardrails and reliability throughout multi-step workflows — not just at inference time — distinguishes this from simpler RAG implementations.
💡 Tip
Kiela's description of early RAG is instructive for any team evaluating their current state: "Early RAG was pretty crude — grab an off-the-shelf retriever, connect it to a generator, hope for the best. Errors compounded through the pipeline. Hallucinations were common because the generator wasn't trained to stay grounded." (VentureBeat, 2026). The question for your organisation is not whether you have RAG — it is whether your RAG system compounds errors or actively manages them.
Advantest: Deploying Across Teams and Customer Ecosystems
Advantest, a global leader in semiconductor test equipment, offers one of the clearest examples of vertical AI agent deployment at scale. Keith Schaub, VP of Technology and Strategy at Advantest, noted that "Contextual AI has been an important part of our AI transformation efforts. The technology has been rolled out to multiple teams across Advantest and select end customers." (PRNewswire, 2026)
This cross-team and customer-facing deployment pattern is significant. It indicates that Agent Composer is being used not merely as an internal productivity tool but as a platform capability embedded into Advantest's own service delivery — a more demanding and governance-intensive deployment model than simple internal chatbots.
High-Performance Early Adopters: Measurable Operational ROI
Three early customer cases from the Agent Composer launch illustrate the range of operational improvements achievable when enterprise RAG systems gain agentic capabilities:
Advanced Manufacturer — Root-Cause Analysis: By automating sensor data parsing and log correlation across multiple systems, root-cause analysis time dropped from 8 hours to 20 minutes (PRNewswire, 2026). This is not a marginal improvement — it represents a 97.5% reduction in time-to-insight for a critical operational workflow.
Tech-Enabled 3PL Provider — Issue Resolution: Using Agent Composer to navigate and act across an internal knowledge base, the provider achieved 60x faster issue resolution (PRNewswire, 2026). At this level, the agent is not supplementing human research — it is replacing the research phase entirely and enabling near-real-time decision support.
Global Strategy Consulting Firm — Research Retrieval: Manual research time for case work retrieval and complex Q&A was reduced from hours to seconds (PRNewswire, 2026). For a sector where analyst time is the primary cost of goods, this directly alters the unit economics of project delivery.
The Hidden Risk: What Most Teams Get Wrong
The dominant misunderstanding in enterprise AI deployment is the conflation of two fundamentally different system architectures: RAG systems that answer questions, and agentic systems that take action. Most organisations building "RAG" today are building the former while expecting the performance characteristics of the latter.
This is what Contextual AI's framing describes as the "hands" problem. A passive RAG system retrieves relevant documents and generates a response. An agentic RAG system can file a support ticket, update a database record, trigger an approval chain, or query a real-time sensor feed — and do so reliably, across multi-step workflows, with audit trails and guardrails.
⚠️ Warning
The majority of organisations describing themselves as "using AI agents" are using software with AI features embedded — not deploying autonomous, action-taking agents connected to production systems. This distinction matters enormously when evaluating your competitive position. PwC's finding that 79% of companies have "adopted" AI agents (PwC, 2025) and Deloitte's concurrent finding that only 11% are in production (Deloitte, 2026) reflect this definitional gap. Do not benchmark against adoption statistics. Benchmark against production deployment rates.
The second critical risk is what Deloitte identifies as the workflow layering fallacy: the tendency to add agents on top of existing processes rather than redesigning operations to be agent-compatible (Deloitte, 2026). This produces AI systems that are structurally constrained by the inefficiencies of the workflows they are meant to improve. The result is incremental productivity gain at best, and costly technical debt at worst.
Gartner's prediction that over 40% of agentic AI projects will fail by 2027 specifically cites legacy infrastructure incompatibility as the primary cause (Gartner, cited by Deloitte, 2026). This is not a failure of model quality or prompt engineering — it is a failure of architecture and governance readiness.
The third risk is workforce readiness, which multiple major research reports identify as underestimated. EY's 2025 agentic AI workplace survey identifies three core barriers to adoption: excitement without reassurance, unclear AI communication to employees, and lack of training. The finding is framed starkly: "Unchanneled worker enthusiasm squanders agentic AI's promise" (EY, 2025). PwC's research corroborates this, finding that 68% of senior executives report that half or fewer of their employees interact with agents in everyday work — despite 75% of those same executives believing agents will reshape the workplace more than the internet (PwC, 2025). Executive conviction and employee reality are not aligned.
⚠️ Warning
Infrastructure investment without workforce engagement programmes produces agents that are technically functional but organisationally unused. The 12x governance multiplier identified by Databricks applies to technical governance — but people governance (clear communication, training, role redesign) is the equally critical parallel investment that most vendor roadmaps omit entirely.
A Framework for Moving Forward
The Five-Layer Agentic Readiness Model
Moving from enterprise RAG to production-ready AI agents requires simultaneous maturity across five interdependent layers. Weakness in any single layer constrains the performance ceiling of the entire system.
| Layer | What Maturity Looks Like | Common Failure Mode |
|---|---|---|
| 1. Context Infrastructure | Proprietary documents, specs, and operational data are indexed, versioned, and accessible to agents in real time | Agents connected to stale, unstructured, or incomplete knowledge bases — the root cause of hallucination in production |
| 2. Orchestration Architecture | Multi-step workflows are defined, monitored, and recoverable; agents can chain tools, APIs, and sub-agents with error handling | Single-step RAG masquerading as agentic capability; no recovery logic for workflow failures |
| 3. Guardrails and Grounding | Hallucination resistance is enforced at inference time; outputs are traceable to source documents; out-of-scope actions are blocked | Agents with access to production systems but no constraints on what they can do or say |
| 4. Governance and Audit | Every agent action is logged; access controls are role-based; compliance requirements are embedded in agent behaviour | Agents deployed without audit trails — ungovernable in regulated industries and a liability exposure everywhere else |
| 5. Workforce Integration | Employees understand what agents do and do not do; workflows are redesigned around agent capabilities; feedback loops exist | Agents deployed into unchanged workflows; low adoption; "pilot purgatory" with no path to scale |
💡 Tip
The organisations achieving 12x more production deployments (Databricks, 2026) are not starting with Layer 1 and hoping Layers 4 and 5 follow. They are building governance infrastructure concurrent with — or before — context infrastructure. Reverse this sequence at your peril.
The Three Horizons of Enterprise RAG to AI Agents
A second useful frame for investment prioritisation is a horizons model that maps the natural progression from passive RAG to fully agentic multi-system operations:
Horizon 1 — Grounded Retrieval (0–6 months): The immediate foundation. Proprietary data is indexed with high-quality embeddings; retrieval is accurate and hallucination-resistant; outputs cite sources. The system answers questions reliably. This is where most organisations currently sit — and where they stall.
Horizon 2 — Agentic Workflow Automation (6–18 months): The agent gains "hands." It can take single-domain actions — filing tickets, querying live data, generating structured reports, triggering approvals — within defined guardrails. Multi-step workflows are possible but scoped. This is the target state for most Agent Composer deployments in 2026.
Horizon 3 — Multi-Agent Orchestration (18–36 months): Specialised agents coordinate across domains. A root-cause analysis agent queries sensor data, a documentation agent retrieves relevant engineering specs, and a workflow agent files the corrective action request — all triggered by a single operational event, with human oversight at defined checkpoints. This is the architecture Gartner forecasts will represent 33% of enterprise software by 2028.
What This Means for Your Organisation
The research is clear and consistent across Deloitte, Databricks, PwC, EY, and Accenture: the organisations that will capture disproportionate value from agentic AI are not those with the most sophisticated models. They are those that have solved the context, orchestration, and governance infrastructure problems first.
Your immediate priorities should be:
1. Audit your current RAG architecture against the Five-Layer Readiness Model. Most enterprise RAG deployments are strong at Layer 1 (basic retrieval) and weak at Layers 3, 4, and 5. The audit will identify your binding constraint. If your agents cannot trace outputs to source documents, you do not have a production-ready system — regardless of what your pilots have demonstrated.
2. Reframe your AI investment thesis around context quality, not model selection. Kiela's observation that models are effectively commoditised is directionally correct and strategically important. Investments in proprietary data infrastructure, knowledge base quality, and retrieval accuracy will compound in value as models continue to improve. Investments in specific model versions will not.
3. Implement governance infrastructure before expanding agent scope. The Databricks 12x multiplier is not a correlation — it reflects a causal mechanism: governance tooling forces organisations to define what agents are permitted to do, which in turn makes those agents deployable in production environments with compliance, audit, and security requirements. Build the accountability layer before you scale the capability layer.
4. Design for workforce integration from day one. EY's finding that "unchanneled enthusiasm squanders agentic AI's promise" (EY, 2025) is a practical directive: your employees likely want to work with AI agents, but without structured training, clear communication about agent roles, and redesigned workflows, that enthusiasm dissipates into shadow usage and workarounds. Treat workforce integration as a technical requirement, not a change management afterthought.
5. Prioritise vertical depth over horizontal breadth in your first production deployments. Contextual AI's deliberate focus on aerospace, semiconductors, and manufacturing — sectors where hallucination has engineering consequences — reflects a broader principle: production-grade agents require domain-specific context quality that horizontal platforms cannot provide out of the box. Identify the two or three workflows in your organisation where agentic RAG has the highest ROI and the most tractable data environment, and build to production depth on those before expanding.
🔴 Important
The competitive window for establishing agentic AI as a core operational capability is narrower than most organisations recognise. Worker access to AI rose 50% in 2025 alone, and the number of companies with 40%+ of AI projects in production is set to double within six months (Deloitte, 2026). The organisations establishing production infrastructure now will have compounding operational advantages — in cycle times, decision quality, and institutional knowledge capture — that become increasingly difficult to close.
Conclusion: The Path Forward
The launch of Contextual AI's Agent Composer is a useful lens on a broader market inflection: enterprise AI is transitioning, irreversibly, from a model-selection problem to an orchestration-and-context problem. The organisations that recognise this shift early — and invest accordingly in the infrastructure layers that connect powerful models to proprietary institutional knowledge — will not merely improve productivity metrics; they will restructure the unit economics of their operations. The production gap is real, but it is not inevitable. It is the direct consequence of underinvestment in the context, governance, and orchestration layers that turn retrieval systems into operational agents. That gap will close in the next 24 months. The question is whether your organisation leads that closure or responds to it.
Sources
- Contextual AI Launches Agent Composer — PRNewswire, 2026: https://www.prnewswire.com/news-releases/contextual-ai-launches-agent-composerai-for-when-it-actually-is-rocket-science-302670574.html
- Contextual AI Launches Agent Composer to Turn Enterprise RAG into Production-Ready AI Agents — VentureBeat, 2026: https://venturebeat.com/technology/contextual-ai-launches-agent-composer-to-turn-enterprise-rag-into-production
- Agentic AI Strategy — Deloitte Insights Tech Trends 2026: https://www.deloitte.com/us/en/insights/topics/technology-management/tech-trends/2026/agentic-ai-strategy.html
- The State of AI in the Enterprise — Deloitte, 2026: https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html
- PwC AI Agent Survey, May 2025 (n=300 senior executives): https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-agent-survey.html
- PwC 2026 AI Business Predictions: https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-predictions.html
- Databricks 2026 State of AI Agents Report (20,000+ organisations, 60%+ of Fortune 500): https://www.databricks.com/resources/ebook/state-of-ai-agents
- Accenture: The New Rules of Platform Strategy in the Age of Agentic AI, 2025: https://www.accenture.com/us-en/insights/strategy/new-rules-platform-strategy-agentic-ai
- Accenture Expands AI Refinery and Launches New Industry Agent Solutions, March 2025: https://newsroom.accenture.com/news/2025/accenture-expands-ai-refinery-and-launches-new-industry-agent-solutions-to-accelerate-agentic-ai-adoption
- EY: Unchanneled Worker Enthusiasm Squanders Agentic AI's Promise, 2025: https://www.ey.com/en_us/consulting/agentic-ai-workplace-survey
- Redis: Agentic RAG — How Enterprises Are Surmounting the Limits of Traditional RAG: https://redis.io/blog/agentic-rag-how-enterprises-are-surmounting-the-limits-of-traditional-rag/
- Redis: AI Agent Architecture — Build Systems That Work in 2026: https://redis.io/blog/ai-agent-architecture/
- Redis: Engineering for AI Agents: https://redis.io/blog/engineering-for-ai-agents/
- Google Cloud: A Developer's Guide to Production-Ready AI Agents: https://cloud.google.com/blog/products/ai-machine-learning/a-devs-guide-to-production-ready-ai-agents
- AWS Prescriptive Guidance: Agentic AI Patterns and Workflows: https://docs.aws.amazon.com/pdfs/prescriptive-guidance/latest/agentic-ai-patterns/agentic-ai-patterns.pdf
- TechFeed24: Contextual AI Launches Agent Composer — Bridging the Gap, 2026: https://www.techfeed24.com/posts/2026-02-03-contextual-ai-launches-agent-composer-bridging-the-gap-betwe
- TechBuddies: Contextual AI's Agent Composer Targets the Real Bottleneck in Enterprise AI, 2026: https://www.techbuddies.io/2026/01/30/contextual-ais-agent-composer-targets-the-real-bottleneck-in-enterprise-ai-context-not-models/
- mem0: Universal Memory Layer for AI Agents — GitHub: https://github.com/mem0ai/mem0
- NirDiamant: Agents Towards Production — GitHub: https://github.com/NirDiamant/agents-towards-production