Most mid-market companies are not short of agentic AI ambition, but are short of a starting position. PwC reports that 79% of organisations have adopted AI agents in some form; KPMG found only 11% run them at full production scale. The gap is not a technology problem. It is a starting-position problem. Drawing on Antoni Kozelski’s session at PyAI Conf 2026 and MIT NANDA research, this article maps the structured path from operational pain to deployed, production-grade agentic system with evidence from Vstorm’s work across 30+ engagements.

From “where do we start?” to production: how mid-market companies actually ship agentic AI

Most conversations about agentic AI implementation for mid-market companies focus on the technology. Which model. Which framework. Which platform. At the PyAI Conf 2026 “AI and ML in Production” panel in San Francisco, something different happened. When host Bryan Bischof asked each panelist to describe the last agent they had put into production, not one of them opened with a technology choice.

“[They] all started from the problem stage.”

Bryan Bischof, at 2:34, in the PyAI Conf 2026 “AI and ML in Production” panel

The companies that ship agentic AI do not begin with a technology decision. They begin with an operational one.

This article is our commentary and expansion on the topics discussed by the panellists. You can watch the full PyAI Conf 2026 “AI and ML in Production” panel here on Youtube or find it here through LinkedIn.

The gap is not a technology problem

The data on agentic AI pilot to production conversion tells a consistent story. According to PwC’s May 2025 survey of 308 US business executives, 79% of respondents said AI agents are already being adopted in their companies. KPMG’s Q1 2025 AI Quarterly Pulse Survey found that only 11% of organisations had deployed agents at full production scale. These are separate surveys with different samples and methodologies, but the direction they point is the same. Adoption is broad. Operational deployment is rare.

Gartner sharpens the picture further. In a press release dated June 25, 2025, the firm predicted that over 40% of agentic AI projects will be cancelled by end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

“Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,”

Anushree Verma, Senior Director Analyst at Gartner

The failure pattern is not primarily technical. The models are capable. The tooling has matured. What is missing, in most cases, is a clear starting position, a disciplined answer to the question of where agentic AI creates genuine operational leverage before any engineering begins.

How mid-market companies currently approach AI adoption

At Vstorm, we observe three recurring patterns in how mid-market organisations typically approach agentic AI implementation, and none of them reliably produces a production system.

The first is technology-first adoption: the organisation selects a platform or LLM, then looks for problems it can solve. The result is solutions in search of use cases.

The second is pilot theater: a proof-of-concept is built that works in isolation, then fails to integrate into existing workflows, and is shelved. This is the pattern Gartner identifies as hype-driven misapplication.

The third is the consultancy handoff gap: a strategy firm produces a roadmap, and a separate engineering partner discovers it cannot be implemented as specified, because the roadmap was built without engineering input. Strategy and execution are two different vendors, and the handoff is where coherence is lost.

What these three patterns share is that they all begin somewhere other than an operational problem. They begin with a tool, a slide deck, or a vendor relationship. The 5% of organisations that cross the divide, to borrow MIT NANDA’s framing, do not begin there.

Start with the problem, not the technology

The PyAI panel observation carries a practical insights. For mid-market organisations attempting agentic AI implementation, the starting point question breaks into three concrete questions that belong in first contact, not at the end of a scoping process.

• The first is: Where is the operational pain? The processes worth automating are cross-departmental, repetitive, dependent on specialised knowledge, and currently slow or error-prone. If the answer to “what breaks?” is “nothing, we just want to be more efficient,” the use case is not ready.
• The second is: What does success look like, measured how? Not “AI will improve our efficiency” but a specific, pre-agreed metric: processing time per case, error rate, throughput per operator. Without a baseline, there is no way to confirm the system is working.
• The third, and most predictive, is: Who on the client side will own the system after it is built? Without a named internal owner before engineering begins, systems are not maintained, not adopted, and not trusted by the people who need to use them.

These three questions are the essence of Transformation Consulting, which, in our experience, is discovery work as much as it is strategy work. Most mid-market companies feel the operational pain but cannot articulate the automation opportunity. That articulation is the first deliverable.

What the path from roadmap to production actually involves

Once the problem is articulated, the path to a deployed system runs through three stages. We follow this structure in every engagement through our TriStorm methodology.

Transformation Consulting identifies where agentic AI creates the highest operational leverage, maps current-state processes in detail, and builds the business case and ROI model per use case before engineering begins. This is not optional, in our opinion, it is where we determine whether the right problem is being solved.

Technology Consulting then translates the roadmap into an architecture blueprint. It selects the right stack, defines integration points with existing infrastructure; CRM, ERP, legacy databases, proprietary systems; and specifies data and security requirements. This stage determines whether the project is actually buildable before a line of code is written. In our experience, this is where most off-the-shelf solutions reveal their limits: as the required process does not fit the predesinged template.

Agentic AI Engineering then builds, deploys, and instruments the production system. Every agent ships with tailored observability so the client team can trace decisions, audit outputs, and debug in production. Knowledge transfer is built into delivery, as it is essentially for running a profitable agentic workflow.

The three stages run under one team, without a handoff gap between the roadmap and the build. That continuity is the structural answer to the consultancy handoff problem that sinks so many promising projects.

Why 95% of pilots produce no measurable return and what the 5% do differently

MIT Media Lab’s Project NANDA published its State of AI in Business 2025 report in July 2025, based on systematic review of over 300 publicly disclosed AI initiatives, 52 structured interviews, and 153 senior leader survey responses. The headline finding: despite $30–40 billion in enterprise spending on generative AI, only 5% of integrated AI pilots extract measurable value. The remaining 95% stall with no measurable P&L impact.

A clarification worth making: this 95% figure refers specifically to pilots failing to deliver measurable financial return, not simply failing to reach production. A system can reach production and still deliver no ROI. The MIT NANDA report identifies the root cause not as model quality, infrastructure, or regulation, but as what it calls the learning gap. Most deployed systems do not retain feedback, adapt to context, or improve over time. They are built to demonstrate a working pilot, not to operate on living workflows.

The structural pattern behind the 5% that succeed is consistent across the research: they start from operational pain, integrate deeply into specific workflows, and are evaluated on business outcomes rather than technical benchmarks. This is the same pattern the PyAI panel demonstrated and the same discipline that the three-stage TriStorm structure is built to enforce.

One finding from the MIT NANDA report is particularly relevant for mid-market buyers: external partnerships succeed roughly twice as often as internal builds.

“Almost everywhere we went, enterprises were trying to build their own tool. But the data showed purchased solutions delivered more reliable results.”

Aditya Challapally, Lead Author of State of AI in Business 2025, in Fortune

The security sign-off is harder than the technical buy-in

One of the PyAI panel’s most practically useful observations came at the 7:07 mark: selling the CTO on an agentic system is relatively straightforward. Getting security sign-off is where projects stall.

For mid-market organisations, this surfaces as data boundary questions, regulatory compliance requirements; the EU AI Act for European organisations; and decisions about what the agent can act on autonomously versus what requires a human in the loop. These are not afterthoughts. They need to be addressed at the architecture stage, before engineering begins. Gartner explicitly lists inadequate risk controls as one of the three primary reasons agentic AI projects will be cancelled by 2027.

The practical implication is that security and governance belong in Technology Consulting, not in a post-deployment audit. When data boundaries, access controls, and observability requirements are defined before the system is built, they become architectural constraints rather than retrofit problems. When they are defined too late, they become blockers.

Why team structure affects agentic AI workflow automation outcomes

A point from Antoni’s panel segment that deserves expansion: the traditional project team structure (front-end developer, back-end developer, QA engineer, project manager) was designed for conventional software delivery and is poorly suited to agentic AI.

The reason is architectural. Agentic systems require someone who understands the business process, the integration surface, the model behaviour, and the failure modes simultaneously. Splitting that understanding across four specialists creates coordination overhead at every decision point, and decisions in agentic AI delivery are frequent, interdependent, and context-sensitive.

A single forward-deployed AI engineer, working directly within the client’s operational context rather than building remotely against a specification, handles a delivery surface that previously required multiple specialists. A Transformation Manager holds the business context. An AI architect supports both in the background. This structure removes the communication lag between what the client needs and what gets built, and produces a system shaped around the actual process, not a brief written before discovery was complete.

Parting thoughts

The gap between agentic AI adoption and production deployment is not a technology problem, it is a starting-position problem. Agentic AI implementation for mid-market companies that succeeds shares three characteristics: it begins with a structured discovery of operational pain rather than a technology selection; it builds the business case and architecture before engineering begins; and it treats security, governance, and internal ownership as design constraints, not afterthoughts. The companies closing the gap fastest are not spending more than the companies that stall. They are spending differently, and starting at the right place.