Enterprise AI adoption has reached a paradox. The tools are better than ever. The infrastructure is available. The budgets have been approved. Yet most organisations remain trapped in a cycle of promising pilots that never reach production. New research from MIT IDE, BCG and Infosys points to a single underlying cause: companies are still treating AI agents as software to be installed, rather than colleagues to be integrated.
The implications for enterprise buyers and suppliers are significant. If the technology is no longer the binding constraint, then the competitive advantage shifts from who has the best model to who can reorganise their business around the new capabilities. That is a very different procurement decision — and a very different selling proposition.
The pilot trap by the numbers
BCG's latest research, cited by SAP's global head of Business AI Platform Ecosystem, finds that 74% of organisations are stuck in pilot mode with no plan to scale. A separate survey by Anthropic found that 42% of enterprises are struggling with data access and quality, the foundational layer without which agents cannot operate reliably. And MIT IDE's analysis of SEC filings reveals that by the end of 2025, only 11% of S&P 500 firms had AI deeply embedded in their business — with two-thirds of that integration happening at technology companies. Financial services, despite heavy pilot activity, achieved deep integration at roughly 4%.
These are not technology failures. They are organisational failures. The agents work. The data is available. The platforms are secure. What is missing is the operating model that turns an agent from a tool into a productive member of the workforce.
From tool to teammate: the digital coworker shift
Satinder Pal Singh, associate vice president for SAP Business AI Platform and SAP AI Strategy at Infosys, frames the central challenge clearly. After surveying more than 500 global enterprises, Infosys and HFS Research found that only 14% have successfully scaled AI. Eighty-six percent remain in early exploration, proof of concept or fragmented pilot phases. For a technology that has topped every boardroom agenda for two years, that is a striking gap.
Singh describes the root cause as the "AI as a department" trap. An organisation creates an AI centre of excellence, walls it off from the business, and tasks it with building technically excellent pilots that the rest of the company does not recognise as solutions to its problems. The result is AI systems searching for problems rather than solving them.
The alternative is what Singh and others are calling the "digital coworker" model. An AI agent is not a piece of software that automates a task. It is a colleague that operates within business processes, requires onboarding, needs access to the right information, must be supervised, and can be held accountable for its outputs. That mindset shift, Singh estimates, accounts for roughly 70% of the implementation challenge.
Why governance speeds adoption, not slows it
Thorsten Leiduck, SVP and global head of SAP Business AI Platform Ecosystem, argues that the absence of governance does not accelerate AI adoption. It creates a risk backlog that eventually forces a more disruptive stop and audit. Clear policies, access controls, monitoring and oversight are necessary not because regulators demand them, but because they are the mechanisms by which an organisation learns to trust an agent enough to give it meaningful work.
The practical implication is that governance is not a compliance layer added after deployment. It is the enabling condition for scale. Enterprises that establish governance early are the ones that move from pilot to production. Those that defer it are the ones that accumulate risk until the programme stalls.
OpenAI's internal lesson: training matters more than capability
MIT IDE research on OpenAI's own adoption of its Codex agentic tool provides vivid evidence. Codex went from under 10% of the average worker's AI output in August 2025 to 99.8% by June 2026. The jump coincided not with a model upgrade but with company-wide training sessions and regular feedback loops. Legal and recruiting teams went from near-zero Codex use to nearly 90% of their AI output in approximately five months.
The lesson is that adoption follows enablement, not availability. Agents that sit in a platform waiting to be discovered will not be used. Agents that are introduced with training, role-specific workflows and feedback mechanisms become the default way of working. For enterprise buyers, this means that the procurement checklist needs to include implementation support, change management and training as first-class requirements, not afterthoughts.
What it means for enterprise buyers
For procurement and operations teams evaluating agentic AI platforms, the research carries three practical messages.
First, evaluate the vendor's implementation model, not just the product features. A platform with excellent technical capabilities but no structured onboarding, training and feedback process will likely join the 74% that never scale. Ask for evidence of how the vendor moves customers from pilot to production, and what governance tooling is included from day one.
Second, plan for workflow redesign, not tool substitution. The largest productivity gains, according to the Codex research, arise when firms reorganise production around the new technology rather than merely substituting it into existing workflows. That requires business process expertise, not just technical expertise. Buyers should ensure their implementation partner or internal team has both.
Third, treat governance as a scale enabler, not a brake. The organisations that slow down to establish clear policies, access controls and monitoring early are the ones that accelerate past the pilot phase. Those that try to add governance later typically find themselves halting programmes that have already run ahead of their risk appetite.
What it means for suppliers
For companies selling agentic AI platforms or services, the research resets expectations in two ways.
The first is that the competitive moat is shifting from model capability to implementation reliability. As frontier models commoditise, buyers will differentiate between vendors based on who can reliably move them from pilot to production — with governance, training and measurable outcomes intact. A vendor that ships software without a structured adoption programme is selling a license to fail.
The second is that the customer success function is now a revenue function. Vendors that invest in training, feedback loops and workflow redesign as core product capabilities will retain customers and expand accounts. Those that treat these as professional services add-ons will churn when the pilot does not convert.
The Agentic Expo angle
Agentic Expo exists because the enterprise AI market has moved past the question of whether agents can perform tasks. The question now is whether enterprises can organise themselves to benefit from what agents can do. The technology is ready. The governance frameworks are forming. The budgets are allocated. What remains is the operational and cultural shift that turns a pilot into a production asset.
At Olympia London on 23 to 24 March 2027, the suppliers who have solved the implementation problem will meet the buyers who need to solve it. The conversations will be about adoption programmes, governance models, change management and the evidence that an agent investment has produced a business outcome. That is the work that separates the 14% who scale from the 86% who do not. It is also the work Agentic Expo is built to advance.
Sources: MIT Technology Review, "Building AI that can work across the enterprise", 16 November 2026; MIT Initiative on the Digital Economy, "AI Adoption in 2026: What MIT Research Tells Us About the State of Enterprise AI", November 2026; BCG, "74% of organisations stuck in pilot mode with no plan to scale", 2026; Genpact / HFS Research enterprise AI adoption survey, 2026.