The numbers tell a striking story. Nearly half of procurement teams are running AI pilots. Across insurance, three quarters of companies have implemented generative AI in at least one business function. Investment is accelerating. Ambition is high.

And yet the results are, by most measures, disappointing. BCG found that 74% of organisations struggle to achieve and scale value from their AI initiatives. In procurement specifically, Deloitte’s 2025 Global CPO Survey found that while 49% of teams are running pilots, only 4% have reached meaningful deployment at scale. In insurance, BCG’s own research shows that just 7% of carriers have successfully scaled AI beyond the pilot stage.

Something is going wrong, and it is not the technology.

The Real Barrier Is Human, Not Technical

The gap between deploying an AI tool and genuinely embedding it into how an organisation works is one of the defining challenges in procurement and insurance right now. And the evidence points consistently to people and process as the culprit, not algorithms.

MIT’s 2025 State of AI in Business report found that 95% of companies fail to achieve meaningful ROI from AI initiatives within six months, with the primary barriers being organisational rather than technical. BCG research confirms that 70–85% of AI projects fail to deliver expected benefits; a rate twice as high as traditional IT projects. Perhaps most telling: only around one third of companies in late 2024 were prioritising change management and training as part of their AI rollouts, despite the evidence that organisations that do invest in culture and change see significantly higher adoption rates.

The technology, in other words, is outpacing the human infrastructure around it.

This pattern plays out in predictable ways across procurement and insurance functions. Understanding why requires looking at the challenge from three distinct vantage points: the teams building AI tools, the professionals using them, and the leaders responsible for making adoption stick.

The Builder’s View: Iteration Is the Point, But Users Don’t See the Roadmap

From a product development standpoint, an AI tool that automates 50% of what used to be done manually represents genuine progress. The roadmap exists. The improvements are planned. Each release will be better than the last.

The problem is that end users do not experience a roadmap, but today’s version of the tool. And if today’s version is slow, inaccurate or clunky, the conclusion most users draw is not “this will get better”, but “I cannot rely on this”.

First impressions in AI adoption are extraordinarily difficult to recover from. In contract management, spend analysis, claims processing or any other data-intensive function, a tool that returns inaccurate results on day one, even at 55% accuracy, can be written off permanently, regardless of how much it improves in subsequent releases. The agile development model assumes iterative improvement. End users, operating under real workload pressure, do not share that frame of reference.

Data quality compounds the problem. AI cannot produce reliable outputs from unreliable inputs. Across procurement and insurance, this is a recurring issue: poorly structured contract repositories, inconsistent spend data, legacy claims records that were never designed to be machine-readable. The ceiling of what any AI tool can achieve is defined by the quality of what goes into it, and that ceiling is often lower than anyone wants to acknowledge at the outset.

The End User’s View: “I Was Working for the AI, Rather Than AI Working For Me”

In our experience working across procurement and insurance functions, this sentiment, articulated by a consultant using an AI contract management tool, captures something that statistics alone do not.

The promise of AI in any operational context is that it reduces burden. It handles the repetitive, time-consuming parts of a task so that human judgement can be applied where it matters most. When the opposite happens, i.e. when a user finds themselves correcting outputs, re-prompting, compensating for gaps or maintaining a parallel manual process just to verify results, the value proposition collapses entirely.

Udacity research bears this out at scale: three out of four workers frequently abandon AI tools mid-task, citing concerns about accuracy, time spent refining outputs and poor workflow fit. Adoption figures look impressive. Trust figures tell a different story. Stack Overflow’s 2025 developer survey found that only 29% of users trust AI outputs to be accurate; down from 40% the previous year, even as usage continued to climb.

This divergence matters enormously in procurement and insurance, where the stakes of inaccuracy are high. A misclassified contract clause, an incorrectly extracted data field, a supplier risk score generated from incomplete data. These are not minor inconveniences. They are the kind of errors that erode professional confidence in a tool quickly and permanently.

The instinct to demand 100% accuracy before trusting a tool is understandable, even if it sets an unrealistic bar. What is often missing is a clear articulation of what the tool is capable of, under what conditions and what the user’s role in that partnership actually looks like.

The Management View: Bias Beware – “Just Because It Has AI Doesn’t Mean It Works”

There is a particular failure mode in AI investment that experienced procurement and technology leaders recognise immediately: the assumption that the label does the work. An AI-powered solution sounds compelling in a procurement brief or an insurance technology roadmap. The promise is real. But unless a tool has been stress-tested against actual workflows, actual data volumes and actual use cases, the gap between what is written on the slide and what happens in practice can be enormous.

The organisations that navigate this well share a few consistent characteristics.

They test before they commit. Whether buying off the shelf or building internally, they define their specific use cases in advance and evaluate tools against them, not against demonstration environments or curated datasets. In insurance, this means running claims extraction tools against the full range of document types and complexities the team actually encounters, not just clean examples. In procurement, it means testing spend categorisation logic against real, messy transactional data.

They manage expectations deliberately. Deloitte’s 2025 CPO research found that procurement digital leaders, i.e. organisations that pair advanced technology with skilled talent, achieve 3.2x ROI on their GenAI investments, compared to 1.5x for followers. The difference is not the tool. It is the rigour of implementation and the quality of the human framework around it. Setting realistic expectations about what an AI tool can do at launch, communicating clearly about the improvement trajectory and being honest about the validation that will still be required, are management responsibilities, not afterthoughts.

They treat iteration as a process, not a failure. The organisations getting the most from AI in both procurement and insurance are those that have built systematic feedback loops: structured testing, documented limitations, clear channels for end user input, and a discipline around improving the next release based on real-world experience rather than theoretical capability.

Why Integration Is Where Most AI Initiatives Break

Even well-managed AI rollouts frequently stall at a more structural level. MIT research found that 95% of enterprise AI pilots fail to deliver measurable impact. Not because of model capability, but because of integration, data and governance gaps. McKinsey’s 2025 data puts it bluntly: only 1% of business leaders say AI is fully integrated into workflows across their enterprise.

The pattern is consistent. Procurement and insurance environments are often fragmented, built on legacy systems that were never designed to work together. When AI outputs sit outside of existing workflows, requiring users to move between platforms, re-enter data, or manually reconcile results, adoption stalls. The issue is not what the technology can do. It is whether the technology connects to where work actually happens.

Successful integration is a design decision, not a technical layer added after deployment. The organisations that move beyond pilots are those that anchor AI directly inside the systems where work already occurs: ERP platforms, contract lifecycle management tools, claims systems; so that users interact with AI as part of a step they already take, not as an additional one. In practice, that typically means identifying where manual effort currently sits in a process and replacing that step, rather than adding another.

The second design question is equally important: defining when human input is actually required. Rather than asking users to validate every output, which recreates the manual burden AI was meant to reduce, high-performing organisations set clear thresholds based on confidence, risk or complexity. AI handles routine, high-volume tasks autonomously. Exceptions and judgement-based decisions escalate to human review. This is what the “human-in-the-loop” model looks like in practice, and research from the insurance sector consistently shows it as a critical factor in sustaining adoption beyond the pilot phase.

The Emerging Capability Gap: Prompting as a Core Professional Skill

Underlying all of this is a shift in what it means to be a skilled procurement or insurance professional in an AI-enabled environment. The mid-stage work: searching, sorting, collating, extracting, is increasingly what AI handles. What it cannot replace is the ability to define a problem precisely and direct an AI system toward a useful output.

Prompting is, in essence, structured thinking applied to a new medium. BCG found that 89% of executives believe their workforce needs improved AI skills to realise the technology’s potential. That gap is particularly acute in procurement and insurance, where domain expertise is deep, but AI literacy is still developing.

Organisations that invest in building this capability, not just deploying tools, but equipping their people to use them well are the ones consistently pulling ahead. Deloitte’s data shows that procurement digital leaders hit or beat cost savings targets 96% of the time, compared to 80% for others. The gap is not technology, but the people, knowledge and implementation discipline.

What Good AI Adoption Actually Looks Like

Across the procurement and insurance work we do at Procurato, the organisations that succeed with AI share a clear set of characteristics.

1) They start with the problem, not the tool

Deploying AI because it sounds strategically credible, or because a competitor has, is a reliable path to wasted investment. The starting point needs to be a specific, well-defined operational problem. The question is whether AI is genuinely the right solution to that problem.

2) They build the human infrastructure first

Change management is not a communications exercise. It is the deliberate work of preparing teams to work differently, setting realistic expectations, and creating the conditions in which people feel equipped rather than threatened. Only a third of organisations currently prioritise this. The ones that do perform measurably better.

3) They calibrate accuracy expectations

The question is not whether an AI tool is perfect, but whether it is accurate enough that the time saved on generation outweighs the time spent on validation. Where that threshold sits varies by function and context; it is different for routine contract extraction than it is for supplier risk classification. Getting this right requires honest assessment, not vendor promises.

4) They invest in prompting capability

Across procurement and insurance, the ability to direct AI toward the right output is becoming as important as the ability to interpret the output itself. This is a skill that can be developed, and organisations that treat it as a training priority are already seeing the difference.

The pilot-to-production gap is real, and it is wide. But it is not inevitable. The organisations closing it are not necessarily the ones with the most sophisticated technology, but the ones taking the human side of AI implementation as seriously as the technical side, and recognising that the two cannot be separated.

Related reading

The adoption challenges described here play out differently depending on the context. For procurement leaders, what CPOs need to know about bridging the AI gap explores the specific use cases where AI is and is not yet delivering in practice, and what a deliberate adoption strategy looks like. For insurance leaders, navigating the specific challenges of AI in insurance maps the regulatory, legacy technology and data quality barriers that shape what is realistically achievable, and when. Both are useful companions to the cross-sector framework set out here.

How Procurato can help

The four principles outlined here: starting with a well-defined problem, building the human infrastructure first, calibrating accuracy expectations honestly and developing prompting capability as a core skill, describe a structured approach to AI investment that most organisations reach only after at least one costly failure. Getting there before that failure requires the same discipline applied to AI tool selection and implementation that good procurement brings to any major commercial decision: rigorous use-case definition, independent vendor evaluation against real data, and a change programme that runs in parallel with the technical deployment, not after it. Procurato’s Technology Sourcing & Procurement Transformation service provides that structured framework, helping organisations in both procurement and insurance select, evaluate and implement technology with the rigour that separates the 4% who scale successfully from the 74% who do not.

Stay ahead of the game with our latest insights

References

Art of Procurement. (2025). State of AI in procurement. https://artofprocurement.com/blog/state-of-ai-in-procurement

BCG. (2024, October 24). AI adoption in 2024: 74% of companies struggle to achieve and scale value. https://www.bcg.com/press/24october2024-ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value

ComplexDiscovery. (2025). Why 95% of corporate AI projects fail: Lessons from MIT’s 2025 study. https://complexdiscovery.com/why-95-of-corporate-ai-projects-fail-lessons-from-mits-2025-study/

Datagrid. (n.d.). AI agent for insurance statistics. https://datagrid.com/blog/ai-agent-for-insurance-statistics

Deloitte. (2025). 2025 global chief procurement officer survey. https://www.deloitte.com/us/en/services/consulting/articles/2025-global-chief-procurement-officer-survey.html

Konnecthouse. (2025). Deloitte’s 2025 CPO survey: Procurement at the tipping point of AI transformation. https://www.konnecthouse.com/deloittes-2025-cpo-survey-procurement-at-the-tipping-point-of-ai-transformation/

Libertify. (2025). State of AI 2025: McKinsey report [Interactive library]. https://www.libertify.com/interactive-library/state-of-ai-2025-mckinsey-report/

Nexos.ai. (n.d.). AI adoption challenges [citing BCG research]. https://nexos.ai/blog/ai-adoption-challenges/

Risk and Insurance. (n.d.). Insurance leads AI revolution but struggles to scale beyond pilots [citing BCG research]. https://riskandinsurance.com/insurance-leads-ai-revolution-but-struggles-to-scale-beyond-pilots/

Roots.ai. (2025). State of AI adoption in insurance 2025 [White paper]. https://www.roots.ai/hubfs/Reports%20and%20Whitepapers/Roots%20-%20State%20of%20AI%20Adoption%20in%20Insurance%202025.pdf

ServicePath. (2025, September). AI integration crisis: Enterprise hybrid AI [citing MIT research]. https://servicepath.co/2025/09/ai-integration-crisis-enterprise-hybrid-ai/

Stack AI. (n.d.). The biggest AI adoption challenges. https://www.stack-ai.com/blog/the-biggest-ai-adoption-challenges

Stack Overflow. (2026, April 2). What the AI trust gap means for enterprise SaaS. https://stackoverflow.blog/2026/04/02/what-the-ai-trust-gap-means-for-enterprise-saas/

Udacity. (2025, September). AI adoption trust gap: Udacity research. https://www.udacity.com/blog/2025/09/ai-adoption-trust-gap-udacity-research.html