Half of enterprise AI projects hit their goals. One in five gets cancelled.

About 50% of enterprise AI initiatives achieve some or all of their business objectives, and one in five delivers no measurable value at all — frequently getting cancelled after deployment. Those numbers come from the AI Business Value Radar, published by the Infosys Knowledge Institute, and they frame an uncomfortable reality for corporate AI programs: the pilot is the easy part.

Sprints are useful. Quick pilots, rapid demos and early wins build the initial confidence that enterprise AI adoption needs. But for many organisations, the effort stops right there, according to the analysis. Companies build momentum, then discover they have no clear path to turn a demo into a lasting advantage. Projects greeted with applause at launch often end up quietly shelved.

The stakes are straightforward. Enterprises have poured budget into AI over the past several years, and the difference between the winners and the also-rans is increasingly about operational endurance — whether a company can compound value year after year — rather than who launches pilots fastest.

Why do so many AI projects stall after the pilot?

The Infosys analysis points to a set of concrete operational hurdles that sit between a successful demo and sustained value:

• Scattered tools and software overload
• Legacy data silos
• Weak governance
• Severe skills shortages

One example illustrates the problem vividly. A mid-sized Australian retailer operates more than 300 software applications across three different eCommerce platforms. Because these systems don't talk to each other, even basic initiatives — a customer loyalty program, for instance — get stuck before they start.

The takeaway: fragmentation, not model quality, is what kills many enterprise AI programs.

What foundational capabilities actually matter?

Sustained AI value rarely comes from a single tool. The analysis argues it requires connecting four things: vision, governance, integration and people.

Strategy comes first. A clear strategy ensures the organisation treats AI as a core business driver, not a routine tech upgrade. Governance must be established early too — clear risk boundaries and ethical rules set from the start, long before the first pilot goes live. As operations grow, those controls need to be built directly into the software tools themselves.

Then there is data readiness, which the analysis calls the true separator between successful deployments and failed ones. AI models are only as effective as the data they can access. A leading Australian retailer shows how this works in practice: online shoppers see up-to-the-minute stock levels and exact aisle locations because the digital store draws straight from live inventory feeds.

On physical sites, the retailer blends visual cameras, IoT sensors and Light Detection and Ranging (LiDAR) equipment to build a far clearer operational picture. The Australian technology firm SenSen.ai applies this same multi-sensor approach to manage urban traffic and public safety — evidence, the analysis suggests, that combined data streams yield far better outcomes than isolated inputs.

The principle in plain terms: live, connected data beats a perfect model fed by broken pipelines.

Where do agentic workflows fit in?

The analysis also pushes organisations to embed AI directly into core business processes through agentic workflows. Instead of operating as standalone tools, intelligent agents handle multi-step actions across operational boundaries — predicting equipment failure in manufacturing, for example, and automatically reconfiguring settings in response.

But architecture means little without people who can run it. The analysis prescribes a layered AI learning path:

• Basic AI literacy for everyone
• Role-based skills for specific functions
• Hands-on agent building for those who will design and run the workflows
• A separate layer for leaders, focused on judgement, risk and accountability rather than tools

Without that progression, the warning goes, companies end up with a few experts and a large workforce unsure how to work with agents.

The point: capability-building is a workforce project as much as a technology project.

How do organisations mature over time?

Building AI capability doesn't happen with the flip of a switch. The Infosys framework describes five distinct phases of maturity: Explore, Enable, Establish, Expand and Excel.

The early stages focus on low-risk proof-of-concept projects and defining clear roadmaps. As organisations move into the Establish and Expand phases, they centralise governance into live platforms and empower individual business units to innovate within shared guardrails.

At the final Excel stage, the organisation operates as a modular system of reusable data, model and agent assets — and leads its industry in agentic process reinvention.

In short: maturity is staged, and skipping stages is how pilots become dead ends.

How should executives translate a plan into action?

A maturity assessment without funding is just a diagnosis, the analysis argues. Rather than spreading resources thin, executives should group AI spending into 10-20 well-defined projects, with every initiative carrying three things: a single owner, a dedicated budget and clear performance metrics.

One major Australian retailer shows the model in action. It replaced isolated trials with a board-approved, fully funded roadmap. By building out the core platform first, the retailer set up enterprise-wide guardrails that now support every new digital rollout.

The lesson: consolidation with accountability beats a wide scatter of unfunded experiments.

What should happen before the next pilot?

Before approving the next pilot, the analysis advises, organisations must be honest about where their real capabilities lie. That means identifying which foundational steps are being overlooked, and confirming they hold a fully funded execution strategy and roadmap — not just a list of isolated trials.

The concluding argument is direct: organisations that win will not be the ones that launch pilots the fastest, but those that systematically build the operational endurance to compound value year after year. Keep sprinting where it helps, in other words — but run the larger race as a marathon.

For enterprise buyers, the message ahead of the next budget cycle is that governance, data plumbing and layered skills programs — the unglamorous work — are what decide whether the 50% success rate improves or the one-in-five cancellation rate does.