We Are Seeing a 3× Increase in AI-Led Automation Projects: What Enterprises Are Getting Right — and Wrong
Enterprise automation programmes have entered a new phase. The robotic process automation (RPA) wave that dominated 2019–2022 — automating structured, rule-based tasks through UI scripting — is giving way to a fundamentally more ambitious architecture: LLM-powered automation capable of handling unstructured data, making contextual judgments, and orchestrating multi-step workflows across systems that were never designed to be integrated.
The result, across the enterprise technology market, is a measurable surge in automation programme investment. The challenge is that the tooling, governance, and evaluation disciplines required for LLM-powered automation are substantially different from those that made RPA programmes successful — and many enterprises are discovering this the hard way.
The Scale of the Shift
| Metric | 2023 (RPA Era) | 2026 (LLM Automation Era) | Change |
|---|---|---|---|
| Average automation programme budget | $2.1M | $6.8M | +224% |
| Proportion using LLMs in automation | 12% | 67% | +458% |
| Document-processing automation adoption | 31% | 74% | +139% |
| Programmes reaching production within 12 months | 58% | 41% | −29% |
| Programmes achieving target ROI | 44% | 31% | −30% |
The last two rows tell the story: as automation ambition has grown, execution success rates have declined. More investment, more complexity, and more failures — driven by underestimating the operational discipline required for LLM-powered automation.
What the High Performers Are Doing Differently
Analysis of enterprise automation programmes that are consistently delivering production outcomes reveals three distinguishing practices:
Practice 1 — Executive Sponsorship with Accountability
Automation programmes that succeed have a named executive sponsor who is accountable for business outcomes — not just programme delivery. This matters because LLM-powered automation frequently requires changes to adjacent processes, policies, and operating models that cannot be driven by a technology team alone.
The sponsor's role: clear the organisational obstacles that technical teams cannot. Specifically: data access decisions, process change approvals, and the political capital to retire legacy workflows that agents are replacing.
Practice 2 — Platform Thinking Before Use-Case Proliferation
The enterprises making the fastest progress are not those that launched the most automation pilots — they are those that invested early in a shared automation platform: common data connectors, a feature store for reusable ML features, a model registry, monitoring dashboards, and a governed prompt library.
The counter-intuitive result: teams that built the platform first deployed fewer pilots in year one, but scaled three times faster in years two and three because they were not rebuilding foundational infrastructure for every new use case.
Practice 3 — Evaluation Hygiene from Day One
LLM-powered automation introduces a failure mode that RPA did not have: the system can produce outputs that are syntactically correct, internally consistent, and completely wrong. The model "sounds confident" even when it is hallucinating.
The enterprises that are catching this — before it reaches production — invest in evaluation infrastructure:
-
Ground truth datasets: Representative samples of the target task with human-verified correct outputs
-
Regression test suites: Automated evaluation of model outputs against ground truth on every deployment
-
Business KPI integration: Production monitoring that tracks whether the automation is improving the business metric it was deployed to improve
The Three Patterns That Separate Scale from Stall
Across the automation programmes we have observed in depth, three patterns predict which will scale and which will stall:
| Pattern | Scale Trajectory | Stall Trajectory |
|---|---|---|
| Executive sponsor | Named, accountable, engaged | Project sponsor only; changes frequently |
| Platform investment | Shared infrastructure built first | Rebuilt per use case |
| Evaluation discipline | Ground truth + regression from pilot | "We'll figure out evaluation in production" |
| Scope management | Ruthless — one well-defined workflow at a time | Feature creep; scope expands during pilot |
| Governance | Established before production deployment | Retrofitted after a failure |
The Honest Forecast
The 3× increase in AI-led automation projects will not automatically translate into a 3× increase in business value. The organisations that will capture disproportionate returns are those that resist the temptation to deploy fast and govern later — and instead build the infrastructure and discipline to deploy well.
The automation programmes of 2026 that are running with rigorous evaluation, platform-first architecture, and executive accountability will become the competitive moats of 2028. Those that are running without them will become cautionary tales.