The AI pilot launched with high expectations is still in the sandbox. The AI initiative is behind schedule. And yet, elsewhere across the industry, organisations are already using agents to automate activities such as order exception handling and customer enquiry routing.
The gap between stalled and scaling is not luck. It is a pattern.
Across the AI assessments, deployment work, data integration programmes and adoption challenges we have supported over the past year, several trends have become difficult to ignore. These are not predictions. They are the shape of what is working now.
Agents are moving from prove-the-point to solve-the-problem
Pilot fatigue is real. Many organisations built their first agents to test the water: a chatbot here, a summarisation flow there.
What is changing is the scope. We are seeing serious agent deployments move towards specific, measurable use cases—pipeline forecasting, order exception handling and field service scheduling.
The difference is simple: a pilot validates that AI can work. A production agent solves a problem that would otherwise require additional headcount or continued manual effort.
This shift also changes what matters when selecting technology.
Model capability matters, but it is no longer the only—or necessarily the primary—selection criterion. The more important question is which combination of model, platform, data and integration can deliver the outcome securely, within a realistic timeframe and with adoption that can be measured.
Data readiness is the real constraint—not model capability
The engineering team always knew this. It turns out boards are learning it too.
You cannot run an effective agent on incomplete, siloed or stale data. Most organisations have all three.
The teams getting value fastest are those that bundle agent deployment with data consolidation and integration—often through MuleSoft, Data 360 or similar technologies.
It is not the most exciting part of the AI story. But it is the difference between "we built an agent" and "we deployed an agent that works."
Data readiness is becoming the gating factor in agent time-to-value. Organisations are increasingly budgeting for it upfront rather than treating it as a downstream technical issue.
The ROI obsession is reshaping how AI gets funded
The days of the loosely defined "AI transformation initiative" with a vague remit and a three-year budget are disappearing.
The CFO is asking for numbers: what problem are we solving, how much does it cost today and when will we see payback?
This is healthy. It reduces speculative deployments and forces organisations to connect AI investment to a meaningful business outcome.
It also means the vendors and consultancies able to connect an agent to measurable value—pipeline velocity, cost per service interaction, fulfilment speed or employee capacity—are more likely to secure investment.
The organisations winning support are those leading with outcomes rather than capability alone.
Outcome definition
The business case now starts with "we will reduce X by Y% in Z months."
Budget accountability
AI is competing for capital against other initiatives. The ROI bar is higher.
Timeline realism
Organisations are learning that data and adoption take time. Fast pilots are common; fast deployment at scale is rare.
Platform coherence is beating fragmented architectures
After a decade of composable, best-of-breed thinking, platform coherence is becoming increasingly important for enterprise AI.
Agents depend on connected data, identity, workflows, integrations, security and governance. Building those foundations across several disconnected vendors is possible, but it is typically slower, more expensive and more difficult to govern.
Where organisations already operate on Salesforce, Agentforce, Data 360 and MuleSoft can provide much of this foundation within a connected ecosystem. Models such as Claude can then provide a powerful intelligence and reasoning layer within the wider solution.
This does not eliminate best-of-breed thinking. It relocates it.
Organisations still need to select the right technology for the job, but the strongest architectures connect those technologies through a coherent operating platform rather than creating another fragmented layer.
The strongest deployments are therefore not treating the model as the entire solution. They are combining capable models such as Claude with enterprise platforms such as Salesforce, alongside the data, integration, security and governance required to operate reliably.
Adoption velocity matters more than feature velocity
The teams creating sustained value are obsessive about adoption, not features.
The question is not, "How many agents can we build?" It is, "How quickly can our teams trust and use the agents we have?"
That requires investment in change, monitoring and feedback loops. It means understanding where workflows will change, where human judgement remains essential and how performance will be measured after launch.
It also means smaller teams doing more—not larger teams building more.
Pick one that solves a real problem and can measure success within 90 days.
Consolidate, clean and integrate what the agent needs to work.
Move beyond the sandbox and put the agent into a real team and workflow. Watch adoption closely.
Measure adoption and outcomes, remove friction, then expand.
What this means if your pilots are stalling
The move from AI as experiment to AI as operation is real. It simply looks different from what many initially expected.
Smaller bets. Tighter integration. Greater discipline around adoption and measurable outcomes.
The debate should not be reduced to model versus platform. Models such as Claude provide increasingly capable intelligence and reasoning. Platforms such as Salesforce provide the customer context, workflow, identity, integration and governance required to put that intelligence to work. The opportunity lies in designing the right combination around a clearly defined business outcome.
If your pilots are stalling, it is rarely because the underlying AI is not capable enough.
It is usually one of three things: the data was not ready, the team was not ready or the business case was not clear.
Fix those, and the path from pilot to production becomes considerably clearer.
Ready to move pilots into production?
We help organisations move from AI experimentation to measurable operational value by addressing the business case, data foundations, technology architecture and adoption together. If your pilots are stalling, let's identify what is holding them back and define a practical path forward.
