The 80% principle

At its core, smart automation is about recognising where AI adds clear, consistent value … and where it does not. In most integrated business operations, roughly 80% of workflows are repetitive and predictable. Order-to-cash cycles, reconciliation between sales and inventory systems, routine compliance checks – these are exactly the kinds of tasks where AI performs reliably, at scale, without fatigue. Automating them delivers the majority of the efficiency gains an organisation is looking for, and frees up the people who were doing them to focus on work that genuinely requires their judgement.

That remaining 20% is where the thinking needs to happen. Custom pricing decisions, regional regulatory nuance, exceptions that fall outside established parameters, commercial calls that carry strategic weight, these are not tasks to hand over to a model. They require context, accountability, and the kind of discretion that humans bring to complex situations. The organisations that understand this boundary, and build their integration architecture around it, are the ones that end up with operating models that actually work.

Where execution goes wrong

The failure modes in AI-powered integration are rarely technical. The systems themselves are, in most cases, capable of doing what they are asked to do. The problems tend to emerge in how the transformation is approached.

Misaligned expectations are perhaps the most common. When leadership frames an AI integration project as a route to eliminating headcount or achieving near-total automation, it sets a course for disappointment and, more damagingly, for resistance. Teams who feel threatened by a programme are unlikely to engage with it constructively. The value that could have been unlocked through their involvement gets lost.

Poor change management compounds this. Deploying AI into a business process without adequately preparing the people who interact with that process is a reliable way to undermine adoption. The technology may function perfectly; the workflow around it may not shift in the ways needed to realise the benefits.

A more effective approach

The organisations getting this right tend to share a common methodology, even if the specifics vary. They start with a single high-impact workflow – order-to-cash is a strong candidate in many businesses – and treat it as a proof of concept rather than a wholesale transformation. Before any deployment, they build what might be described as a digital twin: a simulated version of the changed workflow that allows the team to stress-test the approach, identify edge cases, and refine the design before it touches live operations.

From there, they deploy in layers. Automation comes first, handling the predictable volume. Exception-handling logic follows, routing the cases that fall outside standard parameters to the right people. Decision support (where AI surfaces relevant data and signals to inform human judgement) comes last, and only where it genuinely adds to the quality of the decision being made.

Success is then measured in terms that reflect operational reality: reduced manual effort, faster cycle times, improved data accuracy. Not the sophistication of the models deployed.

The longer view

The competitive advantage in AI-powered integration will not belong to the firms running the most advanced technology. It will belong to the firms that have embedded it most effectively; that have built operating models where AI handles what it handles well, people focus on what they do best, and the boundary between the two is understood, maintained, and continuously refined.

That kind of adaptive operating model does not happen by accident. It requires disciplined thinking about process design, a realistic view of what automation can and cannot do, and an organisational culture willing to evolve alongside the tools it adopts. The technology, in that context, becomes what it should always have been: an enabler, not an end in itself.