Where AI actually creates value (and where it doesn't)

Most companies can now say they use AI. Far fewer can point to the number it moved. The teams getting real value did not have the boldest strategy; they picked one unglamorous, repetitive workflow and let the technology do it well. The difference between AI that pays for itself and AI that becomes shelfware is almost never the model. It is the problem you point it at.

Adoption is not value

McKinsey's 2025 research found that most organisations report no tangible impact on enterprise-level profit from generative AI, and only about a third have scaled it at all.¹ The reason is rarely the tooling. "Adopt AI" is a mandate, not a problem statement, and mandates produce pilots and demos, not outcomes. The work looks like progress and changes nothing, because it was never aimed at a defined target.

What the workflows that pay off have in common

When AI moves a real number, the task underneath it is usually bounded (you can define what "done correctly" means), high-volume (it happens thousands of times, so small savings compound), measurable (a metric existed before the AI arrived), and owned (someone is accountable for running it after handover). None of that is about the model. It is about the operation.

Where it works

Our specialist operators have seen the same shape across very different settings.

• Public-sector document processing. A secure classification and extraction pipeline cut processing time by 78%, dropped the data-extraction error rate below 2%, and has held a full audit trail with zero compliance issues since deployment.
• Enterprise IT service management. An AI triage and routing layer lifted SLA compliance from 71% to 94%, cut average ticket resolution time by 52%, and saved each operations manager around six hours a week of reporting.
• B2B revenue operations. An automated CRM-audit layer took report delivery from three days to 45 minutes and doubled the engagements a team could run at once, from four to eight a month.

Three settings, one shape: a narrow, repetitive, measurable job with a clear owner. Where the problem is undefined, has no owner, or had no baseline to measure against, the same technology quietly becomes slideware.

Scope the problem before you pick the model

So before funding another pilot, decide the specific task, who will own it, and how you will know it worked. The model question usually answers itself after that. It is the sequence we apply to any system: find the workflow where AI will actually pay off, not the one that demos well, and build it to be owned by your team after we leave. Our specialist operators work hands-on across the tools that matter, from Clay and n8n to leading models such as Claude and GPT-4o, and have shipped enterprise-grade AI in regulated environments. You can read more on our AI Implementation page.

If you are under pressure to have an AI story, the most credible one is not that you adopted AI. It is that you found the one workflow where it changed a number, proved it, and handed it over.