Generative AI in Operations: Where It Actually Earns Its Keep Today

Generative AI has been the subject of unusually strong claims about what it will replace and how quickly. Some of those claims are turning out to be accurate. Others are not. For operations leaders trying to decide where to invest attention, the useful distinction is not what AI can do in theory but what it is measurably doing in real deployments today.

This piece is a candid tour of what generative AI has been earning its keep on in operations — where the deployments are stable, the ROI is real, and the failure modes are understood. It is also an honest account of where the technology is not yet ready to be relied on, so you can plan accordingly.

The clearest and most durable value is in operations dominated by unstructured text. Reading long documents and returning structured summaries. Extracting fields from invoices, contracts, or forms. Generating responses to routine correspondence. Categorising incoming messages so they route to the right queue. Rewriting internal documentation into public-facing copy or the reverse.

In these workflows, modern models routinely produce quality that matches or exceeds trained humans on speed and often matches them on accuracy for the routine cases. The value comes not from replacing the humans on the difficult cases but from removing the routine work that used to consume most of their attention. Operations teams reallocate that attention to the exceptions, and total throughput rises sharply.

The second domain of clear value is conversational work — service, qualification, coordination, follow-up. This is the domain where generative AI most visibly changes the customer experience. Response times collapse. Consistency of tone improves. Escalations become faster because the context is already compiled.

The workflows that benefit most are those where the conversation follows a recognisable shape most of the time and needs judgment only occasionally. Customer service on a well-documented product. Lead qualification against a defined ideal customer profile. Appointment coordination across a calendar and a set of policies. The AI handles the shape, the humans handle the edges.

Related to but distinct from text summarisation is structured extraction: taking unstructured input and producing validated structured output. Reading a photographed receipt and turning it into a line-item expense claim with a category, amount, tax code, and cost centre. Reading a scanned contract and turning it into a set of dates, parties, obligations, and risks. Reading an inbound email and turning it into a case with a priority, category, and next action.

This work used to require rule-based tools that were fragile and expensive to maintain. Modern models with proper prompting and validation are far more robust, and they generalise to new document types without a rewrite. This is one of the highest-ROI applications of generative AI in mid-market operations today.

Numerical operations — forecasting, optimisation, precise calculation — are still better served by classical techniques. Generative AI can describe a number well; it is less reliable at producing one from complex inputs when precision matters. Where a workflow requires exact numerical output, generative AI should be treated as a helper that surfaces context, not as the calculator.

Long-running autonomous work with high stakes is another area where the technology is meaningfully more capable than it was but not yet ready for hands-off deployment. Agents that plan and execute multi-day projects with real consequences are working in narrow domains today, but the general case still needs meaningful human supervision.

And highly creative work — genuinely original strategy, novel product design, high-craft writing — remains a domain where AI is a productive collaborator rather than a replacement. The output improves with a human in the loop; it does not yet stand alone at the highest levels.

Across every one of these domains, the pattern that separates the operations teams getting outsized value from those getting incremental value is workflow design, not model choice. High performers pick a workflow, redesign it around what AI is good at, and rebuild the surrounding operational rhythm to match. Low performers try to bolt AI onto an unchanged workflow and are disappointed that the improvement is marginal.

The lesson for operations leaders is that generative AI is a workflow redesign tool disguised as a technology purchase. If you buy the technology without the redesign, you get a fraction of the value. If you do the redesign with the technology as an enabler, you get compounding value across every subsequent workflow.

The starting point that has proven most robust across our engagements is inbound triage — of messages, of documents, of cases. It is high volume, it has a clear success metric, it is easy to instrument, and the improvement is immediately visible to the operations team.

From triage, the natural next step is response drafting: the AI drafts the reply, the human reviews and sends. This teaches the team what the AI is good at and where it needs help, and it builds the shared judgement that later supports higher-autonomy workflows.

By the third or fourth workflow, the team is comfortable enough to move toward end-to-end automation for the well-understood cases. This is where operational metrics start to change materially. Get the first two workflows right, and everything downstream gets easier.