Enterprise AI ROI Clusters in Five Narrow Operations Workflows

Narrow AI deployments are paying off

After two years of broad experiments, a clearer pattern in enterprise AI is starting to show. The deployments with measurable returns tend to be narrow, high-volume and checkable. Assistants that are open to everything and owned by no one tend to stall. Coverage across vendors, consultancies and academic researchers keeps pointing to the same short list of operational workflows: customer service, finance back-office work, software engineering, HR helpdesks and supply chain logistics.

That list comes with a caveat that is hard to ignore. The most widely cited figure on enterprise generative AI is still MIT Project NANDA's finding that about 95% of pilots produced no measurable profit-and-loss impact.2427 Both findings describe the market accurately. The useful question for buyers is no longer whether AI can pay back. It is which conditions separate the minority of projects that do from the majority that don't.

Use case one: customer service resolution

Customer operations is consistently described as the fastest payback in enterprise AI. One analysis argues that companies often reach positive returns in under six months. The reasons it gives are high ticket volume, answers that can be checked against a knowledge base, and an obvious human fallback for any case the system can't close.3 One compilation of Bain's 2026 benchmarks puts median payback for customer-service agents at 4.1 months, the shortest of any category it tracked.4

Klarna is the flagship case. Its agent is reported to do work equivalent to 853 full-time staff.1 On pricing, Intercom's Fin is reported to charge $0.99 per autonomous resolution. A per-ticket price like that gives CFOs a direct way to compare against what a human agent costs.4 Adoption data fits the story: one 2026 tracker found that 56% of enterprises use AI for customer support, more than any other use case.8

Use case two: finance and the back office

Finance is the second area where the reporting broadly agrees. Invoice processing, three-way matching, reconciliation and expense compliance suit automation because the data is structured, the rules are explicit and every action leaves an audit trail the system can be graded against.3 One analysis estimates that close cycles shrink by 20–35% where financial data is clean. It also notes that finance projects cost more and take longer to pay off, because the data needs heavy preparation and there is no tolerance for errors.2

This matters because the MIT research behind the 95% figure found that spending skews the wrong way. Successful deployments tended to target one narrow back-office pain point. Stalled ones often put 50–70% of their budgets into high-visibility front-office pilots in sales and marketing.27 Other commentary on the same findings says back-office work showed higher ROI even though most budgets went to sales and marketing.26 In short, the less glamorous projects appear to be the ones that pay.

Use case three: software engineering, where the data disagrees most

Coding assistants are the most widely deployed enterprise generative AI tool10, and the evidence on them is the most contested. The well-known GitHub-backed controlled trial found developers finished a narrow task 55.8% faster with Copilot.16 MIT Sloan coverage of three field experiments reported a 26% average increase in completed weekly tasks. Junior developers gained 27% to 39%, compared with 8% to 13% for senior engineers.14

Results from inside real organizations are more sobering. An Okta field study published in Communications of the ACM found that engineers' weekly hours fell from 44.77 to 40.83 and that motivation rose. It found no statistically significant increase in monthly pull requests or lines of code, even in follow-up monitoring through May 2025.11 The authors call this the "efficiency-throughput gap": individuals do the same work in less time, but the organization doesn't produce more.11 An Uplevel study of nearly 800 developers similarly found no meaningful change in PR cycle time and a 41% increase in bugs.16

That clashes with marketing-style claims of a 42% improvement in code quality among Copilot users.19 The more credible reading favors the field studies. Copilot makes individual coding faster, but the time saved goes to engineers' lives and calendars unless the company redesigns its processes, reviews and infrastructure to use it.11 Forrester figures cited in one analysis put an agent-run pull-request review at about $0.72, against $48 of senior-engineer time.9 Savings like that per task only reach the bottom line if the review queue was actually slowing delivery.

Use case four: HR helpdesks

HR self-service is quieter but well documented. AMD's agentic HR system, built on Kore AI for a workforce of about 30,000, reportedly cut HR resolution time by 80%, handled half of queries through self-service, and raised employee satisfaction scores by 70%.1 Broader benchmarks suggest policy and handbook question-answering agents can deflect 60–80% of helpdesk tickets.10 Forrester's commissioned Copilot analysis cited a 25% reduction in onboarding time.18 The use case fits the same profile as customer service: repetitive questions, a defined body of policy, and escalation to people for sensitive cases.

Use case five: supply chain decisions

The fifth use case is the most directly tied to financial results. General Mills' CFO told investors that AI models evaluate more than 5,000 daily shipments from plants to warehouses, producing more than $20 million in savings since fiscal 2024 through lower transport costs and better service levels.1 Demand forecasting and predictive maintenance show up repeatedly in other ROI lists. Their success depends on connected ERP and warehouse data and on monitoring for model drift.6 Manufacturing payback runs longer, typically 12 to 18 months in one benchmark, and McKinsey's work with Jubilant Ingrevia tied 15–25% EBITDA gains to pairing digital changes with operational and workforce changes.5

The copilot question: adoption is not productivity

General-purpose office copilots sit across all five areas, and the reporting on them deserves careful reading. Microsoft's own figures are mostly self-reported: 70% of early users said Copilot made them more productive.18 Measured time savings vary widely. Microsoft's early-user study found 14 minutes a day, while a UK government trial reported 26 minutes.18 A later HMRC evaluation, which randomly allocated 3,000 licenses, estimated about 60 minutes saved per employee per week. That figure already includes a roughly 20% discount for inactive users and response bias.15

EY, one of the largest deployments, which is rolling Copilot out to more than 400,000 people, shows the same gap. Microsoft and EY report a 15% productivity gain but have published little of the method behind it. EY Canada reports 10%.15 One analyst notes that EY's strongest finance, tax and assurance results depended on custom automation and connected systems, not standard licenses.15 Another compilation found general-purpose assistants averaging roughly 90% ROI, compared with 340% for targeted fraud detection.7 The overall lesson is that the more specific the problem, the bigger the return.

Why most pilots still fail

Coverage of the 95% figure largely agrees on the causes, though it is less united on the number itself. MIT's researchers blamed weak integration into workflows and data, not model quality.24 Critics note that the report was preliminary, relied on self-reported data, and used a measurement window of about six months, so the figure is better read as a direction than a precise rate.27 Other studies give a range rather than one number. McKinsey found 39% of organizations reporting any enterprise-wide EBIT impact. IBM's CEO study found only 25% of AI initiatives delivered expected ROI. S&P Global recorded abandonment before production rising from 17% to 42%.2529

On the other side, a Wharton survey found nearly three-quarters of leaders reporting positive ROI.25 That suggests the same companies can count time savings as wins while seeing no change in the P&L, a contradiction practitioners have pointed out themselves.30 Spending is outrunning results. Forrester reportedly found enterprises pushing a quarter of planned AI spending into 2027.22 Accenture Federal Services' CEO has argued that every dollar of AI technology needs about nine dollars of change management, training and workflow redesign.5

The verdict

Taken together, the evidence supports a firm conclusion. Enterprise AI returns are real, but they are concentrated, and they go to companies that treat AI as process redesign rather than buying licenses. The five use cases that keep producing numbers share the same traits: high volume, structured inputs, outputs that can be verified, a human fallback, and an owner accountable for a metric. MIT's data reportedly shows that purchased vendor solutions succeed about twice as often as internal builds.27 Bain's benchmark found vendor-deployed agents reaching positive ROI 2.4 times faster than custom builds.9

The practical advice for companies running AI programs follows from that. Pick one workflow with a cost that can actually be cut, set a baseline before launch, and measure results at the level of the P&L rather than per person. Companies that skip those steps are likely to end up with faster employees and no change in the business, a pattern the field studies have already documented.1121