AI in operations: where it really pays, with real numbers

Where does AI already pay in operations?

In 2026 AI pays reliably in 3 places: high-volume customer support, software development and repetitive back-office work such as document processing and reconciliation, and it pays much less reliably in sales, where the bottleneck is trust rather than throughput.

The macro picture is sobering. According to McKinsey's State of AI 2026 survey of 1,719 respondents in 97 countries, 80% say AI has improved their individual productivity, but only 37% report that it contributed anything to their organisation's earnings before interest and taxes (EBIT), essentially unchanged from a year earlier, and only about 6% qualify as high performers with at least 5% of EBIT attributable to AI. Individuals are faster; companies mostly are not.

My reading of that gap is simple. Time saved by a person evaporates unless you change what the process expects of them: fewer people on the queue, a shorter service level, a smaller vendor bill. AI pays where the workflow is redesigned around it, not just assisted by it.

What does AI actually deliver in customer support?

A well-fed AI agent on a mature knowledge base resolves 30% to 60% of inbound conversations without a human, and the honest number for a company starting from scratch is the low end of that range.

The reference case is Klarna. According to its February 2024 announcement, the assistant handled 2.3 million conversations in its first month, two thirds of all support chats, cut average resolution time from 11 minutes to under 2, reduced repeat inquiries by 25% and did the work of about 700 full-time agents, with an estimated $40 million profit improvement for 2024. The second half of the story matters as much: in May 2025 Klarna's CEO told Bloomberg that cost had become "a too predominant evaluation factor" and that the result was lower quality, and the company started hiring humans back for complex cases.

Pricing has moved to outcomes, which helps the business case. Intercom's Fin charges $0.99 per resolution, where a resolution is counted when the customer confirms the answer helped or leaves without asking for more. In my companies we compare that against a fully loaded human cost per ticket of roughly €4 to €12 depending on the country and the product, so the arithmetic works if, and only if, the AI resolution is a real resolution. Track customer satisfaction and reopen rate on AI-handled tickets separately, or you will be paying for silence.

Does AI pay in sales and marketing?

AI pays in sales operations, not in selling: research, data entry, call summaries, proposal drafts and lead routing shrink by hours per rep per week, while conversion rates barely move when the AI talks to the buyer directly.

In my companies the measurable gains are in the boring layer. Account research that took 20 minutes before a call takes 2. Meeting notes and CRM updates that reps skipped now happen automatically, which improves forecast quality more than any dashboard did. Outbound copy gets drafted in the buyer's language for 20+ countries without a translation queue.

Where I stay careful is automated outreach and AI sales agents. Volume is cheap now, so buyers are flooded and reply rates fall for everyone. The advantage goes to whoever uses AI to be more relevant, not more frequent. I measure sales AI on hours returned to selling and on pipeline data completeness, not on revenue, for the first 2 quarters.

Does AI pay in finance and the back office?

Yes, and quietly: invoice matching, expense classification, contract review and month-end reconciliation are structured, repetitive and verifiable, which is exactly where language models are strongest and safest.

This is the function most boards underestimate because it has no demo appeal. In my companies the finance team uses AI to pre-classify transactions, extract fields from supplier invoices, draft management commentary and flag anomalies before the accountant looks. The human still signs. The gain is a shorter close and fewer errors rather than fewer people.

Two rules keep it safe. First, the AI proposes and a person approves anything that touches money, tax or a contract. Second, every AI-produced number must be traceable to its source document in 1 click, because an auditor will ask.

How much faster does AI make software teams?

Between much faster and slightly slower, depending on the task: greenfield, well-specified work speeds up dramatically, while changes to large, mature codebases by experienced engineers can take longer once you count review.

The 2 most cited controlled studies point in opposite directions and both are right. In the GitHub Copilot experiment published on arXiv in 2023, developers implementing an HTTP server in JavaScript finished 55.8% faster with the assistant. In METR's randomised trial published in July 2025, 16 experienced open-source developers working on 246 real issues in mature repositories took 19% longer when allowed to use AI tools, mostly Cursor with Claude 3.5 and 3.7 Sonnet. They had expected a 24% speedup, and even afterwards believed they had been 20% faster. Perception is not a metric.

In my companies the 2026 pattern with agentic tools such as Claude Code is clear: prototypes, tests, migrations, documentation and internal tools move several times faster; core product changes move maybe 20% to 30% faster once review is included; and review has become the bottleneck, so we invested in automated review, stricter continuous integration and smaller pull requests.

How do I measure whether AI pays?

Measure AI in operations with 1 unit-cost metric, 1 speed metric and 1 quality metric per process, compared against a baseline you froze before the rollout.

This is the scorecard we use across my companies. It fits on 1 page and it is reviewed monthly by the person who owns the process, not by the AI team.

• Cost per resolution or per document processed, fully loaded, before and after, including the AI vendor bill.
• Resolution or deflection rate paired with reopen rate and customer satisfaction on AI-handled cases only.
• Cycle time: first response, time to close, days to month-end close, lead time from pull request to production.
• Quality: error rate found downstream, escalations, defects per release, audit findings.
• Human hours returned, and what those hours were actually reallocated to.
• Adoption: share of the team using the tool weekly, because a licence nobody opens is pure cost.

What fails most often?

AI projects in operations fail for 5 reasons, and none of them is the model: bad knowledge bases, no process redesign, no owner, vanity metrics and skipping the human fallback.

I have made most of these mistakes myself. The pattern is always the same: the pilot works because the founder babysits it, and production fails because nobody owns the knowledge base 6 months later.

• The knowledge base is stale or contradictory, so the agent answers confidently and wrongly; fix the content before the model.
• The process is untouched, so the time saved is absorbed and nothing changes in cost or service level.
• Nobody owns it after launch; every AI system needs a named operator who reviews failures weekly.
• Deflection is celebrated while satisfaction and reopens are ignored, which is how Klarna ended up rehiring.
• No clean path to a human, which turns a cost saving into churn.