AI agents: 8 concrete use cases for business departments (and why more than 40% of projects will be abandoned)

AI agents change the nature of working with AI. You no longer ask them for an answer: you give them an objective. In practice, they plan the steps, query the company’s tools, act, then report back. For a finance, procurement or customer service department, the question becomes very concrete. Which part of my process can I delegate, and with what level of control?

The movement is real, but uneven. Large organisations are accelerating, while most companies are still exploring. In addition, some of the projects launched in 2025 will be abandoned. Here is where companies stand, eight concrete use cases and a method for choosing the ones that will hold up in production.

Assistants, automations, AI agents: three tools, three uses

First, an assistant responds to a request. For example, it summarises a document, drafts an email or finds a piece of information. An automation runs a fixed, predictable sequence. A software robot that copies data from one screen to another is a typical case. Finally, an agent combines the two. It is given a goal, chooses its actions and uses tools such as email, the ERP or the CRM. Above all, it adapts to what it encounters.

The distinction matters. Gartner advises using agents when a decision is needed. Automation, by contrast, suits routine workflows, and an assistant suits simple information retrieval. Many projects fail for this very reason: they bring in an agent, which is more expensive and harder to control, where a rule would have done. The firm says it plainly: many use cases presented as agentic do not need to be.

The market also feeds the confusion. Gartner talks about “agent washing”: vendors rebranding existing chatbots or automation tools as agents. Of the thousands of players claiming the term, only around 130 are thought to offer real agents. Before any demo, ask three questions. What does the tool decide on its own? With which data? And what does it do when it gets it wrong?

Where companies stand with AI agents in 2026

McKinsey’s global survey, published in August 2026, shows a clear acceleration. It covers 1,719 respondents surveyed in May and June 2026. 40% of respondents from companies with more than one billion dollars in revenue are scaling AI agents. A year earlier, the figure was only 27%. In smaller organisations, by contrast, the share is flat at 22%.

IT is leading the way. Today, around two in ten respondents are scaling software development agents. In large companies the figure reaches 31%, particularly for development. In addition, 32% of respondents have decided against buying at least one piece of software because they could build it in-house.

A minority is industrialising, the majority is testing

Across all companies, the picture remains cautious. The Capgemini Research Institute surveyed 1,500 executives in 14 countries. Only 2% of organisations have deployed agents at scale. A further 12% have deployed them at partial scale and 23% are running pilots. Finally, 61% are still exploring.

Admittedly, these two surveys do not measure the same thing. They do agree on one point: the gap is widening between those who industrialise and those who test. Financial value is not following yet either. At McKinsey, for example, 37% of respondents attribute an EBIT impact to AI, a stable figure. As for the highest-performing companies, they level off at around 6%.

8 concrete AI agent use cases, department by department

The use cases that hold up in production have one thing in common: they combine high volume, explicit rules, accessible data and a reversible action. The table below presents eight of them. For each one, it shows what the agent does and what must remain human.

Department	What the agent does	What remains human
Customer service	Qualifies the request, resolves simple cases (tracking, changes, standard refunds), passes on the others with context	Sensitive complaints, goodwill gestures above a threshold, guaranteed access to an adviser
Finance	Matches payments to invoices, prepares escalating reminders for overdue payments, flags anomalies	Legal proceedings, decisions on key accounts
Procurement	Compares supplier bids against a scoring grid, checks documents and certificates, prepares the summary	Supplier selection and negotiation
Human resources	Answers routine questions (leave, health cover, payroll), orchestrates the arrival of a new employee	Any decision that concerns a person: recruitment, appraisal, mobility
Sales	Prepares meetings (history, client news), updates the CRM after the meeting, suggests a follow-up	The relationship, sending messages to the client, the commercial proposal
Supply chain	Monitors stock-out risks, proposes rescheduling or a replenishment order	Approval of orders above a defined amount or risk level
Legal and compliance	Carries out a first review of contracts against a clause checklist, flags deviations	The final analysis and the lawyer’s opinion
IT	Handles level 1 support (access, resets), assists development and testing	Architecture, security, approval of production releases

1. Customer service: the most mature case

This is also the function where the promise is most quantified. Gartner predicts that by 2029, agentic AI will autonomously resolve 80% of common customer service issues. Operational costs would then fall by 30%.

The Klarna case, however, shows both the potential and the limit. In February 2024, its assistant handled 2.3 million conversations in one month, two-thirds of all chats. It was doing the work of 700 full-time agents. Resolution time also fell from 11 minutes to under 2. Yet in May 2025, its chief executive acknowledged to Bloomberg that cost had weighed too heavily in the decisions, and quality had suffered as a result. The company is now hiring again to guarantee every customer access to a human.

The lesson is simple. The agent should absorb the repetitive volume without removing the human fallback. The right indicator is therefore not the automation rate. It is first-contact resolution, and satisfaction on the cases handed over.

2. Finance: accounts receivable before the close

Collections are ideal ground. Volumes are high. The rules are known: due dates, amounts, payment history. And the action is gradual. In practice, the agent matches incoming payments and spots late ones. It then drafts reminders tailored to the customer’s profile. If needed, it prepares the escalation file. The accounting close will come later, once data quality has been demonstrated.

3. Procurement and legal: the first review

In these functions, the agent does not decide: it prepares. For example, it compares three bids on twenty criteria. It checks that a supplier has sent its certificates. It also reviews a contract against a clause checklist. These tasks take time, and a reliable first pass frees up expert time. One condition applies, however: an analysis grid written and validated by the experts.

4. Human resources: useful, but regulated

Answering employees’ routine questions remains a simple and well-liked use case. The same goes for onboarding new joiners. However, the AI Act classifies systems used to recruit or evaluate people as high-risk. Their obligations will apply from 2 December 2027: documentation, logging and effective human oversight. We set out this timeline in our guide to the AI Act.

5. Sales, supply chain and IT: the diffuse gain

Preparing a meeting takes a few minutes. So does updating the CRM, monitoring stock or handling an access request. Yet these tasks are repeated hundreds of times a week, which is why the cumulative gain becomes significant. This is also where development agents are progressing fastest.

Why more than 40% of AI agent projects will be abandoned

Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027. The firm gives three reasons: escalating costs, unclear business value and inadequate risk controls. Other surveys confirm these three weaknesses.

This drop in trust is not a bad sign. On the contrary, it marks the shift from demonstration to operation. Companies are discovering that an agent acting in an ERP engages their responsibility. The same is true when it writes to a customer. Successful projects therefore do not aim for maximum autonomy. They aim for the right level of autonomy at each step.

Choosing where to deploy AI agents: a 5-question grid

Before launching an agent, score each candidate use case against five criteria. If a case does not tick at least four boxes, a simpler solution is enough: a rule, an automation or an assistant will do the job.

Question	Favourable signal
Does the volume justify the effort?	Hundreds of occurrences a week, a measured time per unit
Are the rules written down?	A procedure exists and the experts agree on the right answer
Is the data accessible?	Systems expose APIs, and the data is reliable and up to date
Is the action reversible?	An error is easy to correct, or human approval comes before the action
Is the result measurable?	A baseline indicator exists: lead time, error rate, cost per case, satisfaction

This grid echoes McKinsey’s finding on the highest-performing companies: they redesign their processes around AI instead of inserting it into what already exists. Conversely, an agent placed on a poorly defined process mostly automates its flaws.

AI agents: human in the loop, by design

Human oversight is not just an “approve” button. It is designed in from the start, step by step. In practice, three levels are enough in most cases. First, the agent proposes and the human decides. Next, the agent acts below a threshold that has been defined (amount, type of customer, level of risk) and reports back. Finally, the agent acts alone on reversible actions, checked by sampling. In every case, each action must leave a trace: who asked for what, with which data, for which decision.

This framework also meets regulatory requirements. For example, a customer must know that they are talking to an AI. Likewise, high-risk uses require effective human oversight. Finally, it supports frugality: limiting unnecessary loops reduces both the bill and the footprint. We set out these levers in our article on frugal AI.

Where to start: 90 days to launch your AI agents

• Weeks 1 to 3: map. List ten candidate processes with the business departments. Score them with the grid, then select one or two. Finally, measure the starting point.
• Weeks 4 to 8: build in “proposal” mode. The agent prepares, then the human approves each action. Collect the gaps between the proposal and the final decision: they make the best test set.
• Weeks 9 to 12: measure and decide. Compare lead time, quality, cost per task and satisfaction with the starting point. Then extend autonomy only to the steps where the agent has proved itself.

In short, the point is not to have AI agents. It is to know precisely what you delegate to them. The companies that ask this question process by process, with the business teams, will move from demonstration to results. This is the approach we take in our AI and Data consulting engagements.

Sources

• McKinsey & Company, “The state of AI in 2026: On the road to ROI”, 25 August 2026 (survey from 4 May to 8 June 2026, 1,719 respondents, 97 countries). mckinsey.com
• Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, 25 June 2025. gartner.com
• Gartner, “Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029”, 5 March 2025. gartner.com
• Capgemini Research Institute, “Rise of agentic AI: How trust is the key to human-AI collaboration” (1,500 executives, 14 countries). capgemini.com
• Deloitte, “The State of AI in the Enterprise 2026”, press release of 21 January 2026 (3,235 executives, 24 countries). deloitte.com
• Klarna, “Klarna AI assistant handles two-thirds of customer service chats in its first month”, 27 February 2024. klarna.com
• TechSpot, “Klarna is hiring humans again as AI replacements offer lower quality output”, 12 May 2025, based on an interview given by Klarna’s chief executive to Bloomberg. techspot.com