On the slides, an AI agent does everything: reads email, orders stock, reconciles invoices and writes the board report. In the companies we talk to, the picture is more modest. The agent handles one process, does it well or so-so, and someone has to watch the results. That is not a bad outcome, as long as you know from the start what to expect.
This text is for people who have to decide whether and where to invest. No hype, no scaremongering.
What the adoption data says
Start with scale. According to McKinsey's "The state of AI in 2025", 62% of surveyed organisations are at least experimenting with AI agents, but only 23% are scaling an agent somewhere in the enterprise. Most of those do it in one or two functions. Nearly everyone experiments, and a handful run agents in production.
Other sources paint a similar picture, though they need careful reading. An MIT NANDA report from 2025 (150 interviews, 300 deployments analysed) claims that around 95% of generative AI pilots deliver no measurable impact on the bottom line. The figure circulates in the press as "95% of projects fail", which is a simplification. The authors blame poor fit between tools and company processes, not the quality of the models. The methodology also raises questions, so I would treat 95% as a signal of direction, not a hard number.
In June 2025 Gartner predicted that more than 40% of agent projects will be cancelled by the end of 2027. The reasons: rising costs, unclear business value and weak safeguards. Analysts also point to "agent washing", where ordinary chatbots and automations get relabelled as "agents".
The conclusion is dull but important: technology rarely fails. Process choice and expectations do.
Where agents actually save time
The best results come where the task is narrow, repetitive and easy to check.
Customer support. A well-known study by Brynjolfsson, Li and Raymond (NBER, over 5,000 agents at a large software company) found that an AI assistant raised the number of issues resolved per hour by about 14%, and by 34% for novices. Note the setup: the AI suggested, the human decided. Agents used fewer than 40% of the suggestions on average.
Document processing. Invoices, contracts, orders, complaints. The agent extracts the fields, compares them with data in the system and flags discrepancies. A person approves the exceptions. The value is countable here: time per document, share of corrections.
Reporting and data preparation. A weekly sales report, a summary pulled from several systems, a draft commentary on the results. The agent collects and describes, the analyst checks and adds conclusions.
Back office. Reconciling data between systems, checking that applications are complete, routing requests to the right department. Dull work, but that is exactly where MIT sees the best return, even though the same report says most AI budgets go to sales and marketing.
What they share: the agent does not make high-stakes decisions. It prepares work for a person, or closes cases where a mistake costs little.
Where they still fall short
The TheAgentCompany benchmark from Carnegie Mellon simulates a software company and gives agents 175 office tasks. The best agent completed about 30% of them on its own (24% in an earlier version of the study). That result is several models old and is probably higher today, but it shows the right order of magnitude: long, multi-step tasks across many systems are still a weak spot.
From what we see, agents fail most often in these situations:
- processes where data sits in five systems with no proper API access,
- tasks that need judgement and context missing from the documents (for example "this client matters because he knows the CEO"),
- long chains of steps, where an error in step three ruins everything up to step ten,
- conversations with customers about emotional or unusual matters.
A loud example is Klarna. In 2024 it boasted of an assistant doing the work of 700 people, and in 2025 it started hiring support staff again because the quality of conversations had dropped. The company did not abandon AI, it changed the mix. The lesson is simple: savings that show up in the cost line can be eaten by quality nobody was measuring.
Risks the board should know
Hallucinations. A model can state something false with full confidence. The cure is not a "better prompt". It is limiting the task to sources the agent can cite, and checking the output.
Prompt injection. An agent reads emails, documents and web pages. Someone can hide an instruction in them that the agent will carry out. In the 2026 OWASP Top 10 for LLM applications, prompt injection is number one, and "excessive agency" moved up from sixth to third place as agents gain more permissions. OWASP's advice is architectural: assume the model can be fooled, and limit what it is able to do.
Permissions. An agent with access to the whole ERP and a mailbox is the same problem as an intern with the administrator password. Every tool should get the minimum permissions.
Costs. An agent calls the model many times per task. At a pilot of 50 cases a day the cost is laughable, at 5,000 it is not. Count the cost of the whole task, not a single request.
No observability. If you do not know what the agent did and why, you also do not know when it stopped working well. This is the most common gap in projects that reach us after a failed pilot.
Patterns that work
In its review of deployments, Anthropic notes that the best systems are built from simple, composable parts rather than heavy frameworks, and advises looking for the simplest solution that works. It also distinguishes a workflow (code fixes the path, the model carries out the steps) from an agent (the model decides the order itself). Most business processes need only a workflow. We use agent-style freedom only where it is truly needed.
The practical set we use:
- A human in the loop wherever an error costs money: approval before sending, paying or changing something in a system.
- Narrow scope. One process, one document type, one role. A broad "do-everything" agent has no good way to prove itself.
- Tools with limited permissions. Read instead of write where possible. Caps on amounts and on the number of operations.
- Evaluations. A set of a few dozen real cases with the expected result, run after every change. Without it, "we are improving the prompt" means "we are guessing".
- Logging. Every step, tool call and human decision, for audit and for learning.
A simple test: is this process right for an agent
Before choosing a process, go through four questions. Score each from 1 to 3.
- Repeatability. Does the process happen at least a few dozen times a month and run in a similar way? (3 = daily and by the book, 1 = every case is different)
- Cost of error. What happens if the agent gets it wrong and nobody notices? (3 = a minor fix, 1 = financial loss, a penalty or a lost client)
- Data availability. Does the agent have access to the information it needs, in digital and up-to-date form? (3 = one system with an API, 1 = paper, email and knowledge in people's heads)
- Measurability. Can you say what a "good result" is and what it costs today? (3 = there is a metric and a baseline, 1 = "it should be better")
Score 10-12: a good candidate. 7-9: tidy up the data first or add human control. Below 7: put it off, this is not a good first pilot. The process that wins is usually the one with high repeatability and a low cost of error, even if it is not the most spectacular.
A 6-week pilot
Six weeks is enough to see how things really are, and too short for the project to sprawl.
Week 1: choice and baseline. One process, picked with the test above. Measure the current state: time per case, error rate, cost. Name an owner on the business side, not only in IT.
Week 2: data and test cases. Collect 50-100 real cases with correct results. Decide what the agent can access and what it cannot.
Weeks 3-4: build and evaluation. A first version, run on the test set. Fixes based on errors, not impressions. Set up logging from day one.
Week 5: parallel run. The agent prepares results, a person approves them, and everything is compared with the existing process. Collect comments from the people who will use it.
Week 6: decision. Compare against the week 1 baseline: time, quality, cost per task (including tokens and the approver's time). Three outcomes are possible: scale up, fix for one more cycle, or stop. A well-run pilot allows the third option without embarrassment. Better to close it after 6 weeks than to keep a ghost project alive for a year.
What to do on Monday
List five processes that wear your team out and score them with the four questions. Pick one. Ask who on the business side would own it and how you measure its result today. If any of those questions has no answer, the first job is putting the process in order, not technology.
In the projects we run, most of the time goes on exactly that: choosing the process, the data and the way of measuring. The agent itself is usually the easiest part. If you want to go through such an assessment together, we are happy to talk.
Sources
- McKinsey, The state of AI in 2025: Agents, innovation, and transformation
- Fortune, MIT report: 95% of generative AI pilots at companies are failing
- CRN, Gartner: 40% of agentic AI projects could be canceled by 2027
- NBER, Brynjolfsson, Li, Raymond: Generative AI at Work
- arXiv, TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks
- SD Times, Prompt Injection tops 2026 OWASP GenAI / LLM Top Ten vulnerabilities
- Anthropic, Building effective agents
- Tech.co, Klarna Reverses AI Customer Service Overhaul
