AI Agents for Business: Where They Work in Production, Where They Burn Budget, and How to Decide
Every software vendor now sells an agent. Most companies are still asking the wrong question: which platform? The right question is which process, at what level of autonomy, at what cost per resolved outcome. This is the filter I use with leadership teams, with the evidence behind it.
Michele Cimmino
CEO & Founder · Lasting Dynamics
AI agents for business pay off today on one kind of work: high-volume, repetitive tasks where a wrong answer can be caught by a check, undone cheaply, and where the agent can reach the data and systems it needs through proper interfaces. Customer support on a well-kept knowledge base, IT service desk tickets, document intake and back-office reconciliation are the clearest cases. They burn budget on open-ended, multi-system work with irreversible consequences, on low-volume processes, and anywhere the policy itself is ambiguous. Decide process by process with five tests, start at a low level of autonomy, and measure cost per resolved outcome, not cost per conversation.
That is the short answer. The long answer matters because the market is loud and the numbers are sobering. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and McKinsey's 2025 survey found that no more than 10% of companies are scaling agents in any single business function. The gap between the demo and the production system is where most of the money goes.
Who this guide is for
You are a COO, a CIO, a CTO or a founder-CEO of a mid-market company, probably in finance, insurance, healthcare or energy. A vendor, a board member or your own team is pushing you to "do something with agents" and you need to decide where to start, how much to spend and what could go wrong. I use agents every day in my own company's software delivery, and the question comes up in almost every conversation I have with leadership teams this year. What follows is a method, not a product.
What an AI Agent Is, and What Vendors Call One
An AI agent is software built on a large language model that receives a goal, decides which steps to take, calls tools (your CRM, your ticketing system, a database, an email account) to carry them out, looks at the results and continues until the goal is met or it gives up. The difference from a chatbot is action: a chatbot answers, an agent does. The difference from classic automation is judgement: a workflow follows the path you drew, an agent chooses its own path within the limits you set.
Type
What it does
When it breaks
Typical use
Rule-based chatbot
Answers from scripted flows
Any question outside the script
FAQ, opening hours, order status
RPA bot
Repeats fixed clicks and keystrokes on screens
When a screen, field or rule changes
Copying data between legacy systems
Workflow automation (Zapier, n8n, Power Automate)
Runs a predefined sequence of API calls
When the case does not fit the sequence
Lead routing, notifications, simple approvals
LLM assistant (copilot)
Drafts text or code for a human to use
When nobody checks the draft
Email drafts, summaries, code suggestions
AI agent
Plans, calls tools, acts, checks results, loops
On long, ambiguous, multi-system tasks
Support resolution, ticket triage, document intake
Be careful with labels. Gartner warns about agent washing: vendors rebranding chatbots, RPA and assistants as agents. Its estimate, reported by Reuters on 25 June 2025, is that only about 130 of the thousands of vendors selling agentic AI offer real agentic capabilities. Ask any vendor to show the agent choosing between two tools on a case it has not seen, and to show the log of what it did. If there is no choice and no log, it is a workflow with a language model on top, which may be exactly what you need, at a lower price.
What the Evidence Says in 2026
The research on agents in business is young and much of it is written by vendors. These are the independent numbers I rely on, each with its date, because this market moves every quarter.
Source
Date
Finding
Gartner (via Reuters)
June 2025
Over 40% of agentic AI projects will be cancelled by end of 2027 because of escalating costs and unclear business value
Gartner (via Reuters)
June 2025
By 2028, 33% of enterprise software applications will include agentic AI and at least 15% of day-to-day work decisions will be made autonomously
McKinsey, The State of AI 2025
Nov 2025 (survey of 1,993 people, June–July 2025)
23% of organisations are scaling an agentic AI system somewhere, 39% are experimenting; no more than 10% are scaling agents in any single function
MIT NANDA, The GenAI Divide (via Fortune)
Aug 2025
About 95% of enterprise generative AI pilots show no rapid revenue acceleration; the biggest returns come from back-office automation
TheAgentCompany benchmark (Carnegie Mellon et al.)
2025
In a simulated software company, the best agent completed 30% of realistic office tasks autonomously
Moffatt v Air Canada, 2024 BCCRT 149
Feb 2024
A company is liable for what its chatbot tells customers, exactly as for any other page of its website
Read together, these numbers say three things. Adoption is real but shallow: experimenting is common, scaling inside a function is rare. Autonomy on complex work is still weak: a 30% completion rate on realistic multi-step tasks is impressive research and an unacceptable production system. And the returns concentrate where the work is narrow and repetitive, which is rarely where the budget goes first. MIT found that more than half of generative AI budgets go to sales and marketing tools, while the biggest returns sit in the back office.
Two Stories Every Board Should Know
Klarna: the success story with a second chapter
In February 2024 Klarna announced that its AI assistant, built with OpenAI, had handled 2.3 million customer conversations in its first four weeks: two-thirds of all its service chats, in 35 languages, doing the work of 700 full-time agents. Resolution time fell from 11 minutes to under 2, repeat inquiries fell by 25%, and the company expected a USD 40 million profit improvement in 2024. It remains one of the best documented agent deployments in customer service.
The second chapter is just as useful. In May 2025 the CEO told Bloomberg that Klarna would make sure customers can always reach a human, and that investing in the quality of human support was the way forward, acknowledging that the AI-first push had cost quality. The lesson is not that the agent failed. It handled the volume. The lesson is that the design goal was cost per conversation, and the business needed quality per customer. Design for the outcome you actually care about, and keep a human route open from day one.
Air Canada: the agent speaks for the company
In Moffatt v Air Canada (British Columbia Civil Resolution Tribunal, 14 February 2024) the airline's website chatbot told a grieving passenger he could apply for a bereavement fare after travelling. The airline's own policy page said the opposite. Air Canada argued it could not be held responsible for what the chatbot said. The tribunal disagreed: a company is responsible for all the information on its website, whether it comes from a static page or a chatbot, and it awarded the passenger CAD 812.02.
The amount is small. The principle is not. Whatever your agent says or does in front of a customer, your company said or did. In Europe the bar is higher still: since 2 August 2026, Article 50 of the EU AI Act requires that people are told when they are interacting with an AI system. Liability and transparency belong in the design of the agent, not in a disclaimer added at the end.
The Five-Test Agent Filter
Before choosing a platform, take your candidate processes, ten of them if you can, and run each one through five tests. I call it the Five-Test Agent Filter. A process that passes all five is a good first agent. A process that fails two or more is a project that will probably join Gartner's 40%.
Volume. Does the process run hundreds or thousands of times a month? Agents have fixed costs (integration, evaluation, monitoring) that only volume pays back. A task done twenty times a month is cheaper done by a person.
Verifiability. Can a rule, a second model or a quick human look tell whether the result is right? If the only way to check an answer is to redo the work, the agent saves nothing.
Reversibility. What does a wrong action cost, and can it be undone? Drafting a reply is reversible. Issuing a refund, changing a policy, sending money or deleting data usually is not.
Reachability. Can the agent reach the data and systems it needs through stable APIs with proper permissions? If the answer is screen scraping a legacy system, fix the plumbing first.
Exception rate. What share of cases fall outside the documented rules? Below roughly one in five, an agent handles the routine and escalates the rest. Above that, you are automating ambiguity.
Test
Passes when
Warning sign
Volume
Hundreds or more cases a month, steady demand
Seasonal spikes only, or under fifty cases a month
Verifiability
Results can be checked by rule, sample or second model
Only an expert redoing the work can judge it
Reversibility
Wrong actions are cheap and can be undone
Money, contracts, health or legal status change on one click
Reachability
APIs, clean data, role-based permissions
Screen scraping, shared logins, data in spreadsheets
Exception rate
Most cases follow documented rules
Policies live in people's heads and change weekly
The test that kills most projects
In my experience, verifiability is the test that decides. Teams get excited about a process with huge volume, then discover that nobody can say quickly whether the agent's output is right. If you cannot measure accuracy on day one, you cannot improve it, you cannot defend it to an auditor and you cannot prove the savings to your CFO. Build the answer key before you build the agent.
Where Agents Work Today, and Where They Burn Budget
Applied to real processes, the filter produces a fairly consistent map. McKinsey's survey reports that agent use is most common in IT, for example service-desk management, and in knowledge management, for example deep research. MIT's study points to back-office automation for the highest returns. My own experience matches both.
Process
Fit today
Why
Tier-1 customer support on a maintained knowledge base
Strong
High volume, answers can be checked against sources, human escalation is easy
IT service desk (password resets, access requests, triage)
Long chains of steps compound errors; benchmarks show low completion
Negotiation, complaints, vulnerable customers
Keep human
Judgement, empathy and accountability matter more than speed
The pattern behind the table is simple. Agents are strong where the work is narrow and the check is cheap. They are weak where errors compound across many steps: if each step of a ten-step task is right 95% of the time, the whole chain is right only about 60% of the time. That arithmetic, not the model's intelligence, is why open-ended autonomy disappoints.
Projects that burn budget usually start from the platform: a licence bought first, a use case found afterwards.
They automate a broken process: the agent inherits unclear rules and reproduces them faster.
They measure activity, not outcomes: conversations handled instead of cases resolved without a second contact.
They skip the integration work: the agent can talk but cannot reach the systems where the work happens.
The Autonomy Ladder: Decide How Much the Agent May Do
The second decision, after which process, is how much freedom the agent gets. Most failures come from starting at the top of the ladder. I use four levels and move a process up one level only when the data from the previous one justifies it.
Level 1 — Suggest. The agent proposes an answer or an action; a person decides and acts. Zero external risk, immediate learning about accuracy.
Level 2 — Draft for approval. The agent prepares the full action (the reply, the ticket update, the extracted record); a person approves with one click. This is where most processes should spend their first months.
Level 3 — Act with an undo window. The agent acts on its own for low-risk cases, inside limits (amounts, customer segments, case types), with logging and an easy way to reverse.
Level 4 — Act within limits, audit after. The agent acts autonomously on a defined class of cases; people review samples and exceptions. Reserve it for processes with months of measured accuracy.
Each level has its own success metric. At levels 1 and 2, measure the approval rate without edits: when people accept the agent's draft unchanged nine times out of ten for a given case type, that case type is a candidate for level 3. At levels 3 and 4, measure reversals, complaints and second contacts. Write the promotion rule down before the pilot starts, so the decision to give more autonomy is based on data and not on enthusiasm.
What an AI Agent Really Costs
Vendor pricing has converged on usage. These are the public list prices of two of the largest platforms, checked on 11 November 2026. They are a useful reference even if you build your own, because they show what the market thinks one unit of agent work is worth.
Platform
Pricing unit
List price
Source, date checked
Intercom Fin (support agent)
Per outcome
USD 0.99
intercom.com/pricing, 11 Nov 2026
Salesforce Agentforce
Per conversation
USD 2.00
salesforce.com/agentforce/pricing, 11 Nov 2026
Salesforce Agentforce (Flex Credits)
Per standard action (20 credits; USD 500 per 100,000 credits)
USD 0.10
salesforce.com/agentforce/pricing, 11 Nov 2026
Salesforce Agentforce (Flex Credits)
Per voice action (30 credits)
USD 0.15
salesforce.com/agentforce/pricing, 11 Nov 2026
The licence is rarely the largest cost. The number that matters is the cost per resolved outcome: everything you spend on the agent in a month, divided by the cases it closed correctly without a second contact or a human redo. A conversation the agent handled badly is not a saving; it is a cost plus a second conversation plus, sometimes, a complaint.
Integration: connecting the agent to your systems with proper permissions. In legacy-heavy companies this is often the largest one-off cost.
Evaluation: building and maintaining a set of real cases with correct answers, and running it after every change of model, prompt or policy.
Human review: the time people spend approving drafts at levels 1 and 2 and sampling at levels 3 and 4.
Escalations and redos: the cases the agent passes on, plus the ones it got wrong and someone had to fix.
Monitoring and ownership: a named owner, dashboards, incident handling and periodic review of the logs.
An illustrative example, with numbers you should replace with your own. A support team receives 10,000 tickets a month. An agent priced per outcome resolves 5,000 of them correctly at USD 0.99 each, about USD 5,000 a month. Add a part-time owner, evaluation runs and the review time of the human team, and the real monthly cost may be double the licence. Compare that figure with your fully loaded cost of resolving 5,000 tickets with people today. If the agent is not clearly cheaper per resolved ticket, at equal or better customer satisfaction, after six months, stop or change the process.
Ask for the denominator
When a vendor quotes a resolution rate, ask how resolution is defined. "The customer did not reply" and "the customer's problem was solved" can differ by tens of percentage points. Agree on the definition in the contract, and measure it yourself with a sample of real cases every month.
Build, Buy or Wait
MIT's study offers one of the few hard data points on this choice: tools bought from specialised vendors, or built in partnership with them, succeeded about 67% of the time, while purely internal builds succeeded about a third as often. That does not mean never build. It means build only where the agent is part of what makes your company different, and where you have the data and the engineers to evaluate it continuously.
Buy when the process is common to your industry (support, IT service desk, invoice processing) and a vendor already integrates with the systems you use.
Build when the process is specific to you, touches proprietary data or core systems, and the agent's quality is a competitive advantage. Budget for evaluation and operations, not just development.
Wait when the process fails two or more of the five tests. Spend the money on the plumbing (APIs, data quality, documented rules) and come back in six months: the agent will be cheaper and the process ready.
For the full method behind this choice, including the five questions I use to score any software investment, see my build vs buy software framework. If your systems cannot expose clean APIs, the agent project is really a modernisation project; my guide on legacy system modernization explains how to sequence it.
Governance Without Bureaucracy
An agent is an employee who never sleeps and never asks before acting unless you make it. It needs the same things any new role needs: a job description, limits, a manager and a record of what it did. In the EU it also needs transparency: since 2 August 2026, people must be told when they are dealing with an AI system, and the fines for breaching Article 50 reach EUR 15 million or 3% of worldwide turnover. If the agent touches credit, insurance pricing, hiring or access to essential services, check whether it falls into the high-risk categories of the AI Act before you deploy it.
A named business owner for each agent, accountable for its results as for any team's.
Written limits: which actions, which amounts, which customer segments, which data.
Complete logs of inputs, tool calls and outputs, kept long enough to answer a complaint or an audit.
A kill switch that someone outside the project team can use, and a tested fallback to the human process.
An evaluation set of real cases, run before every change and reviewed monthly.
Clear disclosure to customers and an always-available route to a person.
Starting with the most impressive use case instead of the most verifiable one.
Giving the agent level 4 autonomy in the pilot because the demo looked good.
Counting conversations instead of resolved outcomes, and declaring savings that the second contact erases.
Hiding the AI from customers, which is now illegal in the EU and destroys trust when discovered.
Removing the human route to hit a cost target, as Klarna's own reversal shows.
No owner after go-live: models, prompts and policies drift, and nobody runs the evaluation set.
A 90-Day Pilot Plan
Weeks 1–2 — shortlist. List ten candidate processes with their monthly volume and run the Five-Test Agent Filter on each. Keep the one or two that pass all five.
Weeks 3–4 — define success. Write the definition of a resolved outcome, today's cost per resolved outcome with people, the target, and the promotion rule between autonomy levels.
Weeks 5–6 — build the answer key. Collect 200 to 500 real past cases with their correct outcomes. This becomes the evaluation set for any vendor or build.
Weeks 7–10 — run at level 2. The agent drafts, people approve. Measure approval without edits, time saved per case and errors by case type, weekly.
Weeks 11–12 — promote or stop. Move the case types that meet the promotion rule to level 3 with an undo window. Leave the rest at level 2 or drop them.
Week 13 — decide. Compare cost per resolved outcome and customer satisfaction with the baseline. Scale, change the process, or stop, and write down why.
Conclusion
AI agents for business are neither a fad nor a replacement for your organisation. They are a new kind of worker that is excellent at narrow, high-volume, checkable work and unreliable at long, ambiguous, irreversible work. The companies that get value from them choose the process before the platform, start at low autonomy, measure cost per resolved outcome and keep a human route open. The ones that do not will be part of the 40% of cancelled projects, having learned the same lessons at a higher price.
If you want an independent second opinion
I help leadership teams run exactly this exercise with their own processes: the five-test shortlist, the cost baseline, the evaluation set and the first ninety days of a pilot. I have no agent platform to sell, and sometimes the right answer is to fix the process first and buy the agent next year.
Unsure whether your systems are ready for agents at all? The usual blocker is not the model but the integration layer, which is why I start most of these conversations with the questions in my legacy modernization guide.
AI-ready answers
Frequently Asked Questions
Which business processes should we give to AI agents first?+
Start with high-volume, repetitive processes where a wrong result can be caught by a rule, a sample or a second model, undone cheaply, and where the agent reaches data through stable APIs: tier-1 customer support on a maintained knowledge base, IT service desk tickets and document intake such as invoices or claims forms. McKinsey's State of AI 2025 found agent use most common in IT service desks and knowledge management, and MIT's GenAI Divide study found the biggest returns in back-office automation. Keep negotiations, complaints and irreversible credit or claims decisions with people.
How much does an AI agent cost to run in production?+
List prices checked in November 2026: Intercom Fin charges USD 0.99 per outcome and Salesforce Agentforce USD 2 per conversation or USD 0.10 per standard action through Flex Credits. The licence is rarely the largest cost: integration, an evaluation set of real cases, human review, escalations and a named owner can double it. Measure cost per resolved outcome, meaning total monthly spend divided by cases closed correctly without a second contact, and compare it with your current cost per case handled by people.
Should we build our own AI agent or buy one?+
Buy when the process is common to your industry and a vendor already integrates with your systems; build only when the process is specific to you, touches proprietary data or core systems and agent quality is a competitive advantage. MIT NANDA's 2025 study found tools bought from specialised vendors or built with partners succeeded about 67% of the time, while purely internal builds succeeded about a third as often. Wait when the process fails two or more of the five tests and fix APIs, data and documented rules first.
I help COOs, CIOs and CEOs pick the right first process, set the cost baseline and run a 90-day pilot that ends with a decision backed by data. No platform to sell.