What AI agents actually do for a business in 2026, and what not

What is an AI agent, in plain terms?

An AI agent is a program built around a large language model (LLM), the kind of model behind Claude or ChatGPT, that can plan a sequence of actions, call tools to execute them and check its own results, instead of just answering a question. A chatbot talks. An agent does.

The difference matters for the budget. A chatbot saves minutes per conversation. An agent removes whole steps from a process: it reads the ticket, looks up the order, issues the refund within its limit, updates the CRM and writes a 3-line summary for the person who reviews exceptions. The value is in the steps it removes, not in the words it writes.

Which business processes do AI agents run today?

Today AI agents run 4 kinds of processes with little supervision: customer support, back office operations, sales operations and software engineering. Everything else is either assisted work or still an experiment.

Customer support is the most mature. Salesforce reported in June 2026 that its own Agentforce handled 4.3 million enquiries on its help site and resolved 70% autonomously, and it now charges only when the agent resolves a case. Klarna's assistant does the work of 853 full-time agents and handles 2 out of 3 enquiries, according to Klarna's Q3 2025 results as reported by CX Dive.

Back office is the quiet winner: invoice matching, expense checks, contract data extraction, monitoring and writing the incident report. Nobody writes a press release about it, and it is where most of our saved hours come from. Software engineering changed the fastest. Agentic coding tools such as Claude Code, Codex and Cursor now write, test and open pull requests on their own. The Stanford AI Index 2026 reports that agent success on real-world terminal tasks rose from 20% in 2025 to 77.3%. I still write code every week, but most of my time goes to reviewing what agents wrote.

• Support: triage, answers with account context, refunds within a limit, escalation with a summary.
• Back office: invoice reconciliation, contract extraction, compliance checks, monitoring and incident notes.
• Sales operations: lead enrichment, CRM hygiene, meeting notes to next steps, first-draft proposals.
• Engineering: features from a ticket, test writing, dependency upgrades, code review.

Where does a person still stay in the loop?

A person stays in the loop wherever an action is irreversible, expensive, legally binding or visible to a customer who expects a human: payments above a limit, contract terms, hiring, pricing, public communication and any decision that would embarrass the company if it went wrong.

This is not a temporary constraint. Agents are good at execution and bad at judgement on incomplete information. When a customer says "the invoice is wrong", the agent can find the invoice, compare it with the contract and propose a correction. Deciding whether to make a goodwill exception for a customer about to churn is still a human call.

The practical design is a review queue: the agent does 100% of the work and a person approves the 5% to 15% of cases it flags. In my companies the person who used to do the task now supervises 10 times the volume. Klarna learned this the hard way: after cutting support staff in 2024, it resumed hiring in 2025 and let customers request a human, because generic answers hurt satisfaction, according to CX Dive.

What do AI agents not do yet?

AI agents do not yet run a whole department, negotiate with unpredictable humans, work reliably across 20 legacy systems without APIs, or stay on task for days without drifting. Anyone selling you that in 2026 is selling a demo.

The data supports the caution. According to McKinsey's State of AI 2025, 23% of organisations are scaling an agentic AI system somewhere in the enterprise, another 39% are experimenting, and in any single business function no more than 10% are scaling agents. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls, and estimated that only about 130 of the thousands of vendors claiming to sell agents were real.

The benchmarks tell the same story from the other side. The Stanford AI Index 2026 reports that agents jumped from 12% to about 66% success on OSWorld, a test of real computer tasks, which also means they still fail roughly 1 in 3 attempts. In a process that runs 1,000 times a day, 1 in 3 is not a rounding error. It is why limits and review queues are not optional.

What are the failure modes of AI agents in a company?

The 5 failure modes I have actually paid for are hallucinated facts presented with confidence, prompt injection through untrusted text, runaway costs from loops, permissions that were too wide, and the absence of evaluations before launch.

Hallucination is the famous one and, in practice, the least dangerous, because a review queue catches it. Prompt injection is the one that keeps me up: an agent reads a customer email or a support ticket that contains hidden instructions and follows them. The fix is architectural, not a better prompt: treat everything the agent reads as data, never as instructions. Runaway costs happen when an agent retries a failing step for 6 hours; budgets per task and per day solved that for us. Wide permissions and missing evaluations are the 2 mistakes I see most in other companies.

How do you start with AI agents in 30 days?

Start with 1 process that runs at least 50 times a week, is already documented, has a clear definition of "done" and a low cost of error. Give it to an agent in read-only mode, measure it for 2 weeks, then allow write actions with human approval.

This is the plan I give to every CEO who asks. It costs 1 person part time for a month and less in inference than 1 monthly salary. The hardest part is not the technology. It is choosing a process narrow enough to finish, and resisting the temptation to automate the whole department in week 2. The signal that it is working is simple: the people who used to do the task spend their time on exceptions and customers, and the metric you chose in week 1 moved.

• Week 1: pick the process, write down the 10 steps, define the metric (time to resolution, cost per case, error rate).
• Week 2: connect the agent read-only and let it draft, not act; compare 100 outputs against what people did.
• Week 3: allow the safe write actions with a person approving each one; set budgets and logs.
• Week 4: review the metric, document the failure cases, decide whether to scale, fix or stop.

Is it worth it for a small or mid-sized company?

Yes, and small companies often win more than large ones, because they have fewer legacy systems, shorter approval chains and processes that 1 person understands end to end.

An agent running support or back office costs a fraction of a salary and works 24 hours a day in every language your customers speak. Anthropic's Economic Index for June 2026 found that 86% of surveyed workers reported gains in speed, 82% in scope and 69% in quality, and 57% said AI made their skills more valuable. The limit for a small company is attention, not money. Somebody has to own the agents, review the queue and keep the evaluations current. Give it an owner and a metric, and it pays for itself within the quarter.