AI agents in companies in 2026: what works today and what is still hype

What an AI agent is, in plain words

A chatbot answers a question. An AI agent receives a goal and a set of tools, and decides by itself which steps to take: read an email, look up the order in the ERP, draft a reply, update a record. It uses a large language model to understand text and choose actions, and ordinary software to call the tools.

The ability to act is what makes agents useful, and also what makes them risky. A wrong answer from a chatbot is a wrong answer. A wrong action from an agent is a wrong invoice, a wrong email to a customer or a deleted record.

The word is also used loosely. In 2025 Gartner warned about "agent washing", the relabelling of existing chatbots and automation tools as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI were real. When someone offers you an agent, ask two questions: what exactly can it do without a person, and what happens when it is wrong?

What works reliably today

The use cases that work in 2026 have something in common. The task is narrow, the input is mostly text or documents, and a person checks the result before it has consequences.

Document processing with human review

Supplier invoices, delivery notes, contracts, customs papers and certificates arrive in many layouts. An agent can read them, extract the fields, compare them with the purchase order and flag differences. A person approves or corrects the result. This removes most of the typing and keeps a human responsible for the final record.

Internal knowledge assistants

Procedures, technical manuals, past support tickets and internal policies are often spread across folders nobody searches. An assistant that answers staff questions from your own documents, and shows which document it used, saves time for new employees and for the experienced colleagues who usually answer those questions. The source reference is important: it lets people check the answer instead of simply trusting it.

Drafting and triage

Sorting incoming emails or support requests by topic and urgency, routing them to the right person and preparing a first draft of the reply works well. The agent prepares, a person reviews and sends.

Data entry between systems

Many companies still copy data by hand from one system to another because there is no integration: from an emailed order into the ERP, from a supplier portal into a spreadsheet. An agent can do this transfer, with validation rules and a review queue for anything unusual. Where a proper API exists, a classic integration is usually cheaper and more predictable. The agent earns its place where the input is unstructured.

Where agents still fail

Unsupervised decisions

Agents are much better than two years ago, but they are not reliable enough to make consequential decisions alone. The Stanford AI Index 2026 reports that on OSWorld, a benchmark of real computer tasks, the best systems went from about 5 percent of tasks solved in 2024 to 66.3 percent in 2025, still below the human baseline of 72.35 percent. That is fast progress, and it also means a meaningful share of multi-step tasks still go wrong. Approving payments, changing prices or making commitments to customers should stay with people.

Messy data

An agent working on inconsistent master data, such as duplicate customers, outdated price lists or product codes typed as free text, will produce confident results based on the wrong records.

Unclear processes

If three employees handle the same request in three different ways, there is no single process for an agent to follow. Automating a process means writing it down first, including the exceptions.

Cost and latency

Every step an agent takes is at least one call to a model. Long chains of steps take seconds or minutes, and the running cost grows with volume. For a high-volume task with clear rules, a normal program is faster and cheaper. Gartner predicted in 2025 that over 40 percent of agentic AI projects will be canceled by the end of 2027, pointing to escalating costs, unclear business value or inadequate risk controls.

Security

Agents read untrusted text: emails, web pages, attachments. OWASP ranks prompt injection, where text in the input changes the model's behaviour, as the first risk for applications built on language models, and notes that no known method fully prevents it. Its recommendations include giving the model only the access it needs and requiring human approval for high-risk actions.

How to start safely

• Pick a narrow task. One document type, one mailbox, or one transfer between two systems.
• Keep a human in the loop. The agent proposes and a person approves, at least until you have evidence that it is reliable.
• Limit permissions. Give the agent its own credentials and access only to the data and actions the task requires.
• Log everything. Inputs, outputs, tools called, and who approved what. Logs are how you find errors and show what happened.
• Protect the data. Know where data is processed and stored, what the model provider may do with it, and whether personal data leaves the EU.
• Set a measurable goal. Hours saved per week, processing time or error rate. Measure before and after, then decide whether to extend.

Data protection and the EU AI Act

If the agent processes personal data, the GDPR applies as it does to any other system. You need a legal basis, a data processing agreement with each provider, and attention to Article 22, which restricts decisions based solely on automated processing that significantly affect people. Real human review often keeps a use case within this rule.

The EU AI Act, Regulation (EU) 2024/1689, applies in stages. Prohibited practices and AI literacy obligations have applied since 2 February 2025, and the rules for general-purpose AI models since 2 August 2025. Most of the Act became applicable on 2 August 2026, including transparency rules such as informing people when they interact with an AI system. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the high-risk rules: to 2 December 2027 for the stand-alone high-risk uses listed in Annex III, such as recruitment or access to essential services, and to 2 August 2028 for AI built into products covered by EU product legislation (Annex I).

Most internal assistants and document tools will not fall into the high-risk category, but check carefully if your use case touches hiring, credit, education or similar areas. The Act can also apply to providers outside the EU when a system or its output is used in the EU, which matters for companies in Moldova that work with EU clients.

How NOWANI TECH can help

We build AI agents and automations for specific, measurable tasks, with human review, logging and access limits designed in from the start. We will also tell you when a normal integration or a simple script will do the job better. If you have a process you would like to test, get in touch.