AI agents in the enterprise: what they are, what they do well and where they fail

An AI agent is a system that receives a goal, consults the data and tools it is allowed to use and carries out the steps needed to achieve it, within defined limits. In the enterprise it works well on text-heavy tasks with clear criteria, such as answering queries, triaging documents or extracting data. It fails when given unlimited autonomy, unreliable data or no supervision.

What exactly is an AI agent

An AI agent combines three parts: a language model that understands and writes text, a set of tools it can use (searching documents, querying an ERP, creating a record) and instructions that define what it may and may not do.

The difference from classic automation lies in the input. Classic automation needs structured data and exact rules. An agent can work with a hastily written email, a scanned contract or a question phrased in any way.

	Classic automation	Conversational assistant	AI agent
Input	Structured data	Natural-language questions	Goals and free text
How it decides	Fixed rules	Generates an answer	Reasons and chooses tools
What it does	Runs defined steps	Answers	Queries, prepares and acts
When to use it	Stable, exact processes	Frequent questions	Tasks needing text and judgement

In practice, the best systems combine all three. The agent interprets, classic automation executes and a person approves what matters.

Which tasks AI agents handle well

Agents perform best where there is a lot of text, a lot of volume and a criterion that can be explained with examples:

• Internal queries. Answering staff questions on policies, procedures or leave, citing the source document.
• Triage. Reading emails, tickets or requests and routing them to the right team with the key information already extracted.
• Document data extraction. Pulling amounts, dates and references from invoices, delivery notes or contracts and recording them.
• Drafts. Preparing customer replies, meeting summaries or reports for a person to review.
• Search across your own documentation. Finding the exact clause, datasheet or procedure among thousands of files.

In all of them the agent does not replace the organisation’s judgement. It applies it faster and more consistently.

Where AI agents fail

A language model can produce an answer that sounds right and is wrong. This is known as hallucination. The risk grows in four situations:

1. When absolute precision is required without verification. An amount or a date should never depend on the model’s wording alone.
2. When source data contradicts itself. If two internal documents disagree, the agent will pick one, and perhaps not the right one.
3. When it has no limits. An agent that can send, pay or delete without approval is an operational risk.
4. When nobody measures quality. Without a weekly reviewed sample, errors pile up unnoticed.

None of these failures is inevitable. All of them are solved in the design, not in the model.

How to design a reliable AI agent

A reliable agent looks more like a good procedure than an experiment. These are the practices we apply:

1. Narrow scope. One specific task with a written success criterion.
2. Answers grounded in your documents. Retrieval-augmented generation (RAG) means the agent first searches the organisation’s documentation and answers from what it finds, citing the source.
3. Minimal permissions. The agent accesses only what it needs and inherits the permissions of the user working with it.
4. Human approval for anything irreversible. Sending, paying or changing master data goes through a person.
5. Tests before every release. A set of real cases with their correct answers that the agent must pass.
6. A log of every decision. What it received, what it consulted and what it answered, so it can be audited.
7. Continuous measurement. Quality, time and escalations, measured against the starting point.

What data an agent works with

A useful agent works with your data: your documentation, your case history and your systems. That is why the question of where that data is processed matters as much as the choice of model.

The recommended approach is to decide with your technology team where the agent is deployed, in your own infrastructure or the provider’s, according to your security policy. In both cases, require models in private environments and contracts that prevent your data from being used to train third-party models. We explain this in AI and GDPR: keeping data in-house and on our security page.

How much autonomy to give

Autonomy is not a switch. It is a scale, and each task sits at the level its risk allows:

• Suggest. The agent proposes; a person decides and acts.
• Prepare. The agent completes the work; a person approves it with one click.
• Act with review. The agent acts; a person reviews a sample afterwards.
• Act. The agent works alone, on low-risk, high-repetition tasks.

The prudent approach is to start one level below what seems possible and move up when the measurements justify it.

How to tell whether an agent is working

Define four indicators before you start and measure them against the previous situation:

• Share of cases resolved without intervention.
• Error rate in a sample reviewed by people.
• Response time compared with the manual process.
• Escalated cases and the reason for each escalation.

If an indicator worsens, the agent drops back to the previous autonomy level. That protects operations while the system improves.

What an organisation needs to deploy an AI agent

Less than is often assumed. You do not need a data science team or a new platform. You need four things:

1. A specific task, with a business owner who can recognise a correct answer.
2. Access to the task’s data: the documentation, case history or systems where the information lives.
3. A set of test cases, drawn from real operations, each with its correct answer.
4. A measurable success criterion, fixed before starting and compared with the starting point.

With that, the agent is designed, tested against those cases and delivered in stages. After go-live, ongoing support means reviewing its quality, updating its sources and adjusting its limits when the process changes.

Where to start

Choose a task with plenty of text, plenty of volume and an owner who can recognise a correct answer. That is the best first agent. To see how we approach it, look at our AI agents service or request a first conversation so we can get to know your company and see where we can help.