AI agent for business: an off-the-shelf tool or one built for you?

In September 2026, 30% of UK businesses were using at least one AI technology. Three years earlier, when the question was first asked, the figure was 9%, according to the Business Insights and Conditions Survey run by the Office for National Statistics, to which nearly 10,000 businesses responded.

A business owner considering an AI agent first wonders whether to choose an off-the-shelf tool, such as ChatGPT Enterprise, Microsoft 365 Copilot or Mistral AI’s Le Chat, or have an agent built into the company’s software.

The answer depends on one thing: does the task need your data and your rules, the ones held in your software? If not, an off-the-shelf tool is enough. If so, the agent is built into the software that holds that data and applies those rules.

AI agent, AI assistant, chatbot: what is the difference?

The three words are often used interchangeably. Yet they describe three different uses.

A chatbot answers questions in a chat window, for example from visitors to a website or from customers who contact customer service.

An AI assistant helps a person with their work: it drafts, summarises, translates, looks up information. The person reads the answer and decides what to do with it.

An AI agent acts. Erik Schluntz and Barry Zhang, engineers at Anthropic, the company behind Claude, give a simple definition: in an agent, the AI model itself decides which steps to take and which tools to use to complete a task. When those steps are fixed in advance in the code, they call it a workflow.

For example, an agent that reads a request received by email, finds the order in your software and drafts the reply, which a person approves before it is sent.

So an agent can create a customer record, send an email or update an order. To do so, it needs access to your software, and limits on what it is allowed to do there.

The same authors recommend starting with the simplest solution, even if that means not building an agent at all. In their view, an agent costs more and responds more slowly; in return, it gives better results on tasks that call for flexibility.

An offer sold as an “AI agent” may therefore be a chatbot, an assistant, a workflow or an agent. Before comparing two offers, ask the seller what the tool does on its own, and what a person approves.

Off-the-shelf AI agent: when it is enough

An off-the-shelf tool can be used from day one, with no development.

It is enough when a person still does the task: drafting an email, summarising meeting notes, finding information in a few documents you give it. The person rereads the answer, then copies it into your software themselves if needed.

Choosing one tool for the whole company also limits shadow AI: when the company chooses none, an employee may paste a document into an assistant nobody chose. With a chosen tool, the company knows who it is entrusting its documents to.

When the AI agent has to be built into your software

An agent is built into your software when the task takes its data from it, or writes its result to it: your customers, your orders, your stock, your files.

Custom software becomes necessary when the process applies your business rules, when several people are involved with different permissions, or when the information has to be checked before it enters your systems. At DEVEDANOS, we develop it for a fixed monthly budget, and the agent handles one specific task in it.

A built-in agent works through your software: it is subject to the same permissions as the person using it. It answers from what that person is allowed to see, and acts only where they are allowed to act.

The software then checks the agent’s output before saving it. Where an error is costly, a person reviews it before it is saved.

An assistant that answers from your documents also works through the software. For each question, it looks in your documents for the relevant passages and passes them to the model: this is the technique of connecting your documents to a model, known by the acronym RAG.

Under each answer, the software can show the documents it draws on, so that the person can check. When your documents do not contain the answer, the assistant must say so, and hand over to a person.

For a higher-education group, we worked on integrating AI into its platform for work-study placements, open only to its students. The software reads the image of the CV and extracts the student’s profile from it, in a format approved with the group’s staff. This is not an agent: the steps are fixed in the code. Automatic CV analysis cut manual data entry time by 80%, above all for the group’s staff.

What you keep when the provider or the model changes

An AI agent rests on a vendor’s model, and vendors retire their models. Anthropic, for example, commits to giving at least 60 days’ notice before retiring a public model. After that, requests to that model fail.

With an off-the-shelf tool, the tool’s vendor chooses the model and changes it. The answer to the same question can therefore change without you having decided it. What you have built in the tool, your instructions, the documents you uploaded, the assistants you set up, stays with that vendor. Switching tools usually means rebuilding all of it in the new one.

With an agent built into your software, your instructions, your documents, their index and the code are kept with you, not with the model’s vendor. A set of test questions, whose right answers are known, comes with the software. It is replayed each time the model changes, so you know, with figures, whether the new model does as well. A second model, from another vendor, takes the same test: the fallback has already been tested the day the first one disappears.

How much does an AI agent cost?

An off-the-shelf tool is paid for every month, usually per user. Spending follows the number of people who use it.

An agent built into your software first costs its development. After that, each task uses a vendor’s model, which charges by the volume of text processed. Spending then follows the number of tasks, not the number of users.

An agent queries the model at each step it decides on. A task handed to an agent therefore takes several calls to the model, where a simple answer takes one. The model’s vendor sets the price of each call, and can change it.

DEVEDANOS: AI in your software, for a fixed monthly budget

When the task needs your data and your rules, we integrate AI into the custom software we develop for you, as part of our offer. We build an agent when the task requires the model to decide its own steps, and a workflow when that is enough.

The model used is named in writing: the vendor, the name, the version, where it is hosted.

Invoicing is monthly and fixed, with no surcharge for licences or number of users. This budget covers development. Using the AI model is a separate expense, which depends on the volume processed. We track it every month.

Your application is usable within a month. After that, we adjust it as your business evolves, for as long as you wish. The source code, infrastructure, deployment pipeline and documentation are your intellectual property once the first monthly invoice is paid.

Our page on business process automation with AI sets out what we automate, and how.

Further reading

Before choosing a tool or an agent, start from a specific task: the one a person redoes by hand, several times a day or a week.

Our guide covers the questions to ask before signing a quote for the development of your custom tool, on a fixed monthly budget.

Tell us about the task you still do by hand today. We will tell you what software can take over, and whether AI has a place in that work. Shall we talk?