AI agents, chatbots and digital twins: what businesses can actually build today

A customer writes late in the evening because they need to move tomorrow’s appointment. Nobody is in the office. That customer does not care whether the company has AI; they care whether somebody can help now, without waiting until morning or repeating the same information several times. That is where the conversation about agents becomes useful, because the question is no longer whether a business can build a chatbot. It can. The harder question is how far the system should be allowed to go once it understands what the customer wants.

A simple chatbot might explain the rescheduling policy. An agent can identify the customer, check the booking system, offer suitable times and update the appointment once the customer confirms. OpenAI’s agent tooling has bundled capabilities such as search, file access and computer use into one API surface since March 2025. OpenAI, New tools for building agents The extra capability creates real value and changes the risk profile completely. A chatbot that gives a poor answer is frustrating. An agent that changes the wrong booking has created an operational problem somebody now has to fix.

More capable is not automatically better

Two positions here are both reasonable. One says the goal should be as much automation as possible: if the system can understand the request and perform the action, why keep a person in the middle? For routine work that argument is strong. Full automation can reduce waiting, cut administration and make service available outside office hours.

The other position becomes more persuasive as permissions grow. Once an agent can change business data, send external messages or commit the company to something, every additional permission raises the cost of a mistake. We generally prefer to start with less authority and add more once the workflow has proved itself. That is not because people should approve every trivial step forever; it is because the real exceptions usually appear only after a system meets real customers and real data.

Human involvement should therefore not be treated as a weakness in the design. With a customer complaint, an agent can read the message, collect the relevant details, check the history and prepare the case. If the conversation reaches compensation, legal responsibility or a relationship worth protecting, handing it to a person may be the right outcome. The automation has still done useful work. It simply knows where its responsibility ends.

The channel is only the front door

Businesses often start by deciding where the agent should live: the website, WhatsApp, Telegram, Instagram or another channel. Those choices matter commercially, but the harder question is what sits behind them. Meta’s Cloud API and Telegram’s Bot API both provide official ways for business messaging or bots to connect to backend systems. Meta, WhatsApp Cloud API Telegram, Bots

If a company builds a separate bot for every platform, each one can slowly develop its own instructions, knowledge and behaviour. The website assistant says one thing, WhatsApp says another, and a third channel is still using an old price list. That is an architecture problem wearing an AI costume. We prefer to treat the channel as the front door and the business logic as the house behind it, so customers arriving through different channels rely on the same approved information and the same rules about what the system may do. At ARKARA, the conversation itself is rarely the difficult part. The difficult part is deciding what the agent can see, which system is authoritative, what happens if an action fails halfway and when a person should take over.

When there is no API

Most good integrations use an API because APIs are designed for systems to communicate with each other. Plenty of smaller companies, however, still depend on older software with no useful API or on supplier portals that were never meant to be integrated. Computer-use agents open another route by working with the graphical interface itself: reading the screen, deciding what to do and issuing clicks or keystrokes.

That can be valuable when an employee spends hours every week copying information into a system the vendor will not expose. The limitation is reliability. Interfaces built for humans move, popups appear, sessions expire and layouts change. OpenAI’s guidance for computer use therefore recommends isolated environments, allow lists, step and cost limits, verification of the real outcome and keeping users in control of actions that are difficult to reverse. OpenAI, Computer use Where a stable API exists, we would normally prefer it. Computer use becomes most interesting where the alternative is continued manual work, not where a dependable integration already exists.

Digital twins are where this gets much bigger

Digital twins are usually introduced as a video feature: record yourself once, type a new script later and generate another video without going back in front of the camera. That is already useful. HeyGen, for example, requires consent for creating a video-based Digital Twin of a person. HeyGen, Digital Twin HeyGen, Avatar consent

For a smaller business, the first use cases are easy to understand. An owner may repeat the same onboarding explanation many times a year, or a specialist may need to update training material in several languages. In those cases, the twin is primarily a production tool: the human decides what will be said and the twin presents it. We think that is the easier business case today because it improves an existing process without giving the digital identity much independent authority.

The more ambitious version is different. HeyGen’s LiveAvatar is designed for real-time two-way conversations, with an AI system behind the avatar responding live rather than presenting a fixed recording. HeyGen, What is a LiveAvatar That could allow employees to ask questions of a digital version of a trainer, or customers to speak with a product specialist’s digital representative in several languages. For businesses where one person’s knowledge or personality is an important part of the experience, the attraction is obvious.

The same quality that makes an interactive twin powerful also makes it sensitive. People attach a face and a voice to the person they belong to. If the twin makes a promise, gives wrong information or says something the real person would never say, customers may not separate the software error from the person’s reputation. That is why we would treat a content twin and an interactive twin as different categories. An interactive twin needs defined knowledge, explicit limits, controlled access and a record of what it says and does. The person whose likeness is used should know where it can appear and what authority it carries, and users should know when they are dealing with an AI-generated representative.

That last point is not only good practice in Europe. Regulation (EU) 2024/1689 requires systems intended to interact directly with people to inform them that they are interacting with AI unless that is obvious, and the Regulation applies generally from 2 August 2026. EUR-Lex, Regulation (EU) 2024/1689, Article 50 and Article 113 The article’s publication date is two days earlier, so for a Swedish business this is an immediate requirement rather than a distant policy discussion.

If a company told us it wanted an AI agent or digital twin, our first question would still be what problem its employees or customers are dealing with today. Hours spent answering the same booking questions may justify an agent connected to the booking system. A specialist re-recording the same training may have a clear case for a digital twin. If customer enquiries arrive across five channels and nobody has a coherent view of them, the first problem may be those channels rather than the absence of AI. The best system is rarely the one that can do the most. It is the one the company is still comfortable trusting after the novelty has worn off.