ET RU EN

Blog

AI agents already work on their own: what small businesses can hand over to them right now

Published · Prepared with AI

Players sent in bug reports, and about an hour and a half later the game had five fixed versions, all shipped by the AI agents that run Matt Shumer's game. A chatbot would have answered such a report with a hint about what to fix, and a person would have done the rest. An AI agent takes on the whole task and sees it through to the result on its own. For small businesses in Estonia this is already a working tool: below are four tasks from sales, SMM and customer service that you can hand to an agent right now, and what is better kept for yourself.

How is an AI agent different from an AI chatbot?

A chatbot answers your question, and that is where its part ends: checking the answer, applying it and finishing the job is left to you. An agent receives the whole task. It breaks it into steps itself, carries them out, checks the outcome and keeps going until the result is ready. The person sets the task and signs off on the outcome.

Matt Shumer's agents worked exactly like this. He released a test version of a game maintained by his AI agents and asked players to send in bug reports. Judging by the timing of his two posts, the agents shipped five new versions in about an hour and a half and fixed every problem the players had reported.

The Perplexity Teammates agent, running on the Fable 5.2 model, took on a bigger task: working on its own over a few hours, it wrote a browser game - an endless zombie shooter. The agent built all the graphics, including the 3D graphics, in code, without any ready-made 3D models.

But both stories are about code and games. Does any of this reach the people who sell, run social media and answer customers?

Are agents already working beyond programming?

Programming was the first to show how fast this is happening. As Ethan Mollick observes, AI usually makes its way into work slowly, but programming has been the exception: less than a year after genuinely workable coding agents appeared, the software developer's profession is already being deeply reshaped.

Agents for everyday tasks are next. The company n8n has released an open-source personal AI agent (anyone can inspect it and run it on their own machine). According to the company itself, in October this agent completed 98% of its tasks running on the GPT-6.1 Sol model.

Sales are moving too. OpenAI has launched the beta of Businesses in ChatGPT: brands set up their own customer-facing AI agents in ChatGPT, get a verified profile and can see how often ChatGPT mentions them. Paid advertising can also bring shoppers to such an agent.

In Estonia, this shift is already visible in economic forecasts. According to an SEB forecast reported by ERR, the development of AI is forcing local companies to rethink how they make money. Even the IT sector, which carried the country's economy through the difficult years, has itself come under pressure from AI.

So for small businesses the question is now a practical one: which tasks can be handed to an agent today.

Which sales, SMM and customer service tasks can you hand to an agent right now?

Agents suit tasks with a clear input, a clear finished result and rules for checking that result. Here are four such tasks:

  • Sales - sorting enquiries. The agent collects enquiries from your website, email and messaging apps, asks the customer for anything missing (timing, scope, contact details) and hands you a ready summary card, flagged with which enquiry to tackle first.
  • Sales - a customer-facing agent inside ChatGPT. The Businesses in ChatGPT beta gives a brand its own agent and a verified profile and shows how often ChatGPT mentions you. For people who look for products and services through chat, this is a new door to your business.
  • SMM - a month of content. The agent puts together a content plan according to your rules, writes post drafts for each social network and queues them up. You approve or edit, and a post goes out only after your 'yes'. If you run social media for clients, the agent takes over the drafting, leaving you time for strategy and conversations with the client.
  • Customer service - routine questions and complaints. The agent answers frequent questions from your knowledge base (opening hours, order status, return terms), passes anything beyond it to a person and gathers complaints into a list of what to fix first - just as Shumer's agents worked through the players' reports.

What is better kept for yourself?

All four tasks share the same structure: the agent completes the repetitive part, and the final say stays with the person who runs it. Keep for yourself anything where a mistake is costly or where the decision rests on trust between people:

  • Money and commitments: discounts in negotiations, contracts, payments. The agent can prepare a draft; the signature and the consent are yours.
  • Difficult conversations: an emotional complaint, a conflict, turning down a regular customer. The agent can pull together the order history and suggest the facts; a person leads the conversation.
  • Anything that goes out in the business's name. At least for the first few weeks, every post and every reply to a customer goes through your approval until you see how well the agent keeps up the quality.
  • Deciding how to make money. The SEB forecast is precisely about business models needing to be rethought, and that work stays with the owner: the agent brings the data, you make the choice.

Where will an agent stumble, and what can you do about it?

The tasks listed above have four weak spots, and each has a workaround.

First: the impressive figures come from demonstrations and programming tasks. n8n's 98% is the company's own data, and the Shumer and Perplexity games were built in code. On your enquiries and posts the result may differ. What to do: run the agent on one task for two weeks alongside a person and compare how much it did on its own and how much had to be redone.

Second: the agent needs context. Without your rules, your frequent questions with answers and examples of posts that worked, it will produce a one-size-fits-all response. What to do: before connecting it, gather all of this into a single document - it will also serve as the agent's instructions.

Third: access. An agent that acts on its own works with your email, spreadsheets and social media, and even developers test how far it can reach. Perplexity ran attack tests on SPACE, its open sandbox for AI agents (a sandbox is a closed environment the agent should not be able to escape): the isolation held, but the team found and closed a network bypass that worked by spoofing DNS, the internet's address book. What to do: give the agent exactly the access the task requires, and a separate account wherever possible.

Fourth: Businesses in ChatGPT is still in beta, and its rules may change. What to do: treat it as an experiment and keep your main sales in the channels that work for you now.

To sum up: an AI agent takes on a whole task and sees it through, programmers are already reshaping their work around this, and for small businesses the first tasks lie in enquiries, content and replies to customers. Money, difficult conversations and the choice of business model stay with you. One step for this week: pick a task from the list that comes up every week for you, and describe it for the agent in three lines - what comes in, what result counts as finished and who checks it. If you would like to work out how an agent would fit your sales or customer service, discuss it with MOST's experts.

Sources

Questions and answers

How is an AI agent different from an ordinary AI chatbot?
A chatbot answers your question, and then the work is up to you. An agent takes on the whole task: it breaks it into steps, carries them out, checks the result and finishes the job. Your part is to set the task and sign off on the outcome.
Which tasks can a small business hand to an AI agent right now?
Those with a clear input and a clear finished result: sorting enquiries from your website and messaging apps, a content plan and post drafts for social media, answers to routine customer questions from your knowledge base, and gathering complaints into a list of fixes. Sales also has a new option - a brand's customer-facing agent inside ChatGPT, currently in beta.
What should you not delegate to an AI agent?
Money and commitments, difficult conversations with customers and the decision about how to make money. The agent will prepare a draft and gather the facts, but the signature and the choice stay with a person. In the early days, posts and outgoing replies are also best released after your own approval.
How reliably do AI agents see tasks through to the end?
The figures are already serious: n8n's personal agent completed 98% of its tasks in October, according to the company itself. That is the developer's own data, and on your tasks the result may differ. Check for yourself: give the agent one task for a couple of weeks alongside a person and compare how much had to be redone.
What is Businesses in ChatGPT?
It is an OpenAI beta for brands. A company sets up its own customer-facing AI agent in ChatGPT, gets a verified profile and can see how often ChatGPT mentions it. Paid advertising can also bring people to such an agent. Since it is still a beta, it makes sense to try it as an experiment alongside your main sales channels.
How is AI affecting business in Estonia?
According to an SEB forecast reported by ERR, the development of AI is forcing Estonian companies to rethink how they make money. Even the IT sector, which supported the economy in the difficult years, has come under pressure, and the country needs new sources of growth. For small businesses it is a reason to look at how much of their routine an agent can already take on.

← All blog posts