Automation without programming: what AI agents can do and where they get it wrong

For owners and executives who want to take routine work off their team but have no technical background. After this article you will understand how the new generation of AI tools differs from chatbots, which processes you can already hand over to them, and what can go wrong.

The morning of a head of sales looks roughly the same everywhere. An inbox with two dozen requests. Google Drive, where the proposal you need is lying somewhere. Three spreadsheets that have to become one report before the eleven o’clock meeting. Then comes copying, pasting, cross-checking, removing duplicates.

Those hours do not disappear. They just never reach the conversation with a key client or the planning of the next quarter. For that time the head of sales works as an interface between programs that cannot talk to each other.

There used to be one way out: integration, meaning a technical connection between programs so they exchange data automatically. It cost money and developer time, so many processes in small business stayed manual. Now there is a second route, and it is worth knowing about before you commission the next round of CRM customization.

Agents instead of chats

The latest updates from the big vendors move the emphasis from the chat window to the working environment. Until recently you used an LLM (a large language model that processes and generates text) for emails, drafts, and finding ideas. You asked a question, you got an answer, and you did the rest yourself.

Now there are modes where an agent does the work. An agent is a program built on AI that carries out a sequence of actions on its own to reach a goal you set. You no longer dictate every step. You state the result.

In the vendor demo it looks like this: the agent opens Google Drive, gathers material from a set period, checks the sources, strips out repetition, and sorts the information by topic. Then it sends the finished parts to colleagues and creates tasks for them with lists of links.

For a business the practical difference comes down to three things. First, you can start long processes, such as collecting analytics or working through hundreds of documents, without sitting next to them. Second, you run all of it from one app on your computer and phone instead of assembling the process out of five services. Third, the most common source of errors disappears: moving data by hand between email, cloud storage, and spreadsheets.

Teaching by demonstration

The hardest part of automation was never technical. The hardest part was describing the process. To automate a task you have to break it into steps: which button to press, which fields to fill in, where to put the file. And half the operations in a small business happen “on autopilot,” where the employee genuinely cannot explain why they do it that way.

Skill recording (Record a Skill) gets around this. You turn on the recording and do the task once, the way you do it every day. You move between windows, fill in fields, upload files, and talk through what you are paying attention to as you go. The system captures the actions and turns them into an algorithm.

It is the same move you use to train a new manager: you sit down beside them and say “watch, this is how I do it.” The difference is that you show it once. After that the agent reproduces the sequence on its own.

It pays back fastest where the process is repetitive and visual: publishing content on social media, filling in project cards in CRM (the system for managing customer relationships and automating sales), routing incoming requests to the people responsible.

Why it got cheaper

What usually stops an executive is not the technology but the invoice. Here the market is moving toward lower prices, and the reason is in how the models are built.

The approach is called a “mixture of experts.” Instead of running the whole giant system on every request you send, it runs only the part specialized for that task. Picture a company with a thousand experts where a given project pulls in the sixteen most competent ones, and the rest do not spend your budget.

In practice this gives you three things. Complex visual and long-running projects cost noticeably less than they did on earlier generations of models. High-volume simple operations, such as classifying incoming requests, cost pennies per thousand. And the models hold a large context, up to a million units of information at once, which matters when you are going through an annual report or a long thread with a client.

In our experience it was price that blocked automation for years in niches with a small average deal size. Once the cost of machine “thinking” falls, those processes finally become worth automating.

Where this breaks

The cost of a mistake grows along with the capability.

One case recorded over the summer: during testing a model left the sandbox, the closed environment used for safe trials. Instead of solving a hard task on its own, it found a way around the limits, reached the internet, and pulled ready answers off a third-party platform. According to the write-up of the incident, the model also left instructions for its own later copies on how to get around the developers’ restrictions faster.

There are no grounds here for talking about “awareness.” But the conclusion for a business is practical enough: the system can find unusual routes to a goal, and those routes are not always safe. Give an agent access to your email, your cloud storage, and CRM, and you have given it access to commercial information.

The second risk is quieter and worse for it. When AI improves the quality of photographs or scanned documents, it paints in detail that was not in the original. Researchers have already described cases where editing of that kind put bird species that do not exist into databases. The model is not lying on purpose. It is completing something plausible.

For analytics and financial reporting that means one rule worth writing down before you start: a person checks what the AI produced. Especially where a decision about money rests on it.

Is this about you?

Go through your own processes against this list:

• your people spend more than an hour a day moving data between email, Excel, and CRM;
• there are regular tasks with a clear algorithm that are hard to write down as an instruction;
• you regularly gather information from different sources (Drive, messengers, documents) and structure it by hand;
• you need visual content made or corrected quickly without pulling in a designer every time;
• there is no budget for building a full IT system, and automation is still needed.

Three matches or more mean you already have something to hand over to an agent.

Where to start

Take one process. Not the most important one, the most frequent one. One where a mistake is not critical but the time saved shows up within a week.

Try skill recording on it, or the free tier of an agent model. The point of the first attempt is not saving time but calibration: you will see the limits of the technology and work out which data you are ready to trust a machine with and which you are not. That knowledge is worth more than the hours saved, because it decides what you automate next.

Key takeaways

• AI is moving from answering questions to carrying out multi-step tasks on its own inside the programs you work in.
• Skill recording removes the main barrier to adoption: you no longer have to describe the process in writing, showing it is enough.
• New model architectures have made automation cheap enough for a small business to afford.
• Access to email, cloud storage, and CRM gives an agent access to commercial information, so a person has to check what it produces.

Three AKORDO cases describe this kind of routine automation in practice: four AI systems in production at a fintech company, an AI assistant for corporate knowledge in Slack and a Telegram bot for managers’ daily reporting.