A New Era of Computer Adoption: AI

We are at the beginning of a new cycle of computers entering everyday life:

  • in the 90s and 00s, computers and automated workstations appeared at every workplace
  • then came mobile phones and the corresponding mobile apps

That’s from the hardware perspective; from the business software perspective:

  • automation projects with custom development
  • SaaS for more standard cases

Naturally, automation projects often lead to staff reductions and/or lower requirements for employees.

By now, most companies have already been automated. All that remains is changing existing systems to follow changes in the business.

But then AI showed up. And it turned out that now even more business processes can be automated.

For example, automatically or semi-automatically answering questions and reviews on marketplaces. Unlike development (where it’s easy to burn 100 million tokens in a few hours), answering user questions costs pennies (especially compared to the cost of any living human).

Also, AI vendors try to sell their products for tokens or credits. And AI-based products are likewise sold for tokens (whether you like it or not).

In terms of hardware, AI brings few changes at the company level: you’ll most likely need to use a cloud solution from a major provider. In time, companies will get their own servers, but first everything has to settle down (AI manufacturers must hit their growth ceiling, and hardware manufacturers must catch up with them). On a global level, AI data centers have a serious impact, but that’s not the level of individual companies for now.

In short:

  • new automation projects with custom development
  • new SaaS, but now with pay-per-use pricing (tokens)
  • new hardware (rented for now, owned later)

What to Automate

It’s important to understand that, as before, the principles of automation stay the same:

  • high savings, low implementation costs, high probability of successful delivery, and low cost of the agent’s mistaken actions – that’s how we pick priorities
  • it’s better if the business itself understands which processes can be automated with AI
  • the processes must be in order first: automated chaos will only kill the company faster
  • you need a domain expert: without one you’ll have to redo things more times, and success is questionable

How to Automate

So far there are few best practices – mostly just directions.

There are 3 degrees of automation:

  • plain chats (with no special configuration), Cowork, and always-on “telegram” agents
  • the same, but with your own and third-party skills using MCP and console utilities (for stable results where no thinking is needed, just execution)
  • workflow systems, where several agents are launched in some sequence

All of this was for personal use. Within a company, the same things are done centrally. Currently popular frameworks: langchain, langgraph, crewai, autogen.

AI should only be added where it’s needed; everything else is better written as regular software (more reliable and cheaper – no tokens every time).

If something can be done without a workflow, through a skill, then it’s better done through a skill: faster, simpler, and cheaper.

The UI can be:

  • web – as before
  • desktop – as before
  • API – as before
  • mobile app – as before
  • chatbot – new; there used to be few of these because they communicated poorly
  • MCP, CLI, skill – for reuse in other AIs

AI Company

An AI company is about transferring control from the individual employee (it’s unclear what and how they use AI) to the company (the company decides which AIs with which skills employees will use – an analogy with software on work computers – and which processes get automated through AI – an analogy with current automation).

AI as an Employee in the Company’s Org Structure

When talking to AI, it’s right to think of it as something that thinks, because that makes it easier to work with.

It’s easy to imagine that individual AI runs are your department: you’re like a manager handing out tasks, checking results, training your employees, and organizing their work.

If you’re not a domain expert, you bring in external expertise via SaaS or custom agents built by experts. Only in that case it’s much harder to verify the quality of the result. This usually works when average or poor quality is good enough.

I can easily imagine a SaaS employee. There are already SaaS programmers, designers, musicians, and writers. Others, more specialized, will appear too.

AI can well distribute tasks within a department. People even like it (supposed impartiality). And there were hints of this before AI. At the same time, the other functions of a department head are better moved into software and the individual employee’s AI assistants (without calling that a department head). In other words, the function breaks apart into separate components.

It would be fun (like in The Sims) to add such processes as “people” to the org chart. Some will like it. Some will think it’s entertainment. But what if you look at it from another angle? An AI agent can be replaced by a human, and vice versa a human by an AI agent. In that case, it makes sense to add AI agents to the org chart. For now, it’s an open question whether companies will actually make decisions this way in practice.

What Are AI’s Limitations

Lack of Information

The main limitation: lack of context. For example, an AI agent is unlikely to learn that the road in front of your little shop has been dug up (and there will be fewer customers for several days). Or what you agreed on with someone in the hallway.

In each specific case, especially after the fact, you can figure out how to feed this information in. But it’s hard to do in advance. And that means losses from wrong decisions.

Security

There’s no such thing as AI having two inputs: a regular prompt and data from an untrusted source. Everything goes into one (to be precise, there’s a stricter system prompt (and not even for all model types), but it doesn’t help much either). And if everything goes into one, then data fetched from some website or received from some person may contain destructive instructions: leak confidential data, delete everything, sell everything for a single penny. You can easily find such stories in the news already.

There is no good protection (at the level of regular software). There are some approaches that reduce the probability of a breach:

  • typed data (especially without arbitrary strings) for exchanging data between agents
  • minimizing and typing agents’ permissions. In particular, if you’re talking to a specific authenticated customer, external services should receive their user id so they return information only about them
  • checking the prompt through a separate AI agent (like auto mode in Claude Code)
  • AI models themselves are trained not to obey every prompt

In general, as always, security is inconvenient: development delays and extra operating costs.

Cost

Compared to regular software and regular servers, AI is very expensive. That’s why it’s sold for tokens. You need to develop an approach so you don’t accidentally spend too much. Accordingly, strict accounting and cost control are needed.

This also includes the sudden shutdown of AI model providers or a significant increase in their prices. Such problems exist and will continue to exist. In the end, you need many providers for reliability. How do you survive this if AI agents play a key role in the company? It’s an open question, roughly like what happened with the blocking of foreign clouds. That is: downtime, migration costs, and potentially spending on your own hardware with open source models.

Who Is Responsible for a Problem

If you’re the “head of a virtual department,” then you are. But if your department head is an AI, it’s unclear who.

To be precise, the search goes in all directions: the human executor is responsible, and if they weren’t given the task, then it’s the first person up the structure who should be double-checking (and building a system of automatic re-checking).

Or as with market leaders: the best available solution was used, and nobody is at fault – the risk just played out.

No Guaranteed Result

A lot of people don’t realize this. Even if everything usually works fine, sometimes the output is just nonsense. If a human is watching, they’ll simply rerun it and there’s no problem. But if the system is automatic, the next layer somehow has to figure that out.