AI Agents for Entrepreneurs - First Lessons
Andy Avramenko 21 November 2026 12 min read 0 views
So far it seems that automating existing roles/processes gives a better chance of a successful agent rollout than these mythical new wishes.
Hi! I'm Andy, one of the two founders of Combobulating. Until September this year I had only been robotising my own business, and when I had practically run out of ideas of what to do for myself, I decided to try rolling this experience out to my peers. Three months is not a very long time, but it feels like time has compressed a lot and some conclusions can already be drawn. This reflection is largely for myself, so that the next projects are more successful:
1. Who needs AI agents
I've worked in many places, including decent outfits like IBM, I've always written quite a lot on my personal social media, plus I have a work Telegram channel with 18 thousand people, so I was a bit surprised how few people I know wrote to me since I raised the "I implement AI agents" flag high. One could assume that I'm just not a very pleasant person (which is true), but the previous big topic I worked on didn't stop anyone, and questions about cross-border transfers and all sorts of residence permits helped me feel important and useful to society for a long time.
The main practical interest so far comes from entrepreneurs who believe that the bright future belongs to AI. Usually they have a lot of experience with chatbots like ChatGPT / Claude, already pay for a subscription to them, follow the AI market closely and want to start implementing it themselves, but don't quite understand where to start.
The main pains they want to solve with AI agents:
- saving their own time (answering their staff's questions, etc.)
- being a better version of a business owner (asking hired management the right questions at the right time)
- replacing individual roles on the payroll with AI agents.
To be honest, alongside my universe described above, there are also processes of routine business automation with LLMs, where nobody really replaces anybody. But they don't reach us, so I know/understand them little.
2. Trust and security
One of the key questions that comes up before the go-ahead to implement AI agents is security in the broad sense of the word. There's a lot of worry that Anthropic or other frontier labs will read everything, steal it and turn you in to the police. My main counter-argument for everyone is - if not Anthropic, then definitely someone else. Unlike many others, Anthropic already has top LLM models, you're unlikely to teach them some clever business tricks or show them some alpha in investing that they can't reach on their own. Whereas the likes of telegram, dropbox, microsoft, apple, telecom providers and even your smart TV - they would all be very motivated to learn something new.
Besides trusting the Anthropics, there's also the general question of security - an agent always has sensitive documents and access you worry about. This one is simpler - nobody will sort out all the potential security risks on its own better than claude code, you just have to ask, or better, add it to the schedule.
3. Connecting to systems - the biggest friction
Choosing a server, learning to connect to it, integrating with all sorts of google drive, corporate ERP and other systems, and also figuring out where and how it's most convenient to talk to the agent - that's probably the main value we bring to most clients so far. Well, except maybe for a general confidence along the lines of "if agents haven't broken anything for us in a year, they probably shouldn't for you either".
Many things seem obvious in theory (any chatgpt will tell you what to do), but in practice difficulties come up:
- the agent eats up the whole context in telegram, and in the end it's not only expensive but also, surprisingly, leads to mistakes or the agent getting lazy
- the agent refuses to integrate with the broker's API and says that you, meatbag, have to enter the passwords everywhere and place the orders yourself
- the video edit looks terrible, because the agent uses whatever it found on the computer, someone forgot to install the tools/libraries for it and didn't tell it "let's install them before we start".
4. Boxed solution or the IKEA effect
It's amazing how many successful, accomplished people around have found a hobby in building AI agents. Day after day they find a new process in their business or personal affairs that can be automated, and rejoice like children at how cool and stable it works. The man has had a whole huge chain of stores for years, and he vibe-codes at night like a kid. I'm sincerely happy for them, and examples like these help me feel I'm not the only one who's lost his mind.
There's a parallel universe where, on the contrary, people want to take ready-made settings https://combobulating.ai/en/system and adapt them to themselves and their processes. I didn't really believe in this business model, first of all because I tend to do everything myself wherever possible. But Slava's idea has clearly taken off. By the way, it's the cheapest way right now to get a working result, because Slava babysits every client, installs everything for everyone himself and has generally gone into perfectionism instead of laughing merrily with me at unemployed programmers on youtube https://www.youtube.com/@combobulating-ai-rus.
Each approach has its advantages, but what is visible right from the doorstep - these universes don't intersect. A person who set up their agent on their own and solved the first business tasks with it will never replace their style and approach with someone else's settings, even if those follow the most advanced practices of Anthropic itself. I'm like that myself, what can I say.
5. About 9 women carrying one baby
One of the biggest tasks, once the agent already has all the access and the rest of the tooling - is how to work with it. I mean, even if you've solved the technical part - we'll talk by email, ok fine, when should it write to you on its own, proactively? about what?
I went through the whole evolution - I ran prompts and watched how it worked. I ran separate agents on a schedule at night and read each agent's diary in the morning. I built a "meta-agent" that summarises all the diaries and prioritises my attention. Then we started simply giving me 5 tasks in the morning, which I answered in a text file. The first time I went on holiday without my laptop, I was thrilled. Now we've switched almost entirely to voice...
I thought my experience could be turned into "do this, this and this and everything will be fine", but the whole beauty of agents is that you can do things differently. I see the first clients going through a similar evolution anyway. Add the agent to a telegram chat with your employees, or have it answer in your own name? Answer always, or only an hour after you haven't answered yourself? Lots of nuances that I personally can't decide for someone else.
6. Automating existing roles is more useful than the mythical business debt
Programmers have the notion of tech debt, when some code really ought to be rewritten but there's no time because urgent real tasks keep pouring in. In the same way, every entrepreneur has areas there's no time and no one to deal with. For example, properly analysing the market and following competitors. It seems like a good idea, but in practice you do it at random or after a client said "it's cheaper at Joe's".
So far it seems that automating existing roles/processes gives a better chance of a successful agent rollout than these mythical new wishes. In existing processes there is always an expert, often the entrepreneur himself. He can easily tell the agent where to get which data, what counts as a good result, etc. You can also look it up in old artifacts - documents, correspondence.
Bring me into any bank of america today, and within a year I'll cut 80% of the staff without hurting revenue! I'm ready to sign that on paper. I'm hyper-confident in the result. At the same time, tell me to start making tiktoks for my business right now - well, who knows if I'd get anything substantial out of it, I don't understand it, and with every day I'd doubt it more and more and shift my focus back to proven mechanisms.
7. Claude really loves rolling out other Claudes
The next cheap product will probably be a Claude that helps other entrepreneurs roll out Claude at their place. It watches the work of third-party agents very subtly, draws value sheets perfectly (what useful things the agent has already done for the owner), adapts the size of the context window and replicates the practices we already work with at home. It also forms best practices perfectly - this client had such and such a difficulty, let's do it differently here from the start.
Traditionally, a clumsy, faceless approach works badly in sales. Everyone has to look at each other, sniff each other out, it can't be done by some magical automated service or an AI avatar, but if you set aside the imperfection of human nature - my Claude could roll out other Claudes much better than me. That's probably how it will be soon, we'll be calling each other and discussing trends and the future, while Claude does the main work in parallel and answers the practical questions.
8. Quick wins vs loss of trust
For one client we were waiting for access to their ERP. To fill the time, I made a simple chatbot that can find things about the client's business from public sources. I didn't set up the system prompt / hooks / context management - I put it off until we had access. In the end the agent got an important fact about the company wrong, one it had right at hand, it was simply too lazy to double-check. It took another week to convince them that Claude doesn't hallucinate. I'm not even sure the trust has been fully restored.
9. Cutting staff vs giving people tools
Giving employees access to all kinds of copilots and implementing AI agents = two completely different projects. The first leads to higher costs in the hope of future productivity. The second - leads to an insane cut in costs, and also to the hope of future productivity. The first more or less works for programmers, but with rather modest results, maybe +20-30% by the estimate of people I know whom I've talked to. The second cuts costs tenfold on average.
Several entrepreneurs I know got disappointed in the first approach but aren't ready for the second. They build parallel teams - they take a couple of people who do the existing tasks many times faster using AI agents. Not quite our path - we're for an autonomous role of agents in the organisation, not a thing to speed employees up, but a serious wake-up call for hired employees who smugly repeat "AI won't replace us".
10. State of the industry
These were a very interesting 3 months, largely because I had spent a long time away from working directly with people and had gone a bit wild. It's nice again to meet new people and find new things to talk about with those I've known for a long time. Thank you all for your trust, I sincerely hope this is all not in vain and your life will change into bc/ac (before anthropic / after anthropic).
My main surprise and disappointment so far is the insanely large gap between what others do and what I do. If I, just during this project, have switched completely to a single-terminal mode that autonomously manages other agents on several servers / several accounts, have largely switched to voice with a personal wearable device, have learned to manage the pace of proactivity in the agents' work depending on tokens/approved features and other circumstances, then most people who have been in this industry for a relatively long time are doing some kind of benchmarking - "tried the new Neon - really cool, but gpt-6.2 - a nightmare".
You'd think I should be happy about this, but honestly, I'd like to pick up something interesting from others myself too. The FOMO from the agents' first success has passed, of course, but it still feels like there's some magick out there that can be borrowed so that the prompt "make me rich fast" finally works :)
On the bright side, since we mostly go to all the calls together, I've picked up a lot of ideological things from Slava that used to be hard for me to put into words/accept - I've stopped doing anything myself except writing texts, like this one. In one of the mailings I got the time zones wrong, and Slava is right - an agent would never make that mistake, you never need to do anything yourself.
Was this article worth your time?
Read next
Cortisol
What AI agents can borrow from living nature besides ants. By default, an agent needs nothing. The job of a smart agent is to synchronize rhythms.
Always-on agents - agents that work all the time
Anthropic published an article on building always-on agents. All good, except the definition. A lot of what I have is built around one Claude Code session.
When the job is too big for your Claude Code server
Some builds need more memory than your machine has. Now Claude Code can borrow a bigger machine for the job: 3 cores and 6 GB, deleted when it is done.
Give Claude a computer of its own
Your own Linux machine with Claude Code on it, working around the clock while your laptop is shut. From $3.49 a month.