Lazy and curious — my baseline stack

"I choose a lazy person to do a hard job. Because a lazy person will find an easy way to do it" © Bill Gates

I have spent my entire working life in the real sector — engineering, manufacturing, design, construction. And all the while my companies have been at the cutting edge of information technology. Because I am a very lazy person, and I have always been too lazy to do by hand what could be simplified or automated.

We moved to hybrid and remote formats back in 2012 — eight years before covid made everyone else do it. With AI it is exactly the same story: we have used it since it appeared, at one point those were the first chatbots, and in this article I will describe the setup as it stands on the date of publication.

AI has always been simply a working tool. For a while the talk of a revolution amused me, but what has happened to the technology recently I would not even call a revolution — we have crossed the event horizon, time has compressed, and there is no way back. So here I will describe how my AI setup works right now.

This is not a manual on how it should be done. Your context is different, and much of what works for me may not apply to you. I do not claim the final word — I am simply describing how things are on my side.

And the main disclaimer: this article was in all likelihood already out of date the moment it was published, but it should still be of some use a while later.

What I have AItomated

Every flesh-and-blood employee has digital employees of their own

In short — all the digital routine. Let me go through the main positions so it is clear what level of adoption we are talking about.

Agendas and meeting minutes. I have an agent that walks through the team's working chats for the week, collects topics, and puts together an agenda for the next meeting. After the meeting it takes the transcript of the recording and issues the minutes with decisions, assignments and deadlines.

All correspondence. Including the layout in Word — with corporate headers and footers, company details, seals, signatures. We used to have assistants for this. Now we don't.

Multi-page contracts. Layout to a strict template, with continuous clause numbering, references to annexes, and a check that there are no contradictions between sections.

Presentations and reports. Internal and external, in corporate style. We used to draw every presentation in Figma; now we only occasionally finish something off there when it needs "decorating".

Plus a long tail of small things I no longer even count — format conversion, data exports, standard certificates, backlogs, documentation templates. And a number of know-hows that deserve their own articles later.

The stack

Autonomous agents using well-honed skills are digital employees

We work on the principle of "a tool for every task". Do not entertain the illusion that one subscription will close everything.

Right now I hold subscriptions to all three leading models — ChatGPT, Gemini, Claude. There are subscriptions to aggregators of generative photo and video models, plus a number of local narrow-purpose ones — for data de-personalisation, for example. The entry subscription of $20 a month is roughly the same everywhere, but under active use you hit the limits quickly, and depending on which model is currently "in its prime" I take the extended $100–200 subscription with that one. ChatGPT is on a corporate subscription with a shared company billing account for PRO requests, which burn separate tokens.

At the same time we almost never go into the models' web chats — we work through an IDE, in my case Microsoft VS Code with extensions for every model. You can test this mode of working through Google's Antigravity, it is much the same; we gave that to the participants of the mastermind as a way into the topic. Even in its free version Google gives you tokens from top-tier models to try.

The IDE has access to the shared drive where we keep the company knowledge base in MD format: internal regulations, standards, standard templates, product descriptions, process descriptions and a lot of other good stuff. Why MD — because at our current volume RAG is not needed yet, though we are moving in that direction. For now we simply run every Word, PDF or Excel document through a home-made local converter into MD and extend the base every day.

On top of the knowledge base sit the skills. Simplified, a skill is a long structured prompt describing who this digital employee is, which folder it looks into for data, and in what format it returns the result. When I tell an agent "issue such-and-such document on such-and-such topic", it works out which skill to invoke, follows the document map into the right folder, gathers the context, and returns the result in the required form.

When an intern joins you — a capable person, but without much experience — naturally you will check their work, explain the context, give them a job description. With a model it is exactly the same. And the job description is your skill. It is the script an employer once wrote for an employee. If you are able to automate your own job description — at least its formal part — you launch an agent and get on with the substantive work yourself. Once the skills are honed and the agents are configured, we automate their autonomy in launching and completing the task.

Three skills that work for me every day

We have dropped lawyers altogether — none in-house, none outsourced. No lawyer managed to handle our tasks at even 30% of the AItomated level

The lawyer skill. We trained it on the relevant regulatory base and our internal documentation. It analyses tenders, contracts, procurement rules. It produces opinions with direct references to clauses of the regulations, contradictions between norms, and specific recommendations. In terms of volume and depth of analysis, I would not have got that result from an in-house lawyer at all. The main strength of the skill is in the analysis of combinations of norms: which clauses together permit something, which prohibit it, and how a workable position emerges from that. A human physically cannot do this work in the same time — they simply do not hold that much context in their head at once.

The procurement skill. It analyses incoming tenders to see whether our goods and services are hidden there under different codes and wordings. This is a typical market story: the buyer describes a position through adjacent codes or buries it inside the volumes of a bundled service so that a particular supplier fits it — and in the live search results you simply will not see your product there. The skill collects all tenders for the period, reads the technical specifications, matches them against our product line, and finds the hidden volumes.

The agenda and minutes skill. Already described — it walks the chats, collects topics, prepares the agenda before the meeting; afterwards it takes the transcript and issues the minutes. This is probably the most mundane of them, but also the most regular saving of time.

Where AItomation has not happened yet

Alongside AItomation you have to change the internal regulations

Our accounting works with very specific raw data: on top of the standard stock counts, ledger entries and foreign trade operations we have SEZ residency, special investment contracts and a pile of attendant nuances. We tried to automate part of the work through skills layered over the existing accounting system. It came out very cumbersome, with a heap of scripts. On raw, unstructured data an agent cannot work autonomously: too much context, too many exceptions, too many places where its context window overflows while it is still reading the data.

We are now writing our own AI-native ERP system https://staiter.io/ — built from the outset around agentic work by models. When the ERP matures, accounting will be AItomated too. We do not even picture yet what it will look like; perhaps only one person will remain on primary documents, and that's it — that is still to be thought through.

What I do not delegate

One day AI leaders will appear, assemble teams of flesh-and-blood agents and take the market, but nothing like that is on the horizon right now

Everything to do with live communication — interviews and hiring, working through difficult situations, conflict, dismissal, support after bad news, gratitude. Wherever empathy and emotional intelligence are required, human contact always beats AI.

In my experience, the strongest teams hold together because the leader genuinely shows up in the work. Through attention, through presence, through the fact that what hurts the company hurts you personally.

The same goes for coaching. I practise executive coaching to ICF standards; it is the work of two people in partnership, and AI has no place there.

Where to start

Talk to people in person, and let AI handle everything else

Do not rely on a single subscription. Take subscriptions to all the top models plus an IDE with agents. Inside any one tool there are always limitations that another tool's capabilities will cover.

Knowledge base first, skills second. Before writing your first agent, gather your company's basic documents in one space in MD format: regulations, standards, templates.

Write the skills with an agent. Open the IDE, ask it to write an instruction for creating skills, then use that instruction to write the skill you need for the task.

Count the hours freed. There is one metric — how many hours a week the team got back, and where those hours went.

780 hours in 8 hours, and other cases

I once fed a strong reasoning model the results I had produced in a single working day and asked it to estimate the team hours the same work would take by hand. I got 780 hours — a hundred working days, or six months with holidays.

And with the AI lawyer we saved forty thousand dollars on customs clearance of equipment, although every living specialist said it was impossible.

The digital products we are shipping right now with our very small team, which does not include a single programmer, are valued by reasoning models at roughly two million dollars of development cost under the old rules.

And all of this thanks to two qualities of mine — laziness and curiosity. Laziness we covered above; curiosity is your guide into the world of AItomation.

Just ask the models how your particular task could be implemented, and together you will find the solution

If you are too lazy even to look for the solution yourself, we can do it together — see the formats I work in.