Almost every business has people who spend hours at their computers every day doing the same things: moving data from one system to another, answering the same customer questions, compiling reports, processing orders, corresponding with suppliers. The owner can see the team is overloaded, but hiring another qualified manager is difficult and expensive.
I want to show why this work is already largely automatable today, what it looks like in practice, and where to start.
Any job has repeatable steps
Even work that requires thinking is made up of repetitive actions. A sales manager asks customers for the same parameters every time, looks up the product in the database, checks availability, prepares a quote, and enters the deal into the CRM. A person makes the decision, but most of the steps around that decision are the same.
And where there is repetition, a process can be automated. In the past, this required rigid scripts that broke at the slightest change. Modern AI agents work differently: they read text, understand context, use tools, and can carry a process from start to finish, stopping where a human decision is needed.
Agents can use a computer
One of the main reasons modern agents are so productive is that they can work with a browser and control a computer. They aren't limited to running individual scripts. An agent can open a website, log in to a system, click the right buttons, fill in a form, move data between tabs, download a file, and check the result, just as a person would.
For a business, this means you don't need to rebuild existing processes or change the tools you're used to. The agent integrates with what you already use: CRM, accounting system, email, spreadsheets, supplier portals. Even if a system has no convenient API, the agent can work with it through the regular interface.
My experience: from 1–2 to 5–8 tasks a day
I'm a Senior Software Engineer with over 9 years of experience, currently working at Apple as an engineer for a contracting company. Most of my workday was made up of bugs: reading the task, finding the cause among dozens of applications, reproducing the problem, fixing it, opening a pull request, waiting for checks, responding to comments, and handing the task off to testing. Plus constant threads in the work chat.
Over several months, I built what I call an agent harness around this process: a system based on Claude Code that describes exactly how I work. It consists of about 60 written instructions (skills) for the agent, connections to the task tracker, work chat, browser, and CI, background scripts for routine steps, and a dashboard that shows the state of each task.
Now it looks like this. I start a task with a single command. The agent reads it along with the screenshots and video recordings, finds the cause in the code, reproduces the problem live in a browser, and prepares a report with evidence. I say "go." The agent makes the fix, verifies the result itself in the browser with screenshots, opens a pull request, monitors the checks and reviewers' comments, and after the merge returns the task to the tester. It only prepares messages to colleagues as drafts: I always send them myself.
The result: my productivity has grown from roughly 1–2 to 5–8 tasks a day, depending on complexity, because several such processes can run in parallel. The tools keep improving, so this is most likely not the limit.
The most important takeaway for me wasn't technical.
The bottleneck was never the writing of code itself. It was all the work around it.
And this applies to far more than software development.
The same thing works in any office
If you replace "bug" with "customer request" and "pull request" with "commercial proposal," you get the process of almost any company. Here are examples of computer-based work that can be largely automated:
- first contact with a customer: collecting the request, clarifying details, answering typical questions;
- order processing: checking availability, placing the order, entering it into the accounting system;
- working with data: transferring it between spreadsheets, the CRM, and accounting, finding errors;
- supplier communication and logistics: requests, confirmations, status tracking;
- reports: daily and weekly summaries that someone is currently compiling by hand.
One of my clients sells tires. His main problem will be familiar to many: there aren't enough managers, and qualified people are scarce on the market, so some customers don't get a timely reply. We don't plan to replace the managers. Our goal is to take the routine off their plates: collecting the request, selecting tires by parameters, checking availability, preparing the quote. Then the same manager can handle more customers, and their time goes to conversation and decisions rather than copying data.
A separate area that is developing very fast right now is conversational agents: an agent you can talk to by voice and, in some solutions, also see as an interactive avatar. Such an agent is well suited to the first line: greeting the customer, collecting information, answering typical questions, and handing a person an already prepared request. When you can see and hear a system, people accept it more easily and better understand what it is doing. These interactive agents are what we work with at OPV Systems.
Hybrid teams are appearing more and more often, where AI agents fill some roles and people fill others, and they work in tandem. In such a team, the agent doesn't just follow instructions but takes part in the process as a colleague: it prepares, checks, reminds, and hands work on.
What automation looks like in practice
Every business is unique, and there is no ready-made, out-of-the-box solution. But the approach is consistent:
- Look at how the work is done now. It works best when a person simply shows their ordinary day: what they open, what they copy, who they write to, where they wait.
- Break the process into steps and find the ones that repeat most often and take the most time.
- Automate one step at a time. Instead of a big system delivered after six months, it's better to add small "bricks" and show results right away. That way progress is visible from the first weeks, not at the end of the project.
- Keep a person in the process. The tool should show what stage it is at, what it is doing, and what result it is delivering. The employee monitors the process and steps in where their decision is needed.
- Train the team. A well-configured tool is easy to use and built for the team's specific requirements, so anyone can learn it.
Some tasks are automated in a week, others take a month or several months. It depends on the process and the systems you need to work with.
The goal is not to get rid of people. The goal is for one employee to be able to do the work of several without being overloaded.
What about data?
This is the first question any owner asks, and the right answer depends on the company.
- Cloud models (for example, Anthropic or OpenAI) offer the best quality and speed. On business plans, providers offer terms under which customer data is not used to train models. The specific terms should be checked for each provider.
- Local models run on the company's own server, and the data never leaves the office at all. There are many open models now (for example, Qwen, Gemma, gpt-oss), and each handles different tasks better. For a business that needs complete confidentiality, you can build your own AI server, a small "mini data center" in the office.
- A combined option: sensitive data is processed locally, and everything else goes through cloud services.
Limits
I'm talking specifically about work done at a computer. Physical processes (production, warehousing, delivery) are a separate field. But around them there is almost always a lot of computer work: orders, documents, communication, bookkeeping. And that is where time is most often lost.
Also, not everything should be fully automated. Decisions a person is responsible for, and conversations where trust and empathy matter, should remain with a person. An agent can prepare everything for such a conversation, but it shouldn't conduct it in place of a person where it matters.
Where to start
Try a simple test: ask everyone on your team to write down what they did at the computer over the course of one day and mark the actions that repeated. Most often, this list contains the first candidates for automation.
If you see that your team is stuck in operations and want to understand what can be automated, write to me or book a call. We'll go through your processes and see where to start.