Can AI really help startups operate with fewer employees? Redwerk founder Konstantin Klyagin shares firsthand lessons from successful automation projects and costly experiments, including which tasks are worth automating, why some AI systems fail, and how small businesses can calculate the true cost of adopting AI.
Redwerk founder Konstantin Klyagin shares firsthand lessons on choosing the right tasks for AI, avoiding costly automation mistakes, and knowing when human judgment matters more than technology.
Key Takeaways: What Startup Founders Should Learn About AI Automation
- Automate individual tasks before considering entire jobs. Many employees handle responsibilities that involve judgment, relationships, and business knowledge that cannot easily be replaced by AI.
- Start with repetitive work that is easy to verify. Document summaries, inbox organization, meeting notes, initial drafts, and information extraction are often more practical starting points than complicated business decisions.
- Get your business processes organized first. Incomplete records, inconsistent procedures, and poorly documented workflows can undermine automation projects and create costly mistakes.
- Calculate the real return on investment. AI subscription fees are only one expense. Founders should also consider integration, usage charges, human review, monitoring, maintenance, and error correction.
- Don’t overlook traditional automation. When a task follows clearly defined rules, existing software or a straightforward integration may be cheaper and more reliable than an AI agent.
- Run a small pilot before expanding. Klyagin recommends testing one workflow for 30 days, measuring time saved and error rates, and keeping a person responsible for reviewing results.
- Use AI to make a small team more capable. For an early-stage startup, helping existing employees spend more time on sales, customers, and product development may provide greater value than cutting positions.
For entrepreneurs starting a business, artificial intelligence offers something that would have been difficult to imagine just a few years ago: the ability to handle a growing workload without immediately hiring more people. AI tools can help write marketing content, organize customer inquiries, prepare proposals, summarize meetings, and complete administrative tasks that might otherwise consume hours of a founder’s day.
For a startup operating on a tight budget, those capabilities are attractive. When every dollar counts and the founder is already juggling sales, customer service, bookkeeping, and business development, the prospect of having an AI assistant handle routine work can seem like an obvious investment. Some AI platforms even promise to perform the responsibilities of entire employees, raising the question of whether startups should rely on automation instead of making their first few hires.
But just because a task can be automated does not mean the automation will save money or produce reliable results.
An AI tool might generate a customer proposal in seconds, but someone still needs to verify the pricing, promises, and contract terms. A customer service chatbot may answer common questions accurately while struggling with unusual complaints or situations that require discretion. An AI agent might complete a series of administrative tasks successfully during a demonstration, only to encounter problems when working with incomplete records, inconsistent procedures, or real customers.
These limitations can become particularly expensive for startups, which often lack the financial resources and technical support needed to recover from a poorly planned technology investment.
Table of Contents
Why Startups Need to Think Differently About AI Automation
One of a young company’s most important decisions is deciding which responsibilities people should handle, which can be automated, and which might benefit from a combination of both.
The answer is rarely as simple as replacing a job title with an AI application. A customer service employee, for example, may also recognize recurring product problems, preserve relationships with dissatisfied customers, and identify opportunities to improve the business. Those contributions can be difficult to capture in a basic comparison of salary costs and AI subscription fees.
The same challenge applies to founders who are tempted to automate every part of their operations before establishing consistent workflows. If customer records are inaccurate, billing procedures differ from one transaction to another, or important information exists only in someone’s memory, automation can reproduce those problems rather than resolve them.
For entrepreneurs exploring how to adopt generative AI in a small business, the more practical starting point is to identify repetitive work, establish reliable processes, and determine how success will be measured.
The National Institute of Standards and Technology’s AI Risk Management Framework also emphasizes evaluating AI systems for reliability and risks within their intended use. Startups may not need a sophisticated corporate AI governance program, but they do need a way to recognize mistakes, protect sensitive information, and decide when human review is necessary.
Learning From Real AI Automation Projects
To better understand what businesses should automate and where they should proceed cautiously, StartingUpTips spoke with Konstantin Klyagin, founder of Redwerk, a software development company with experience building business applications, automation systems, and AI-assisted workflows.
Klyagin’s perspective comes from both client projects and Redwerk’s own experiments with artificial intelligence. His experiences illustrate why successful automation often depends less on the sophistication of the AI tool than on understanding the business process it is supposed to improve.
In one project involving an Australian sign-language interpreting agency, Redwerk helped reduce manual administrative work by a reported 40%. Rather than automating the entire operation, the company first addressed existing workflow problems and retained human involvement in decisions where errors could have serious consequences.
Not every experiment produced similar results. Klyagin also describes an attempt to build an AI-powered social CRM that struggled with lost data, inconsistent behavior, and token expenses reaching approximately $100 per day. The experience reinforced the importance of testing, managing costs, and understanding the limitations of increasingly autonomous AI systems.
His central advice is especially relevant to startups deciding whether to invest in AI instead of hiring employees.
“The first question should be: ‘Which tasks inside this role are repeatable and easy to check, and who catches the mistakes when AI gets one wrong?'” Klyagin explains.
For startup founders, the distinction matters. The objective should not be to automate as much work as possible, but to find opportunities where technology can improve productivity without creating new expenses, operational risks, or customer service problems.
In the following interview, Klyagin discusses the business tasks best suited for AI, the hidden costs of automation, mistakes companies make when eliminating positions, and how a small company with five, ten, or twenty employees can begin experimenting with AI without making risky staffing decisions.
Interview With Konstantin Klyagin, Founder of Redwerk
Editor’s note: The following interview was conducted in writing for StartingUpTips. Klyagin’s responses discuss lessons from Redwerk’s client engagements, internal AI experiments, and experience implementing business automation.
When companies say they want AI to “replace a role,” what is the first question they should be asking instead?
The first question should be: “Which tasks inside this role are repeatable and easy to check, and who catches the mistakes when AI gets one wrong?” A role is a bundle of tasks. Usually only some of them are good candidates for automation. When a company maps the work task by task, it typically finds that AI can absorb part of the volume, while the exceptions, judgment calls and relationships stay with a person. That map tells you whether you are redesigning a job or removing one, and in most cases it points to a redesign.
Can you give an example of a job that looks automatable until you break down what the person actually does?
Customer support is the classic one. On paper, the job is answering tickets, and AI handles a lot of that well. Look closer and a good support rep is also spotting patterns across complaints and flagging bugs to engineering, deciding when to bend a policy for a high-value customer, calming down someone who is about to cancel, and recognizing when an issue carries legal or safety risk.
The ticket answering can be automated. Most of the role’s value sits in everything else. Companies that cut the whole role tend to lose the bug reports and the retention saves first, and they find out months later when it shows up in churn.
What tasks are currently best candidates for AI automation, and what still require too much human judgment?
The best candidates are high-volume, rule-based and quick to verify: summarizing documents, pulling data out of forms, sorting and routing requests, writing first drafts, and generating test scenarios. In our QA work, AI helps with localization checks, accessibility validation and regression testing, and engineers review the results.
Human judgment is still needed for decisions with legal or financial consequences, trade-offs without a clear right answer, architecture choices, negotiations, and any task where a mistake is expensive and hard to spot. My rule of thumb is simple: if checking the AI’s output takes almost as long as doing the work, the savings disappear.
What do companies most often misjudge before eliminating a role because they believe AI can handle it?
They misjudge the work nobody wrote down. Job descriptions list the visible tasks, but experienced employees carry a lot of informal knowledge: which client needs extra care, which system breaks on the last day of the month, who to call when something goes sideways.
Companies also underestimate the exception rate. AI might handle 80% of cases well, and the remaining 20% can take up most of a person’s time. The third misjudgment is review cost. Someone still has to check the output, and that time belongs in the plan from day one.
Can you share a real example where a client wanted to automate more of a process than you ultimately recommended? What changed their thinking?
We worked with an Australian sign-language interpreting agency that wanted to reduce the manual work involved in bookings, payroll, and invoicing. They were still relying on spreadsheets alongside their existing software. The bigger goal was to automate as much of that workflow as possible.
But the platform already had issues with staff availability, booking notifications, and billing. Automating everything at once would have made those problems worse. We recommended fixing the underlying issues first, then introducing automation step by step.
We automated cancellation fees and notifications, while keeping staff involved in processes requiring review and approval. The important consideration was what happens when something goes wrong. A scheduling mistake could leave someone without an interpreter for a medical appointment. An incorrect cancellation fee could affect someone’s income.
Ultimately, we reduced manual administrative work by 40%. The lesson was that you don’t have to automate every decision to make a big difference.
How can a business tell the difference between an AI task that works well in a demo and one that can perform reliably in day-to-day operations?
A demo runs on clean inputs and the happy path. Production brings messy data, edge cases, volume and change over time. To test reliability, run the system on a month of real historical data and measure the error rate. Then check what happens when it fails: who gets alerted, how fast someone notices, and what the fix costs.
AI outputs also vary from run to run, so the testing has to account for that. We built a test automation framework for an AI note-taking app where the summaries came out different every time, which made standard exact-match checks useless. We built AI into our own test scripts to check structure, keywords and meaning, so we could confirm the app captured the right context and action items. That is the kind of verification production AI needs.
What role does data quality play, and what happens when a company tries to automate before its workflows or internal data are ready?
Data quality decides whether the project works. AI speeds up mistakes as fast as it speeds up output, so automating a messy workflow gives you the same mess, faster and at larger scale. If customer records are duplicated, processes differ by department, or key steps live only in someone’s head, the AI will learn those inconsistencies and repeat them with confidence.
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. The order that works is to document the workflow, clean the data, and then automate.
How should a business calculate the real ROI of AI once it includes software costs, monitoring, corrections and human review?
Start with the full cost of the AI setup: licenses or usage credits, integration and build time, ongoing monitoring, the human hours spent reviewing outputs, the cost of fixing errors that reach customers, and the upkeep when models or vendors change. Compare that to the honest, fully loaded cost of the current process.
The number most companies skip is review time per output. If a person spends ten minutes checking something that used to take fifteen minutes to produce, the real saving is five minutes, minus everything else on that list. Also decide up front whether a human needs to stay in the loop, because that one decision changes the math more than anything else.
Have you ever advised a client not to use AI for something they specifically wanted to automate? Why?
Yes, and it happens more often than people expect. AI has become the default answer to almost every business problem, and I understand the appeal. But the tool has to fit the task. When the rules are clear and the right answer is always the same, regular software does the job better, faster and for a fraction of the cost.
Take a logistics platform we worked on. The company needed to identify duplicate customer records, track driver availability, and update order statuses automatically. These were predictable processes with clear business rules. We solved them with traditional software development, database checks, and straightforward automation.
You don’t need an AI agent to figure out that a completed delivery should be marked as delivered. The same goes for calculating cancellation fees, reconciling payments, or sending reminders based on fixed conditions.
AI makes sense when the task involves interpreting unstructured information or handling situations that are difficult to describe with rules. But when the logic is clear, traditional software is usually cheaper, faster, and easier to maintain. I would rather save a client money than sell them AI they don’t need.
For a small business with 5, 10 or 20 employees, where would you start looking for AI opportunities without making risky staffing decisions?
Start with time. Have everyone track for one week where their hours go, then look for repetitive admin work: inbox sorting, scheduling, meeting notes, invoicing, first drafts of proposals, answers to common customer questions. Pick one workflow, run AI on it for 30 days with a person reviewing every output, and measure the hours saved and the error rate.
At that size, the smartest move is to give your current team hours back and put those hours toward sales, service or product. The people who already know your business are the ones who can tell you whether the AI is getting it right.
How has Redwerk itself changed the way it works because of AI? Are there tasks your own team now automates that employees previously handled manually?
The most obvious is routine work: task handovers, comment checks, reports, and info-chasing now run on their own.
The second is based on our super-structured marketing workflow. We built automations and custom skills that cut out the middleman in project management, publishing, and newsjacking, so our marketing team now moves eight times faster.
The third is delivery, and here’s the twist: AI made it cheap to build, but expensive to fix. Clients come to us with apps AI built in a weekend that look great in a demo and fall apart in production, and a growing part of Redwerk’s work is cleaning up or rebuilding that code. Automation testing at QAwerk is nothing new for us, we’ve done it forever, but the more code AI writes, the more you need a human tester who actually thinks like a user.
Have any of Redwerk’s own AI experiments failed or produced worse results than you expected? What did you learn?
Back in the early days of experimenting, I tried building a social CRM for myself on OpenClaw with no code at all, and within a day it was losing its database, forgetting context, changing behavior whenever it switched models, and burning around $100 a day in tokens. It ended up like a low-skill assistant with a high-skill price tag.
Since then, the landscape has shifted fast: the length of tasks AI agents can handle has been doubling roughly every seven months, and memory, the exact thing that broke for me, has turned into a discipline of its own, with dedicated architectures that keep agents on track across sessions. Cost is still the part people underestimate, since the same task can burn wildly different amounts of tokens and many long-running agent tasks still need a human to step in. That experiment shaped how we build automation at Redwerk today: clear instructions, managed memory, and a real cost check before we call anything “automated.”
Final Thoughts: Start Small, Measure Results, and Keep Humans Involved
Artificial intelligence can give startups access to capabilities that once required larger teams and more substantial budgets. However, technology that appears inexpensive at first can become costly when founders underestimate the work required to integrate, monitor, and maintain it.
Klyagin’s experience reminds us that successful automation isn’t measured by how many responsibilities a company can hand off to AI. It is measured by whether the technology improves a business process, produces dependable results, and creates enough value to justify its costs.
For founders considering their first AI investment, the safest approach is often to begin with one repetitive activity, establish clear expectations, and evaluate its performance before expanding. That makes it possible to benefit from automation while preserving the experience, judgment, and customer relationships that young businesses depend on.
For a broader discussion of how AI affects staffing and workforce decisions, read our companion feature on PowerHomeBiz: Before You Replace Employees With AI, Find Out What the Technology Can Actually Do.
