Start with what wears you down, not with the technology
The most common mistake I see: a business wants to automate whatever sounds impressive. “We want an AI that writes our sales proposals.” But the real win almost always sits where nobody wants to look: in the boring day-to-day work someone does by hand ten, twenty, fifty times a day.
My advice is to write it down. For one week, note where your team spends the most hours, and next to each item, whether that work could be written up as a recipe. Not a feeling, but steps. Two or three clear favorites almost always stand out. And they're rarely the ones you thought of at first.
Three questions that separate a good candidate
A good candidate gets a “yes” on all three:
- Does it repeat? A customer writes “where's my order?”, and that happens fifty times a day. Same question, same path to the answer.
- Does it follow rules? If the decision can be written as “if this, then that”, an agent will repeat it. “Look up the number in the system, report the status, and if it's late, offer compensation” is a recipe, not magic.
- Can you tell when it's done right? It has to be clear what a good result looks like. An entered invoice either matches the PDF or it doesn't.
If even one answer is “no”, stop. That's not the right work yet.
What to hand off first
Concrete examples that pay off fastest:
- Repeat customer questions. Order status, opening hours, return policy, prices. The agent answers instantly, any time, in whatever language the customer wrote in.
- Moving data between systems. An invoice arrives as a PDF, and someone opens it and retypes the numbers into the accounting system. An agent does the same: without typos and without a lunch break.
- Qualifying inquiries. You get a hundred inquiries, but only twenty are worth a call. The agent sorts them by your criteria and leaves sales only the warm ones.
Notice the pattern? These aren't the biggest jobs. They're the most frequent.
If you work in accounting, I wrote up the same principles with concrete document volumes separately: AI for accounting and bookkeeping. For carriers and freight forwarders: AI for logistics and transport companies.
What not to touch (at least for now)
This is where the most money gets burned, so draw the line right away.
Leave to people whatever needs negotiation, empathy or a rare exception. An unhappy customer threatening to leave is not a job for an agent. A messy case that comes up once a month and is different every time isn't either. “Automating” that costs more than you'll ever save, because there's no end to the exceptions you'd have to anticipate.
And one more thing few people mention: don't touch a process that changes every week. Automation loves stability. If the way you do things isn't settled yet, sort it out with people first, and only then hand it to an agent.
Sensitive customer data is a separate question. I don't refuse to automate it, but it's worth knowing the rules first: AI and GDPR.
The smartest model is a hybrid
It's rarely all or nothing. What works best is a simple split: the agent handles the repetitive majority, and politely hands you the few cases that really need a person.
An example from customer service: out of a hundred messages, most are “where's my order” or “what are your hours”, and the agent handles those. The rest, where the customer is angry or the case is messy, it flags and passes to a person with the full context. The person is left with only the interesting part.
That's exactly why what usually gets automated isn't a job, but a few hours of it. More on that: will AI replace employees.
Don't try to automate the biggest task. Automate the most frequent one.
The real cost isn't building it, it's keeping it running
This is where most people get it wrong. Building an agent is a one-off job. But processes are alive: prices changed, a new service appeared, you migrated to a different system. If nobody maintains it, the agent quietly starts making mistakes. So don't just count “how much will it cost to build”, but also “who will keep it alive”. For us, that's part of the solution, not an add-on to the bill. I collected concrete market numbers in a separate article: how much AI automation costs in Lithuania in 2026.
How to test it without taking a risk
You don't need to rebuild your whole business at once. Pick one clear candidate, the one that repeats most often. Run a pilot on your own real data. Measure over a couple of weeks: how much time you saved, how many fewer mistakes there were.
If it matches expectations, you expand. If it doesn't, you lost a couple of weeks, not half a year. That's exactly how we work: we find the repetitive work, build an agent that does it from start to finish, and prove the value with your numbers. If we can't prove it, you pay nothing.