Many conversations about artificial intelligence begin with product names. Someone demonstrates a new language model, image generator or automation that looks impressive in a presentation. A few weeks later, however, the team still works almost exactly as before, while the software bill has grown. For a small or medium-sized company in Gdańsk, Gdynia or Sopot, this matters because every unnecessary subscription competes with sales, advertising and operating budgets.
A useful AI project should therefore start with a process, not a tool. The company first needs to identify where time, money or information is being lost. Only then should it decide whether artificial intelligence is actually the right solution.
Start with the business problem, not the model
A practical first step is to list tasks repeated every week. In a service company these may include preparing similar proposals, copying data from forms, answering recurring questions, organising meeting notes or producing first drafts. In e-commerce, useful candidates include product descriptions, enquiry classification, review analysis and creating communication variants.
Not every task is worth automating. If an activity happens once a month and takes twenty minutes, the automation may cost more than the manual work. If five people do a similar task every day, even a modest saving per person quickly becomes meaningful.
Map the process before automating it
Describe four elements: what starts the task, which data is required, who makes the decision and what the final result should be. This simple map reveals which stages can be solved with ordinary rules, which require integrations and where a language model is genuinely useful.
- Input: form, e-mail, document, recording or CRM data.
- Processing: classification, summary, response drafting, analysis or tagging.
- Control: a human decision on whether the result is acceptable.
- Output: customer reply, task, report, document or database record.
AI should not be a layer placed on top of organisational chaos. If the company has inconsistent data, unclear responsibility and several conflicting spreadsheets, a model will not fix the underlying process.
Where the first measurable gains usually appear
The most useful early projects are often unglamorous: summarising incoming sales enquiries, classifying messages, drafting a first response, searching an internal knowledge base or creating a first version of a report from supplied data. Marketing teams can use AI for research, headline variants, comment analysis, video transcription and brief preparation. The closer a piece of content is to publication, however, the more important editorial review, brand consistency and local knowledge become.
Measure time, quality and corrections
Time is the simplest metric. Measure how long the task takes for a week before the change and compare it with the pilot. Add three additional metrics: errors, human corrections and customer-response time.
If automation reduces a task from thirty minutes to eight but requires another twenty minutes of corrections, the business gain is mostly an illusion.
A pilot should have a defined period, an owner and a success criterion. For example: “for four weeks we will test automatic summaries of website enquiries; the goal is to reduce lead-qualification time by at least 30 percent without reducing data quality.” That is much more useful than “we are implementing AI in sales”.
Data security has to be part of the design
Companies working with customer records, medical data, contracts, financial information or internal know-how should establish rules before launching the pilot. Employees need to know what may not be pasted into public tools, who can access conversation history and where outputs are stored.
A short internal policy can define approved tools, types of data, anonymisation, verification and publishing responsibility. It does not need to become a fifty-page document to be effective.
A practical 30-day plan
- Select one process performed frequently by at least two people.
- Measure its time and common errors.
- Map input, decisions and output.
- Run a limited pilot on a controlled number of cases.
- Assign a human reviewer.
- After two weeks compare time, quality and correction rate.
- Integrate it with other systems only after the pilot produces a positive result.
The local competitive advantage
The advantage is not that a company “has AI”. That will soon be as unremarkable as having e-mail. The difference is how quickly a company can translate technology into a better process. In the Tricity that can be especially valuable for service businesses, healthcare, real estate, tourism, logistics, retail and creative companies, where much of the work consists of communication and information handling.
Practical takeaway: choose one repetitive task and treat AI like a new employee on probation. Give it clear inputs, expected outputs and a human responsible for control. Scale only after the pilot works.
Example: a service company that needs a better information flow, not an “autonomous agent”
Imagine a local company receiving dozens of enquiries every week through forms, e-mail and social media. Sales staff read each message, copy information into a spreadsheet and then decide who should respond. A useful first project is much simpler than a fully autonomous agent: collect messages in one place, generate a short summary, classify the topic and suggest a category. A person approves the result.
This can reduce manual work while keeping the risk low. After several weeks, the company also gains structured information about the most common enquiries, timing and sales outcomes. That evidence can guide the next automation.
Calculate the business case
If a task consumes twelve hours per month, the relevant hourly cost is 80 PLN and automation reduces the time by half, the potential monthly saving is 480 PLN. Subscriptions, implementation, maintenance and human review still need to be deducted. The calculation is simple, but it helps distinguish a business project from a technology experiment.
FAQ before implementation
Do we need our own model?
Usually not at the beginning. Custom infrastructure becomes relevant when scale, security, cost or specialised data create a clear reason.
Can AI operate without human review?
In some low-risk processes, yes. A safer route is to first understand typical errors and then gradually reduce human involvement.
How do we encourage adoption?
Do not begin with a presentation about the future of work. Demonstrate one frustrating task that becomes faster and easier. Adoption follows practical value.