Before anyone selects a tool, select the process. Companies usually start the other way around: with an impressive demo, then searching for a place to force it in.
Models in 2026 are solid. The bottleneck is knowing how much the target process costs today.
Below is how we run the first few weeks. It is no secret and you can apply it without us.
Three questions that settle most decisions
Is the process repeatable?
Dozens of times a month, in the exact same format: a form inquiry, a supplier invoice, a resume for a job posting. A process that looks different every time is not suitable for a first implementation, no matter how painful it is.
Does anyone measure it today?
If no one knows how many cases arrive each week, how many minutes each takes, and how many customer errors return, ROI cannot be calculated. A week of measurement before deciding is cheaper than any single day of implementation.
Can an error be caught before it has an impact?
A process where a human approves the output before sending is ready to start. Leave processes where the system automatically sends an offer, invoice, or client reply for stage two. The first stage will reveal how often the model makes errors and where.
What the first eight weeks look like
From workshop to production deployment usually takes 4–8 weeks. Stages are billed separately, and each subsequent stage starts only after the previous one is fully accounted for in hours.
- Week 1: workshop and process selectionNinety minutes with the founder. Map out pain-point processes and select one based on the three questions above. You leave with a list of systems to integrate and clear ownership on your side.
- Week 2: measurementWe measure: ticket volume, minutes per ticket, revision counts, cases pending over 24 hours. Without baseline metrics, post-launch comparisons rely purely on impressions, making ROI discussions subjective.
- Weeks 3–6: development on a data copyWe build the system in a sandbox using a copy of your data, dedicated credentials, and permissions defined prior to kickoff. Writing data to live systems requires manual human approval.
- Week 7: live case testingA benchmark of at least thirty real-world cases from recent months, including complex edge cases. We evaluate the results by precise hit-or-miss rates, not subjective impressions.
- Week 8: decisionProduction release, revision, or cancellation. We include the exit clause directly in the contract to ensure nobody defends an unviable project.
When a subscription is enough, and when it is not
Claude for Small Business, ChatGPT, and Copilot offer pre-built workflows included in their subscription plans. For many companies, this suffices, and we state this openly during workshops. The real distinction lies in the workflow architecture surrounding the tool.
Subscription suffices
- A single user runs the tool for individual tasks
- Non-sensitive or public data
- A human verifies the output before use
- The tool features a native integration with your systems
- No documentation requirements
- Funded directly from operational expenses
Implementation required
- Multiple users collaborate within one process, passing outputs across departments
- Personal customer data, contracts, HR records, trade secrets
- The output routes directly to clients, onto invoices, or into accounting software
- System without an API: warehouse software, HR and payroll, legacy CRM
- Required: AI acceptable use policy, Article 4 training, Article 50 disclosure, GDPR register entry
- Implementation funded by a grant, requiring a contractor, schedule, and KPIs
If all points match the left column, buy a subscription and return to us in a quarter when the right column appears.
What to bring to the workshop
Three painful processes
No descriptions needed. A name and the person running them today are enough.
Who and how much time
Even roughly: "Anna, about two hours a day entering invoices into the system".
List of systems
Accounting software, warehouse, CRM, email. Include names, as integration costs depend on them: from 2,000 PLN per integration, based on our implementations.
Does the process include personal data
This determines whether the model can run in the cloud or must stay on a server in your rack.
Budget or grant program
If the implementation relies on PARP or KPO, application deadlines dictate the schedule, not the other way around.
Questions and answers
Frequently asked questions
Where to start with an AI implementation in a company?
Start by choosing one process, not a tool. The process must be repeatable, measurable, and supervised so a human approves the output before it takes effect. Then take one week to measure baseline metrics, followed by implementation. Going from the workshop to production usually takes 4–8 weeks.
Isn't ChatGPT or Claude enough?
Often yes: when one person uses the tool, the data is not sensitive, and a human reviews the result before use. A custom implementation is needed when multiple people share the process, personal data is involved, systems lack APIs, or the output goes directly to a client or onto an invoice.
How much does the workshop cost?
A 30-minute call and a 90-minute workshop are free of charge and booked via calendar. The implementation quote is prepared after the workshop, always with price ranges provided before the start. Price ranges from our implementations are published on the pricing page.
What if the pilot shows it does not pay off?
Then we wrap up after the pilot. This option is in the contract from day one. Thresholds below which we advise against implementation are published: for example, document workflow does not pay off below 200 invoices per month.
Other paths
What next
We built it ourselvesReview of an AI implementation you built yourselfAn agent set up in Claude Code, n8n, or Make works until it does not. We check what happens when it stops.
Law and oversightAI Act and Polish law: company obligationsArticle 4, Article 50, Polish law from August 11, penalties from October 28. What we check and what documentation we deliver.
GrantsAI implementation grants for companies in 2026PARP, KPO, FENG: which programs, what thresholds, and what we prepare for the application to make it submission-ready.- GeneratorImplementation plan in nine questionsIndustry, size, process, data, systems. In response, you receive recommendations, stages, cost benchmarks, responsibilities, and grants. Every item cites a source.
No-obligation call
90-minute workshop with the founder
A process map, one selected process, and a list of systems to integrate. You leave with an actionable plan, whether you choose to work with us or not.