EMAIL · ORDERS · DOCUMENTS
AI agent development for business
We build AI agents that read emails, check your CRM and ERP, and prepare each case for a decision. Your team approves instead of retyping. We start with one process and a test on 20–50 of your real tasks.
Fixed price after the audit, before we write the first line of code.A person always approves irreversible actions: an email to a customer, a refund, a payment.Code and model instructions in your repository after every project stage.
Call: 15 minutes, video or phone · Prefer to call? +48 663 877 600
WHERE AN AGENT PAYS OFF FASTEST · 4 PROCESSES
Processes where an AI agent pays off fastest
An AI agent pays off fastest in quote requests, complaints and returns, orders from email into the ERP, and lead qualification with meeting booking — wherever an employee reads text, checks two or three systems and only then decides.
Quote requests
400 requests a month, 15 minutes each
A salesperson looks up the client in the CRM, checks the price list and writes a reply. The agent does the same in the background and leaves a ready draft. The employee reads it, fixes one sentence and sends it.
Complaints and returns
Every case is different, the terms run to 40 clauses
The agent gathers the order data, the customer's history and the photos from the email, matches the right clause of the terms and proposes a decision with its reasoning. A person decides on the refund.
Orders from email into the ERP
2–3 hours a day of retyping
Orders arrive by email, as PDFs and in the message body, in a different layout every time. The agent reads the line items, checks the product codes in the ERP and fills in the order. An employee approves it with one click.
Qualification and meetings
A Friday 7 p.m. enquiry waits until Monday
The agent asks about scope, budget and timing, records the contact in the CRM and offers a free slot from the calendar. That is how Vi works — our own assistant in the Polish version of this site.
↳ Can the process be written down as rules? See process automation — PLN 15,000–30,000 net per process, 2–4 weeks.
// OPTIONS · 3 WAYS TO START · NET PRICES + 23% VAT
Three ways to your first agent — with prices
There are three ways to start: an AI audit at PLN 3,900 (5 working days), a supporting agent from PLN 20,000 (5–8 weeks) or an agent with integrations from PLN 35,000 net (8–12 weeks). After the audit or specification you get a fixed price before we start building.
ENTRY · AI AUDIT
Not sure whether the process suits an agent
Included: a 2-hour workshop with your team, a review of 3 processes, a recommendation for each (agent, automation, chat or nothing) and an agent specification with a fixed price.
Order the build within 60 days and we credit 100% of the PLN 3,900 towards it.
OPTION 1 · SUPPORTING AGENT
The agent reads and prepares, a person sends
Who it's for: one process, 1–2 systems.
Included: read tools, drafts (reply, note, proposed decision), an approval panel, a test on 20–50 tasks, limits and action logs. An employee approves every send.
OPTION 2 · AGENT WITH INTEGRATIONS
Agent with integrations — CRM, ERP, email, calendar
Who it's for: a process that runs across 2–4 systems, where the agent also has to write to them, not just read.
Included: everything in option 1, write tools with permissions enforced by the system, approval of irreversible actions, hand-over to a person with the full history, tests of refusals and of attempts to hijack the agent.
METHOD · HOW WE CHOOSE THE SOLUTION
Agent, automation or chat — how we choose
Of the three, an agent has the highest cost per task and is the least predictable. So we start by asking whether you need one at all. One column decides it: what happens in the process.
What happens in the process | Solution | Start-up cost (net) | Cost per task | Who approves |
|---|---|---|---|---|
The rule is complete: “if the invoice is from X, book it to Y” | Automation | PLN 15,000–30,000 | a few grosz, often PLN 0 | nobody, the rule runs on its own |
The answers are in documents, nothing changes in your systems | Chat or AI assistant | from PLN 8,500 | approx. PLN 0.36 per conversation | nobody, the chat only answers |
Someone has to read text, check 2+ systems and propose a decision | AI agent | from PLN 20,000 | approx. PLN 0.82 per 10 steps | an employee, for every irreversible action |
In none of the options do we leave decisions about people, such as in recruitment, or payments to an agent without a human involved. How an agent differs technically from a chat and from automation, we explain in the article AI agent in business. All our AI services with prices are listed on the AI implementation page.
// OUR OWN DEPLOYMENT · LIVE SINCE OCTOBER 2026
Vi: how we put limits on our own AI that writes to the CRM and books meetings
Vi is the AI chat in the Polish version of digitalvantage.pl. It doesn't plan multi-step tasks like a full agent, but it uses tools that change data in our systems. We used it to test the rules we apply for clients.
WRITE TOOLS
Writes to the CRM
It creates a contact and a sales opportunity, adds a note with a link to the conversation and a task for the team — through 5 operations of our CRM's API. The key it connects with opens only those operations and reaches nothing else in the system. The integration is covered by 25 contract tests and 33 API tests.
BOOKING WITHOUT GUESSING
Books a slot from the calendar
It shows only slots returned by the calendar and books exactly the one picked from the list. If the slot disappears in the meantime, it fetches the list again and doesn't claim success. It doesn't book without an email address. It takes the time zone from the browser.
HAND-OVER TO A PERSON
Hands the case over with its history
It hands over in 6 situations, including when the visitor asks for it, when it has no answer, for complaints and for a hot sales lead. After hand-over the bot stays silent. Outside working hours it offers a call slot instead of promising a quick reply.
BRAKE
Has limits set in the system
A monthly token budget — once it is exceeded, the bot stops using the model and hands the case to a person; outside working hours it sends one pre-written message. The booking system accepts at most 10 bookings per hour from one address, and all of Vi's bookings come from one server — that is the ceiling for the whole chat. A cheaper model extracts the data for the CRM, because it is good enough for a fixed format.
↳ See it for yourself: pick a time for a call in the form below — the slots offered come from our calendar.
WHAT YOU GET · 4 THINGS INCLUDED
What you get with every agent — included in the price
Every agent includes four things, not as extras: a specification with a “task done” criterion, 20–50 test tasks from your own cases, step and budget limits, and action logs with 30 days of support after launch. Code, instructions and tool definitions go into your repository after every project stage.
Specification
“Task done” defined as a state in the system
“The agent wrote ‘booked’” is not a criterion. “The booking exists in the calendar” is. The specification also lists the tools with their risk level and the actions that need approval.
Tests
20–50 test tasks from your own cases
The set is built from real cases, including those where the agent should refuse. We run each task several times, because one successful attempt proves nothing. You get the results in writing.
Limits
Steps, budget and rate written into the specification
A step limit per task, a monthly budget and a request limit per user. Once a limit is exceeded, a person takes over. What happens after the limit doesn't depend on the model.
Logs and support
Action logs and 30 days of support after launch
Every tool call, its parameters and its result go into the logs. For 30 days after launch we read them with you and refine the instructions and tools.
// PROCESS · 5 STAGES · 5–12 WEEKS
How an AI agent is built — 5 stages, 5–12 weeks
Each stage ends with something you sign off. The time in brackets applies to an agent with integrations.
5 working days
Audit and specification
What you get: a document with the task, the “task done” criterion, the tools with their risk level, the actions to approve, the limits and a fixed price.
1–2 weeks (2–3 weeks)
Prototype on the best model, read-only
What you get: a working prototype plus the accuracy, time and cost of a single task. We swap in a cheaper model where the test shows it is enough — as with data extraction in Vi.
1–2 weeks (2–3 weeks)
Tests on 20–50 tasks
What you get: results from several runs of each task, refusal tests and a list of what the agent does well and what it can't do yet.
2–3 weeks (3–5 weeks)
Launch on part of the cases
What you get: the agent working on part of the cases, a weekly log review and fixes to the instructions and tools.
Ongoing · PLN 900–1,500/month
Extending permissions and maintenance
What you get: a new tool only once the logs show the agent handles the current ones, and tests before every model change.
In total: supporting agent 5–8 weeks, agent with integrations 8–12 weeks. Payment 30/40/30: 30% when the prototype starts, 40% at testing, 30% at launch on part of the cases. The PLN 3,900 audit is credited towards the build.
- AGENT BUILD · ONE-OFF
- from PLN 20,000
- TOKENS · 10-STEP TASK
- PLN 0.82
- HOSTING · MONTHLY
- from PLN 150
- MAINTENANCE · MONTHLY
- PLN 900–1,500
↳ The token calculation step by step: how much an AI agent costs. The wider picture — where to start with AI and what it costs: AI in business. Model pricing: Anthropic, Pricing.
Got a process for your first agent?
In 15 minutes we'll check whether it needs an agent, automation or a chat. You leave with an answer, not a sales pitch.
↳ The call takes 15 minutes, by video or phone. Prefer to call? +48 663 877 600
RISKS · 3 CONCERNS AND WHAT WE DO ABOUT THEM
Three risks of an AI agent and how we remove them
Risk
The agent does something irreversible
A separate account with minimal permissions, read-only at first. An email to a customer, an order cancellation, a refund and a payment always wait for an employee's approval.
Risk
It gets stuck in a loop and runs up the bill
A step limit per task, a monthly budget and a request limit per user. Once a limit is exceeded, a person takes over. When the agent doesn't know something, it hands the case over instead of guessing — more in the article AI agent in business.
Risk
The platform disappears or changes its settings
The code and instructions are yours. We build in n8n or Node.js, not in a model vendor's closed builder. Before every model or tool change we run regression tests.
NOT A FIT
When an AI agent doesn't pay off
An AI agent does not pay off when the “if… then…” rule is complete (automation at PLN 15,000–30,000 will do), when it is only about answers from documents (a chat or assistant from PLN 8,500), or when systems have no API and the data sits in spreadsheets.
The rule is complete
If every case can be described as “if… then…”, automation at PLN 15,000–30,000 will be cheaper and predictable.
It's only about answers from documents
If nothing needs to change in your systems, a chat or an assistant with a knowledge base is enough, from PLN 8,500. We describe this option on the page AI implementation.
Systems without an API, data in spreadsheets
The agent has nothing to reach for. First comes order in your data and systems — that is where technology consulting starts.
TECHNOLOGY · WHAT EACH IS FOR
What we build agents with
- n8n — when your team needs to see the process. The AI Agent node can require approval before a chosen tool. We also use it in process automation.
- Node.js — when you need full control over the tools and tests.
- Claude and OpenAI via API — you pay for usage and switch providers in the configuration. Tests show whether the agent works the same on a new model.
- MCP (Model Context Protocol) — a standard for connecting a model to tools. We set permissions the same way: in the system, not in the prompt.
FAQ · 8 QUESTIONS
Questions before building an AI agent
How much does it cost to build an AI agent, and how much per month?
A supporting agent costs from PLN 20,000, an agent with integrations from PLN 35,000 net. The entry point is a PLN 3,900 AI audit, credited towards the build if you order it within 60 days.
Monthly: tokens approx. PLN 0.82 per 10-step task (1,000 tasks ≈ PLN 823), hosting from PLN 150, maintenance PLN 900–1,500.
How long does it take to build an agent?
A supporting agent takes 5–8 weeks, an agent with integrations 8–12 weeks, counting from the 5-day audit. It depends most on the number of systems and on how quickly we get access to their APIs and 20–50 sample cases for testing.
What if the test on 20–50 tasks goes badly?
You end the project after stage 03. You pay 30% and 40% of the price, i.e. for the prototype and the tests. The code, instructions, test set and results stay with you. The results also show which part of the process can be automated with rules alone.
Who owns the code, instructions and tools?
You do. The code, model instructions, tool definitions and test set go into your repository after every project stage. You can develop the agent further with another contractor or your own team.
What does the agent do with personal data?
It processes only the data the task needs, from the systems listed in the specification. We host the application and logs on our own infrastructure in the EU or on your infrastructure. The model region is written into the specification, and you sign a data processing agreement with us (Art. 28 GDPR). We agree the log retention period before launch.
Does the agent have to say it is AI?
Yes, when it contacts people. Art. 50 of the AI Act applies from 2 August 2026, and the European Commission's non-binding guidelines of 20 July 2026 explicitly cover agents that handle correspondence or book appointments. The agent has to disclose that it is AI and on whose behalf it acts. Emails from our agents carry an AI label at the top. Details in the article EU AI Act for business.
How is this different from an agent in ChatGPT or Copilot?
A subscription agent works on whatever you paste in or share with it, with the vendor's settings. Ours works in your 2–4 systems, on a separate account, with limits and a log of every step. When a subscription is enough, we explain in the article ChatGPT Business, Copilot or Gemini.
What happens after launch, and how do I end maintenance?
The first 30 days of support are included. After that, maintenance costs PLN 900–1,500 a month and covers log reviews and tests before every model change. The notice period goes into the contract before launch. When maintenance ends, the agent keeps working and the code stays with you.
RESOURCES · BEFORE WE TALK
Before we talk
ARTICLE
AI agent in business — when it makes sense
How an agent differs from a chat and from automation, and what it costs in tokens.
ARTICLE
EU AI Act for business
Duties, deadlines and fines — including telling people they are dealing with AI.
// Next step
Let’s talk about AI in your company
Describe the process that takes the most time. On the call we tell you whether an agent makes sense there and where to start.
What happens next
- We confirm the call within one working day.
- On the call we go through the process and the data an agent would work with.
- You get a recommendation on where to start and what it costs, or an honest “not worth it”.