---
title: "AI agents"
url: "https://wzor.biz/en/services/ai-agents/"
description: "We design and deploy AI agents inside products: scope of actions, conversation UI, RAG with source citations, guardrails and a control set per release."
lang: "en"
---
# AI agents

At Wzór UX agency you receive an AI agent designed around your product, deployed by our team or by your own developers. We add the agent to a product you already run, or design it together with a new [web application](https://wzor.biz/en/services/web-app/).

[Free consultation](https://wzor.biz/en/free-consultation/) [Projects](https://wzor.biz/en/projects/)

-   [105+ projects](https://wzor.biz/en/projects/)
-   [115+ clients](https://wzor.biz/en/clients/)
-   [27 testimonials](https://wzor.biz/en/testimonials/)

Challenges

## What challenges do **AI agents** help with?

###  ![](https://wzor.biz/_astro/challenge-04.DJ5e_wQQ_2aL6Xb.webp) “We want to roll it out, but _I am afraid to give an agent access to our systems…_”

Our approach

Before the first screen exists, we agree for every task what the agent does on its own, what it hands to a person for approval and what it does not touch in any scenario. An operation that cannot be undone goes through a confirmation, and every risk on the list gets its own brake.

Scope of services

-   [Technical Audit](https://wzor.biz/en/services/technical-audit/)
-   [Custom software](https://wzor.biz/en/services/custom-software-development/)

###  ![](https://wzor.biz/_astro/challenge-13.K1R-BuTQ_1RkR8F.webp) “I do not want it _making up answers for our customers…_”

Our approach

We teach the agent to say only what is in your documentation and to show the source with every answer. We design the "I do not know" state as carefully as a correct answer, because that is what saves the conversation when a question goes beyond what the agent knows.

Scope of services

-   [UX Research](https://wzor.biz/en/services/user-research/)
-   [UX consulting](https://wzor.biz/en/services/ux-consulting/)

###  ![](https://wzor.biz/_astro/challenge-08.CpUOXUL6_Z1GH2pM.webp) “We have been through a pilot that _ended in a presentation…_”

Our approach

We start from the tasks the agent is meant to take over and from the number that will tell you it has taken them over. We build the set of control questions with your experts and run every release through it before launch, so at the end you have a result, not an impression from a demo.

Scope of services

-   [Product discovery](https://wzor.biz/en/services/product-discovery/)
-   [Product strategy](https://wzor.biz/en/services/product-strategy/)

Who it is for

## Which companies do we build **AI agents** for?

Stores Customer support Documents SaaS Sales

### Online stores

We design agents that stand next to the customer at the moment of choice, not in a little window at the bottom of the screen. The agent has to know stock levels, variants and delivery times, because without them it asks the same questions as a contact form, only slower.

### Companies with customer support

We take on support teams that answer the same questions over and over. We start by separating the tickets the agent can close on its own from the ones it should prepare for a person, because that split decides the result, not the quality of answers in a demo.

### Companies with knowledge in documents

We meet companies where the terms and conditions sit in one place, the arrangements in mailboxes and the procedures on an intranet nobody opens. The agent then answers from your documents and shows the source for every sentence, so an employee can check where it came from.

### SaaS products and dashboards

We build the agent into a product where a simple task costs the user five screens. Instead of adding another view, we give them one place where they say what they want to do, and the product does it and shows what has changed.

### Sales teams

We set up agents for the work a salesperson does before and after a call: gathering what is known about the company, preparing a note, reminding them of what was agreed. We leave the conversation to the person, because it is the one part nobody wants to get from a machine.

-   [E-commerce](https://wzor.biz/en/industries/e-commerce/)
-   [Software](https://wzor.biz/en/industries/saas/)
-   [Banking & Finance](https://wzor.biz/en/industries/fintech/)
-   [Healthcare](https://wzor.biz/en/industries/medtech/)
-   [Energy](https://wzor.biz/en/industries/energy/)
-   [Media](https://wzor.biz/en/industries/media/)
-   [Aerospace & Defense](https://wzor.biz/en/industries/aerospace-and-defense/)
-   [Automotive & Mobility](https://wzor.biz/en/industries/automotive-and-mobility/)
-   [Crypto & Blockchain](https://wzor.biz/en/industries/crypto-and-blockchain/)
-   [Education](https://wzor.biz/en/industries/education/)
-   [Entertainment](https://wzor.biz/en/industries/entertainment/)
-   [Fashion & Accessories](https://wzor.biz/en/industries/fashion-and-accessories/)
-   [Grocery Retail](https://wzor.biz/en/industries/grocery-retail/)
-   [Gaming](https://wzor.biz/en/industries/gaming/)
-   [Public Sector](https://wzor.biz/en/industries/public-sector/)
-   [Insurance](https://wzor.biz/en/industries/insurance/)
-   [Logistics & Fulfilment](https://wzor.biz/en/industries/logistics-and-fulfilment/)
-   [Industrial Manufacturing](https://wzor.biz/en/industries/industrial-manufacturing/)
-   [Pharma](https://wzor.biz/en/industries/pharma/)
-   [Real Estate](https://wzor.biz/en/industries/real-estate/)
-   [Retail & FMCG](https://wzor.biz/en/industries/retail-and-fmcg/)
-   [Cybersecurity](https://wzor.biz/en/industries/cybersecurity/)
-   [Services](https://wzor.biz/en/industries/services/)
-   [Sport](https://wzor.biz/en/industries/sport/)
-   [Telecommunications](https://wzor.biz/en/industries/telecommunications/)
-   [Transport & Travel](https://wzor.biz/en/industries/transport-and-travel/)

First conversation

## Free consultation

In the first conversation we walk through the process the agent would take over and agree what it will do on its own and what stays with a person. You come away with the scope of the first version of your AI agents and what needs to be ready on your side. To start, a description of the process is enough, even in bullet points.

[Free consultation](https://wzor.biz/en/free-consultation/) [Projects](https://wzor.biz/en/projects/)

![](https://wzor.biz/_astro/portrait.Cn4p6oJS_Z1ryFk.webp)

Patryk Korycki CEO, Business Analyst

Why us?

## Why choose our **UX agency**?

-   ### The agent does what we agreed on
    
    The agent's scope is written down and does not drift during the rollout. You have a document to show your security team and your board, instead of explaining out loud what the tool will not do.
    
-   ### Answers from your documents, with the source
    
    Every answer can be checked, because the agent names the document it took it from. Your team does not have to guess whether it is true, and a customer never gets a sentence nobody at your company has ever said.
    
-   ### Its own account, thresholds and a kill switch
    
    The agent gets its own identity and permissions narrower than an employee's, plus risk thresholds, a trail of the actions it took and a way to stop it in an instant. Without that, no board will agree to give it access to its systems.
    
-   ### A pilot that ends with a number
    
    The first version goes live on one task and works in production, instead of waiting for the full feature set. After a month the numbers show how many tickets it closed on its own, and only that is the basis for expanding it.
    
-   ### Product designers, not a plug-in
    
    The agent is designed by a team that has been designing interfaces for nine years, so talking to it is part of the product, not a window stuck in the corner of the screen.
    
-   ### Proven client trust
    
    We have designed for [115 brands](https://wzor.biz/en/clients/), and the same craft goes into the agent: the user's task first, then the technology.
    

Process

## How will we build your **AI agent**, step by step?

1.  ### A conversation with your team
    
    ![](https://wzor.biz/_astro/workshop-session.Hnn780ic_mnpiQ.webp)
    
    We choose one process for the agent to take over
    
    We break the candidate processes down: how many people use each one, what a mistake costs and whether there is a source the agent could draw its answers from at all. Some ideas are dropped here on purpose, because a rule or better search will handle them, not a model. What remains is one process and the reason it is that one.
    
    Key steps
    
    -   Reviewing the processes that are candidates for automation
    -   The cost of a mistake and the number of people behind each one
    -   Checking that a source of truth exists
    -   The criteria that will tell us the pilot worked
    
    Deliverables
    
    -   The chosen process with the reasoning
    -   List of rejected ideas and why
    -   Map of the agent's tasks and limits
    -   Pilot exit criteria
    
2.  ### Data audit and the control set
    
    We check what the agent is going to learn from
    
    Before the first prompt is written, we go through the documents: what is out of date, what is duplicated, where two versions say different things and which data has to be cut off before anything goes to the model. Conflicting sources are more dangerous here than missing ones, because the agent will answer confidently from whatever happened to win the search.
    
    Key steps
    
    -   Reviewing the knowledge sources and how current they are
    -   Duplicates and conflicting versions of documents
    -   Cutting off sensitive data
    -   Control questions from your experts
    
    Deliverables
    
    -   Report on the state of the knowledge base
    -   An index ready for search
    -   A set of control questions
    -   List of data excluded from processing
    
3.  ### Prototype of the conversation and edge states
    
    The conversation is designed before the backend is built
    
    A clickable prototype shows not only the happy path but also a refusal, a misread intent, an interrupted action and a hand-off to a support agent. We test it on a few people from your team, because at this stage fixing the script costs hours, not a sprint.
    
    Key steps
    
    -   A catalogue of mistakes and how the interface behaves for each
    -   The conversation script and the point at which the agent asks for approval
    -   The agent interface in your design system
    -   Testing the prototype with your team
    
    Deliverables
    
    -   Clickable prototype of the conversation
    -   Catalogue of edge states
    -   Interface design
    -   List of operations the agent is allowed to perform
    
4.  ### Pilot and production rollout
    
    A small group first, then everyone else, step by step
    
    The pilot goes to a limited group of users with the conversation log switched on and ends at a gate: we move on, or we close the topic with a report on what specifically did not work. Once past the gate, the release goes out behind a flag, with usage limits and a kill switch on your side, so its reach grows step by step rather than jumping to your whole user base.
    
    Key steps
    
    -   A pilot with a small group and a conversation log
    -   Measuring against the control set and the share of hand-offs to a person
    -   A gate: we go further or we stop
    -   Release behind a flag, with limits and an off switch
    
    Deliverables
    
    -   Pilot result
    -   Go-live decision with the reasoning
    -   The agent in your product
    -   An emergency off switch on your side
    
5.  ### Measurement, tuning and hand-over
    
    We improve it on real conversations, not on workshop assumptions
    
    We tune the prompts, tool descriptions and index on what people actually ask, because that always looks different from the workshop. Your team gets the documentation and a procedure for adding the next task, along with instructions for extending the control set so that a new feature does not break an old one.
    
    Key steps
    
    -   Tuning the prompts, tool descriptions and index
    -   Report on accuracy, latency and cost per conversation
    -   Handing the log and the access over to your team
    -   A procedure for adding the next task
    
    Deliverables
    
    -   Measurement report
    -   Documentation and access to the log
    -   Procedure for extending the agent
    

Methods and technologies

## Methods and safeguards we use when building **AI agents**

-   ### RAG with source citation
    
    RAG looks up the answer in an index of your documents, so an updated policy reaches the agent the same day, without touching the model. When it finds nothing that fits, the agent says "I do not know".
    
-   ### Control question set
    
    The set of control questions is put together by your experts, because they know which answer is correct and which only sounds credible. After we leave, you keep a tool for checking every change that follows.
    
-   ### Confirmation of irreversible actions
    
    We design the confirmation before an irreversible action as a screen that shows exactly what will happen. The interface asks for consent instead of reporting after the fact that the matter has been dealt with.
    
-   ### Conversation log
    
    The conversation log records the cost, latency and rating of every answer, so an argument about how well the agent works does not end in trading impressions. It also yields a list of topics the agent cannot handle yet.
    
-   ### Data minimisation
    
    Data minimisation means the model sees only the fields the task needs, and we cut off the rest before sending. Traffic goes through your backend, so keys and logs stay on your side.
    
-   ### AI Act and WCAG 2.2 AA
    
    The AI Act obligation, that is clear information that a system is running the conversation, applies from August 2026. We design it together with WCAG 2.2 AA, because the regulation requires this information to be accessible as well.
    

Selected work

## Our projects

-   [
    
    ![WygodnaDieta - a phone showing the app's home screen: an "Excellent" Trustpilot rating and a "Menu-Choice Diet" plan card priced from PLN 80.90/day, against a red gradient with vegetables.](https://wzor.biz/_astro/cover.BfmwIntu_1CFKMk.webp)
    
    Grocery Retail
    
    WygodnaDieta
    
    View project
    
    The WygodnaDieta.pl app and online store - a meal-box diet service with menu-choice diets, ready-made plans and gift vouchers. Diet selection and setup with a dietitian consultation, account management (delivery calendar, balance, notifications), picking add-ons for kids, and finalising an order with a choice of payment method (Blik, card, bank transfer, GooglePay), across mobile and web.
    
    
    
    ](https://wzor.biz/en/projects/wygodna-dieta/)
-   [
    
    ![NeoBank - three phones on a green background with the brand wordmark behind them: payment cards, a dashboard showing 27,568.89 PLN available with transfer shortcuts, and a savings goal at 61%.](https://wzor.biz/_astro/cover.cTNj7kPE_Z2t5E6w.webp)
    
    Banking & Finance
    
    NeoBank
    
    View project
    
    The NeoBank banking app: a dashboard with the balance and shortcuts, payment cards, transfers, a currency exchange, a loan repayment schedule and savings goals with progress bars. Mobile and desktop.
    
    
    
    ](https://wzor.biz/en/projects/neobank/)
-   [
    
    ![myAzymut - a monitor with a map of Europe and shipment routes between cities, a "Fleet signal" panel listing vehicles with no signal and with a delayed signal, and a timeline at the bottom; a night-time motorway junction and a large myAzymut wordmark in the background.](https://wzor.biz/_astro/cover.CMSoXGtF_Kvont.webp)
    
    myAzymut
    
    Client portal with a network map and shipment tracking
    
    
    
    ](https://wzor.biz/en/projects/myazymut/)
-   [
    
    ![Puls CRM - the wordmark with a blue pulse line above a monitor showing the "Deals" screen: a side menu, the sales pipeline kanban with the columns New, In talks, Proposal and Won, and deal cards with company, value and days in stage; light blue waves in the background.](https://wzor.biz/_astro/cover.B_Q_VQbe_1ERf64.webp)
    
    Puls CRM
    
    CRM app with a kanban sales pipeline
    
    
    
    ](https://wzor.biz/en/projects/puls-crm/)
-   [
    
    ![Flotino - a loan calculator and the "Who is our online loan for?" screen on phone mock-ups, against blurred colourful shapes.](https://wzor.biz/_astro/cover.BIgugYRo_ZeHyg7.webp)
    
    Flotino
    
    Online cash loan service
    
    
    
    ](https://wzor.biz/en/projects/flotino/)
-   [
    
    ![PWS Konstanta - the homepage on a laptop mock-up: a hero "The highest quality of insurance backed by experience" with a photo of a column capital and a "30 years" badge.](https://wzor.biz/_astro/cover.CtYGmRYj_Z2gVAJr.webp)
    
    PWS Konstanta
    
    Website for an insurance broker
    
    
    
    ](https://wzor.biz/en/projects/pw-konstanta/)
-   [
    
    ![Fonia - two phones on a teal background: the cashback app home screen with a 325.53 PLN balance and a list of partner stores, next to a store search with suggestions for Empik, Booking, AliExpress and Play.](https://wzor.biz/_astro/cover.DdIowm6o_Z12riqx.webp)
    
    Fonia
    
    A cashback app for online shoppers
    
    
    
    ](https://wzor.biz/en/projects/fonia/)
-   [
    
    ![EzzyGuide - three phones on a dark background: a sign-up confirmation, guide Małgorzata Iwańska's profile with a star rating, and tour details for "In the Footsteps of the Wrocław Dwarves" with the price and contact details.](https://wzor.biz/_astro/cover.DW1rNgrl_ca7w6.webp)
    
    EzzyGuide
    
    A tours app connecting guides with travellers
    
    
    
    ](https://wzor.biz/en/projects/ezzyguide/)
-   [
    
    ![VERSA - the B2B panel for an automotive parts distributor open on a hand-held tablet: the account overview, a shipment tracking map, the partner programme at 30% off, order history and a list of popular products.](https://wzor.biz/_astro/cover.DWv2CO5n_Z27qwnC.webp)
    
    VERSA
    
    B2B panel with billing, shipment tracking and discounts
    
    
    
    ](https://wzor.biz/en/projects/panel-b2b-versa/)
-   [See what else we build Projects](https://wzor.biz/en/projects/)

## Commission an **AI agent**

One process, a pilot with criteria agreed before it starts, and only then production.

![](https://wzor.biz/_astro/collaborate-square-2.B9Ey0d4M_Zwonzw.webp)

[Let us talk Write to us](https://wzor.biz/en/free-consultation/)

Client voices

## What our clients say

Bosfor Group Agencja XO Media Sellace iLeopard

 ![](https://wzor.biz/_astro/opinia.BrbzyDCf_Z2rQFd5.webp)Maja Wieloch-Silecka Operations Director, Bosfor Group

> “Working with the Wzór team was a genuinely great experience! They always answered our questions quickly and were flexible about our comments. They have no shortage of creative ideas and are at the same time thoroughly reliable and on time. The end result fully met our expectations; we recommend them to anyone looking for professionals who care.”

 ![](https://wzor.biz/_astro/piotr-alberski.B09Ae7iY_PJKgl.webp)Piotr Alberski Chief Operating Officer, Agencja XO Media

> “Our clients expect us to act fast and effectively across a very wide scope - from building a brand, through a website with a store, to running an effective sales campaign. Not an easy task, but doable when you have a tight-knit team and partners who take on challenges on the fly. For us, quality, accountability and commitment are key, and working with UX Agency Wzór gives us that 100%.”

 ![](https://wzor.biz/_astro/mateusz-swol.DkFG1xeC_ZygxNl.webp)Mateusz Swół Chief Operating Officer, Sellace

> “Our cooperation with UX Agency Wzór started with our own website. We really liked the way they work, so we decided to subcontract projects for our clients to them on a White Label basis. We loved their workflow, the way they hand projects over to our team, their documentation and their work on the design system. Now we can take on bigger risks.”

 ![](https://wzor.biz/_astro/dawid-smiglarski.RgmAcSCC_Z1JjuzK.webp)Dawid Śmiglarski Area Manager, iLeopard

> “After migrating our store to PrestaShop and changing the template, we decided to have UX Agency Wzór run a UX audit. It surprised us to learn there were over 40 UX defects in our system, 5 of them critical. The Wzór team not only identified the problems but also set out practical recommendations and concrete examples of how to fix them. At the meeting where we went through the audit we also received valuable input and together chose a new module for the checkout process. We believe the changes will benefit both our customers and the growth of our brand. We thank UX Agency Wzór for their commitment and professionalism, and we will certainly use their services again.”

[Meet our clients (115)](https://wzor.biz/en/clients/)

The team

## Who will build your **AI agent**?

-   ![Aleksandra Bondar](https://wzor.biz/_astro/portrait.D0iAts48_Z2ryhWH.webp)
    
    ### Aleksandra Bondar
    
    Senior UX/UI Designer
    
    [View profile](https://wzor.biz/en/team-and-partners/aleksandra-bondar/)
    
    How can I help you?
    
    > “I will design the conversation with the agent in your design system and test it on a clickable prototype with your staff before anyone writes the backend.”
    
-   ![Katarzyna Adamczuk](https://wzor.biz/_astro/portrait.BFsC1rhN_Z17TUuY.webp)
    
    ### Katarzyna Adamczuk
    
    Prezes, Analityk Biznesowy
    
    [View profile](https://wzor.biz/en/team-and-partners/katarzyna-adamczuk/)
    
    How can I help you?
    
    > “I will work out with you what the process the agent is taking over costs today, so after the pilot there is something to compare the result with.”
    
-   ![Patryk Korycki](https://wzor.biz/_astro/portrait.Cn4p6oJS_Z1ryFk.webp)
    
    ### Patryk Korycki
    
    CEO, Analityk Biznesowy
    
    [View profile](https://wzor.biz/en/team-and-partners/patryk-korycki/)
    
    How can I help you?
    
    > “I will agree with you the scope of the first rollout and the criteria for deciding whether the agent stays. I will coordinate the pilot and the production rollout.”
    

FAQ

## Frequently asked questions about **AI agents**

What is an AI agent?

An AI agent is a program built on a language model that is given a goal, a set of tools and limits on what it may do, and then chooses for itself which tool to reach for to finish the task. In a company it usually takes over repetitive work: answering from documentation, changing appointments, preparing summaries.

How is an AI agent different from a chatbot?

An AI agent differs from a chatbot in that it carries out operations in your system, while a chatbot follows a conversation tree designed in advance. A chatbot ends where its script ends, whereas an agent picks the tool for the task itself and stops when it lacks data or permissions.

Can an AI agent make up answers?

An AI agent can make mistakes like any model, which is why it answers only from your documents and gives the source with every answer. The risk cannot be brought down to zero, so we measure it openly: every release runs through the set of control questions, and the result is a number.

Does our data go to the AI model?

Your data goes to the AI model only to the extent a given task requires, and the rest, personal data included, is filtered out beforehand. The processing location, retention and a ban on training the model on your data are written into the data processing agreement, and if processing has to stay in the EU, we choose a provider that meets that condition.

Which AI model do you use?

We choose the AI model by measuring it in the pilot, not up front: the winner is the one that performs best on your set of control questions at an acceptable cost and latency. The model call layer is kept separate, so switching providers is a configuration change, not a rewrite of the deployment.

Can an AI agent carry out actions in our system?

An AI agent carries out actions in your system through the endpoints you already have, and only from an explicit list of operations it has been cleared for. We do not add new business logic, and every call is recorded together with the conversation that led to it.

How do you build an AI agent for a company?

You build an AI agent for a company starting from one process, not from choosing a model. Then we tidy up the documents it will draw on, build a prototype of the conversation and run a pilot with a narrow group before the agent reaches all users.

Do we need to build an agent from scratch if we already have a chatbot?

A poorly performing chatbot usually does not need rebuilding from scratch, because the problem rarely lies in the model itself. We start with the conversation log, which shows where people give up, and then fix the script, the knowledge sources and the hand-off to a person.

What if the AI agent pilot does not work out?

A failed AI agent pilot ends with a report and a decision, not a request for an extension. We set the criteria before the start, and the report says what was missing: data, a stable process or a return on the cost per conversation, and under what condition the topic is worth coming back to.

![](https://wzor.biz/_astro/portrait.Cn4p6oJS_Z1ryFk.webp)

Patryk Korycki CEO, Business Analyst

[Send us a message](https://wzor.biz/en/contact/)

Related services

## Services related to **AI agents**

### You want to adopt AI but do not yet know _what for_.

We will find the process with you where automation will really pay off.

[Product discovery](https://wzor.biz/en/services/product-discovery/) [UX consulting](https://wzor.biz/en/services/ux-consulting/)

### You have a product the agent is to go into, but _the product itself does not deliver_.

We will check where users get lost before we add an agent to it.

[UX Audit](https://wzor.biz/en/services/ux-audit/) [UX Research](https://wzor.biz/en/services/user-research/)

### You need _a whole system_, not an assistant inside it.

We will design and build software for your process or the front end of a web app.

[Custom software](https://wzor.biz/en/services/custom-software-development/) [Web app development](https://wzor.biz/en/services/web-app-development/)

Next step

## Free consultation

At the meeting we walk through the process the agent is to take over and split it into a part for the machine and a part for people. You get the scope of the first version of your AI agents and the conditions that need to be met on your side. A description of the process in bullet points is enough.

[Free consultation](https://wzor.biz/en/free-consultation/) [Projects](https://wzor.biz/en/projects/)

![](https://wzor.biz/_astro/portrait.Cn4p6oJS_Z1ryFk.webp)

Patryk Korycki CEO, Business Analyst
