AI agent development
AI agent development: agents that work inside your systems
An AI agent completes tasks: it plans several steps, uses your systems and presents the result for approval. CIRO EFFORT builds such agents for mid-sized companies, from order intake and invoice matching to market and user research. We use them ourselves every day, including in our own development work, which runs two to three times as fast as the traditional way. First clickable version after two to three weeks, fixed price per project phase.
Across Germany for startups and mid-sized companies, with team members in Munich and Dortmund. Updated:
- faster and cheaper than traditional development
- 2–3×
- to the first agent working with your real data
- 2–3 weeks
- by people wherever a mistake would be costly
- Approval
AI agent, chatbot or automation?
The three terms are often confused. Which approach fits depends on your process:
| Chatbot | Rule-based automation | AI agent | |
|---|---|---|---|
| What it does | Answers questions in a dialogue | Runs fixed steps the same way every time | Plans steps itself and completes a task |
| Handling exceptions | Refers to a person | Stops | Adapts, asks back or submits for review |
| System access | Usually none | Hard-wired | Through approved tools with clear permissions |
| Best for | FAQs, first contact | Stable, unambiguous workflows | Workflows with text, variants and judgement |
In practice we often combine them: fixed rules for unambiguous cases, an agent for the cases in between, a person for approval.
The AI advantage
AI agents: two to three times as fast as the traditional way
2–3×
less effort and time than traditional development
For AI agents this means: we build them with the same technology our own agents use every day to test, check and research.
Effort for a new feature
For example invoicing or time tracking in existing software. Index, traditional agency = 100.
- In-house team
- 120
- Traditional agency
- 100
- Offshore development
- 75
- CIRO EFFORT
- 35
What are AI agents suited for?
Good first candidates are workflows that occur often, process text and tie someone up today:
01
Order and enquiry intake
Reading emails and PDFs, transferring data into the ERP, flagging ambiguities and preparing follow-up questions.
02
Invoice and receipt checks
Matching invoices with orders and delivery notes, submitting deviations with reasons for approval.
03
Customer enquiries
Understanding the request, looking up data in the system, drafting the reply. Your team reviews and sends.
04
Research and reporting
Monitoring markets, competitors and tenders, summarising findings and reporting with sources.
05
UX and audience research
Agents with persona profiles explore websites and products from the customer's view, for user testing and audience analysis. More below.
06
Internal assistance
Agents that prepare meetings, gather documents and write minutes into the right system.
AI agents for UX research, user testing and audiences
We use the same agents for clients who want to know their brand, product or audience better:
01
Audience analysis
Which groups look for your offering, with which terms, questions and objections? From search, market and your own data.
02
AI personas
Personas built from data, which your team can question and use to test ideas.
03
Synthetic user tests
Persona agents go through your website or app step by step and log where they get stuck.
04
Brand perception
How does your brand come across to decision-makers, existing customers and new prospects? First impression, value proposition, trust signals.
05
Competitor analysis
How do your competitors position themselves, which questions do they answer, and which gaps can you fill?
06
Content and search analysis
What your customers search for and whether your content answers it, on Google and in AI search engines.
Agents that work for us around the clock
Besides writing code with AI, we run our own agents that test, check and analyse our products and projects. Here is what three of these runs look like:
Agent …
- Owner signs up, code confirmed
- Sets up the company, books 4 seats
- Invites sales, support and finance
- Support answers a customer request
- Permission checker: sales cannot see payroll data
Agent …
- Opens the page on an iPhone
- Checks contrast, focus and reading order
- Runs the statement as a new landlord
- Report with screenshots to the board
Agent …
- Reads website, services and legal notice
- Detects logo, colours and fonts
- Derives the current situation
- Proposes the structure of the pitch page
Swarm tests that behave like real companies
A whole company of agents uses the product like real customers: every role has its own tasks, permissions and routine, from sign-up to the paid invoice. Checkers run alongside for what people easily overlook: a permission checker makes sure nobody sees data that is none of their business, a lawyer reads the copy, a financial auditor the invoices, an editor the language.
Outcome: every error lands as a task with a step log on our board before a customer notices it.
UX analyses in a real browser
Agents operate the interface on phone and desktop as a specific person: here a landlord creating a service-charge statement for the first time. They assess whether every step is understandable, whether feedback gets through and whether the page is accessible.
Outcome: a report with screenshots for each hurdle and a recommendation, prioritised by impact.
Needs analysis before the first call
Before we talk to a company, agents analyse its presence: offering, audiences, brand and visible issues. Our team takes the result into the initial call, so we can skip the standard questions.
Outcome: a starting point and a proposal for pitch page and draft quote, checked and adapted by people.
What happens in an agent run
- 01
Brief
A person defines what to check: a workflow, a page, a company.
- 02
Roles
The agent gets personas with goals, prior knowledge and permissions, such as owner, support or finance.
- 03
Execution
It works in a real browser, on phone and desktop, with real clicks and inputs.
- 04
Log
Every step is recorded with a screenshot and reasoning, traceable for the team.
- 05
Finding
Whatever stands out lands as a task with its log on the board, sorted by severity.
- 06
Review
A person on the team assesses the finding and decides what gets fixed.
- 07
Fix
The fix goes into the code; automated tests run with every change.
- 08
Retest
The agent runs the workflow again until everything is green.
How reliable is synthetic user testing?
Synthetic tests are good for hypotheses and obvious hurdles, but they do not replace real users. Synthetic users tend to be too agreeable and take cleaner paths than people, so we combine both:
| Synthetic user tests | Tests with real users | |
|---|---|---|
| Speed | Hours to days | Weeks, including recruiting |
| Strength | Breadth: many profiles, many paths, early in the project | Depth: real emotions, real behaviour, surprises |
| Weakness | Too polite, too logical, stereotypes for rare groups | Costly, few participants per round |
| How we use it | Forming hypotheses, pre-testing concepts, finding hurdles | Securing decisions before major budget is committed |
Our personas draw on your data in addition to a language model's general knowledge.
How reliable are AI agents?
Staff only use an agent they trust, so every agent we deliver includes:
- Strictly limited tools and permissions: the agent may only do what its task requires
- Human approval for everything that goes out or moves money
- A complete log of every step, traceable for your team
- Testing on real cases before launch, against your quality standards
- Ongoing measurement of error rate, duration and cost in operation
- Protection against prompt injection and rate limiting; our security agents review the code
- A swappable model, so you are not tied to one provider
From workflow to production agent
Each phase has a verifiable outcome and its own fixed price:
- 01Week 1
Understand
Workshop with your team, processes and data. We prioritise use cases by impact and feasibility.
- 02Weeks 1–2
Design
Solution concept, model choice, data protection and AI Act classification. Then the fixed price for implementation is set.
- 03Weeks 2–9
Build
First clickable version with your real data, evaluated against your quality standards, in short cycles.
- 04Weeks 10–12
Deliver
Go-live, training for your team, monitoring. Then operations and ongoing development from one team.
Agents your staff actually use
For an agent to be used in daily work, we design the interface and measurement from the start:
- Consulting
- Which workflow is suitable, what may the agent do alone, where does a person need to approve?
- UX and design
- Agents your staff actually use, because their results are understandable, they ask when unsure and people stay in control.
- Development
- Tools, APIs and permissions that let the agent work in your systems, with strict access controls.
- Analytics
- Every step the agent takes is logged and analysed: cost, duration, error rate.
- AI
- Model choice, planning of multi-step tasks and ongoing evaluation against your quality standards.
What the cost of an AI agent is made of
Three cost components we quantify for your case in the concept:
- Build: analysis, development, integration and testing, as a fixed price per phase
- Operations: hosting, monitoring and ongoing development, on request for a monthly retainer
- Usage: AI model costs per processed task, depending on volume and model choice
AI agents FAQ
What companies most often want to know before their first agent:
A chatbot answers questions in a dialogue. An agent receives a task, plans the steps itself, uses approved tools such as your ERP or mailbox and delivers a result a person can check.
Synthetic users are AI agents with a persona profile: goals, prior knowledge, objections and typical behaviour of an audience. They receive a task, such as finding an offer or sending an enquiry, and document every step. That reveals hurdles before real users experience them.
No. Synthetic users are fast and broad, but too polite and too logical: they agree more readily and find paths where real people fail. We use them for hypotheses and early tests and verify important decisions with real users from your audience.
The cost consists of build, operations and usage. We charge the build as a fixed price per project phase, operations on request as a monthly retainer, and AI model usage depends on how many tasks the agent completes. In the concept we calculate all three components for your volume before you commission implementation. We name concrete figures in the initial call.
You decide. We recommend that everything that goes out or moves money is approved by a person, at least at the beginning. As experience grows, approvals can be relaxed for unambiguous cases. Every step the agent takes is logged.
Yes, through their APIs. The agent receives only the tools and permissions its task requires. Where a system has no suitable API, we build a separate, isolated intermediate layer.
After two to three weeks a first version works with your real data, initially in a test run alongside the existing workflow. Production use typically follows after two to three months.
For some cases that is right, and then we say so. Ready-made platforms hit limits once the agent needs to go deep into your own systems, your data should not leave the EU, or per-user costs grow with the team. Your own agent belongs to you, including code and configuration.
Yes, like any AI system. Agents for office workflows are generally not high-risk systems, but transparency obligations and the AI literacy obligation for staff apply. We classify every agent at project start and document it in writing. As of October 2026, not legal advice.
Let's talk about your project.
We take care of strategy, design, development and data. Your intro call is directly with the founders: 45 minutes, no obligation.
Or directly: 089 24 88 31 22
