Choosing the process
Together we choose a process that repeats often, has a clear "done" criterion and has its data in your systems. We define what the agent can do on its own and what only after a person approves.
Claude agents
An agent is not a chat window but a coworker without hands: it receives a task, gathers data from your systems, takes action and reports back. We implement Claude agents with clear limits, logs and human approval where there is risk.
Built for ISO 27001NIS2 DORA
- Sales enquiry agent
- Customer service agent
- Invoice processing agent
The agent works, a human approves
- 1 process first agent in 2–4 weeks
- Limits permitted actions and human approval
- Log every agent action is recorded
- Claude Anthropic API, Model Context Protocol (MCP) and Claude Agent SDK
An assistant answers a question. An agent does the work: it reads the request, checks the CRM, calculates a quote, drafts an email and creates tasks for the team. The difference is that an agent can act in your systems, so it needs not only a model but also rules: what it can do on its own, what only with approval, and where everything is recorded.
We build Claude agents with the Claude Agent SDK and tool use, MCP servers for your systems and long context for large documents. Every agent has a defined purpose, a list of permitted actions, least-privilege access and an action log that can be reviewed. We start with one process, measure the result and only then expand.
Use cases
The scenarios we implement most often with Claude. Each one starts with a single process and is measured.
Sales enquiry agent
Receives an enquiry from email or a form, checks the customer in the CRM, drafts a quote from your price list and creates a task for the sales rep. Claude works from your price list, not from the internet.
Customer service agent
Classifies requests, answers standard ones from the knowledge base and hands complex ones to a person with the information already gathered and a suggested reply.
Invoice processing agent
Reads incoming invoices, matches them against orders, flags discrepancies and prepares entries for accounting. The accountant approves instead of keying in data.
IT support agent
Handles first-level requests such as password resets, access requests and standard faults, following your procedures, and creates a task for IT when it cannot resolve the issue itself.
Procurement comparison agent
Collects supplier quotes, compares prices and terms in a single table and gives a recommendation with sources.
Meeting agent
Builds a list of decisions and tasks from a meeting recording or notes, assigns owners, creates the tasks and reminds people of deadlines.
Compliance evidence agent
On a schedule, collects ISO 27001, NIS2 or DORA evidence from your systems (access reviews, backup reports, updates) and files it in the audit folder.
How we implement it
From the first conversation to a working solution, with measurement and ongoing support.
Access and limits
We create a separate identity for the agent with the minimum permissions it needs, turn on the action log and add approval steps for risky actions such as payments, sending to customers and deletion.
Build and test
We build the agent with the Claude Agent SDK and tool use, MCP servers for your systems and long context for large documents. We test with real but historical cases and measure accuracy until the result meets the agreed threshold.
Launch and support
We launch with human oversight, review the logs and adjust the rules. Once the first process runs reliably, we move on to the next.
Frequently asked questions
Can an agent do something irreversible?
Only if it is configured that way. By default, risky actions such as payments, sending to customers and deletion require human approval, and the agent has only the permissions its task needs.
Where does the agent run and where does the data go?
Through the Anthropic API or Amazon Bedrock in an EU region (Frankfurt), with your keys and logs. On commercial plans and through the API, your data is not used to train models; with Amazon Bedrock the model can run in an EU region, and access keys and logs stay under your control.
How long does the first agent take?
Usually 2–4 weeks from process selection to launch with support. After that each new agent is faster, because access, logs and rules are already in place.
How much does it cost?
You pay by usage (API) or per user (Claude Team / Enterprise); we help choose the cheapest combination for your workload. We give the implementation price after the first conversation, once it is clear how many systems need to be connected.
Related services
Claude integrations
AI is most useful not in a separate window but in your systems: ERP, CRM, helpdesk, document management.
Claude assistants
An assistant that knows your policies, price lists and contracts, not the internet.
Process automation with Claude
We automate what repeats: invoices, orders, reports, contract reviews.
Expertise confirmed by certifications
This service is delivered by specialists with valid international certifications. Their expertise is confirmed by independent organisations, not by us.
-
Artificial Intelligence Security Professional
AI systems are deployed with protection against data leakage and model abuse.
-
ISACA Advanced in AI Security Management
The AI security programme and its controls are managed following the ISACA methodology.
-
ISACA Advanced in AI Risk
AI risks are identified and managed like any other IT risk.
Get a quote for agents
Tell us which process or system you want to improve with Claude, and we will send a quote within 1 business day.
We reply within 1 business day