Services · Building block

An agent for work that looks different every time.

AI agents read enquiries, look things up in ERP and CRM and create a draft. Anything binding that leaves the company is approved by your team.

Quote agentLog

Draft quote for the enquiry from Brandt Anlagenbau

  1. Mailboxdone09:12:04Enquiry read: 3 items, drawing attached
  2. CRMdone09:12:06Customer Brandt Anlagenbau found, no. 20817
  3. ERPpartial09:12:092 of 3 items matched to the item master
  4. Knowledge systemcited09:12:11Framework contract: 8% discount, source p. 4
  5. Saleshanded over09:12:12Item 3 has no item number: handed to Ms Keller with the current state
May
  • Read mailbox, CRM, ERP
  • Create a draft
May not
  • Send the quote
  • Change prices
  • Cases stop waiting

    The agent works on enquiries as they arrive, even while the experienced colleague is in a meeting. She reviews a draft instead of starting from scratch.

  • Works in your systems

    It reads the mailbox, ERP and CRM and puts the result there. No copying between a chat window and your application.

  • Fixed limits, full trace

    What it may do is set in advance, and every step is logged. Anything binding that leaves the company is approved by your team.

  • Hands over instead of guessing

    If a detail is missing, the case goes to a person with everything found so far. Nobody starts again from the beginning.

Where it gets stuck today.

  1. The case waits for one person.

    Only one experienced colleague can handle quotes or enquiries, because it takes three systems and expert knowledge. When she is in a meeting, the case sits.

  2. The chat that takes nothing off your plate.

    The clerk copies the enquiry into the chat and the answer back into the system. The step that costs the time is still there.

  3. The automation that keeps breaking.

    When an email does not fit the pattern, it lands in an error folder nobody empties. The cases that need judgement stay put.

From enquiry to draft

Its own route per case, every step logged.

The agent reads the enquiry and decides where to look: the customer in the CRM, the items in the ERP, the discount in the framework contract. At the end there is a draft in review, with a note on what is still open.

SalesOutlook
  • Brandt Anlagenbau09:11Enquiry for 3 itemsto agent
  • Lange & Söhne08:47Delivery time querydraft ready
  • Wagner Logistik08:30Price request flangesdraft ready
Quote agentLog

Draft quote for Brandt Anlagenbau

  1. Mailboxdone09:12:04Enquiry read: 3 items, drawing attached
  2. CRMdone09:12:06Customer found, no. 20817
  3. ERPpartly09:12:092 of 3 items matched
  4. Knowledge systemcited09:12:11Framework contract: 8 % discount, p. 4
Quote AN-2026-0412in review
CustomerBrandt Anlagenbau, 20817
Items2 matched, 1 open
Discount8 % per framework contract
Valid until6 Nov 2026
Agent's noteItem 3 has no item number, please add it.
Approve

Limits

What the agent may do is set before it starts.

One task, only the access it needs, and a clear line where it hands over. Anything binding that leaves the company or moves money is always approved by a person.

  • May: read and create drafts

    Read enquiries, look up customers and items, create drafts in review. With the permissions of the person it works for, no more.

  • May not: send or change prices

    Sending, price changes and payments go through approval. That limit is the default and only changes on purpose.

  • Hands over when something is missing

    An item without a number, a special discount, a conflict with the contract: the case goes to sales with what it has. We measure how often it hands over correctly.

Task: draft quoteLimits
  1. MayRead enquiries, look up customers and itemson its own
  2. MayCreate the quote as a drafton its own
  3. May notSend the quote, change pricesapproval only
  4. Hands over onItem without number, special discountto sales
ToolsQuote agent
  1. CRMFind customerread
  2. ERPRead items, create draftread, draft
  3. SendingSend quote to customerno access

Approaches

What an agent can be built with.

There are several common routes. Which one fits depends on where your systems live, who runs the agent afterwards and which model is an option. On every route, connecting ERP and CRM is the real work.

  • Our own agent in TypeScript

    Our default

    Our stack, models from EU endpoints

    Fits when
    the agent should work inside your systems and trigger nothing without approval. It gets its tools per request, and anything with an effect outside needs sign-off. Models come from European providers such as Mistral, IONOS, STACKIT or OVHcloud, or from your Azure, with no fixed vendor.
    Mind
    More building than a kit, but the code and the rules are yours.
  • Microsoft Copilot Studio

    Managed, Microsoft 365 and Power Platform

    Fits when
    you work in Microsoft 365 and the agent should work in Teams or Outlook. Power Platform connectors and sign-in through Entra ID are already there.
    Mind
    Usage-based billing on top of the licences. Data sits in the chosen Azure region; according to Microsoft, some features also store data outside it.
  • OpenAI Agents SDK

    Code framework on the OpenAI Responses API

    Fits when
    the agent should run as its own service and OpenAI models are a given. The SDK brings tools, handoffs between agents and tracing.
    Mind
    Ties you closely to OpenAI interfaces. The older Assistants API was shut down on 26 August 2026; new agents build on the Responses API.
  • LangGraph and similar frameworks

    Code framework, open source

    Fits when
    state, loops and approvals need precise control and the model should stay replaceable. Runs on your own servers.
    Mind
    More development effort, and someone has to look after the code for the long run.
  • Agent nodes in n8n

    Workflow platform, self-hosted

    Fits when
    the agent takes on only the one step that needs judgement inside a fixed flow. The AI Agent node picks from connected tools, the rest stays rules.
    Mind
    Long, branching agent logic gets hard to follow in a workflow editor; then code is the better choice.
  • Model Context Protocol (MCP)

    Open standard for connecting systems

    Fits when
    several agents or AI applications should use the same systems. An MCP server for ERP or CRM is built once and used by every compatible agent.
    Mind
    An MCP server exposes tools. Permissions and approvals have to sit there, not only in the agent.

Tasks and limits.

What the system takes on

  • Reads enquiries, emails and attachments and picks the tools each case needs
  • Looks things up in ERP, CRM and the knowledge system and creates drafts
  • Hands over when information is missing, instead of guessing
  • Logs every step for every case
  • Is measured on real cases: finished drafts, correct hand-overs, false confidence

What your team takes on

  • Sets where the boundary sits and what needs approval
  • Approves anything binding that leaves the company or moves money
  • Takes over every case handed to them
  • Checks every draft at the start; the boundary is adjusted from the corrections
  • Grants the permissions; the agent never has more than the person it works for

AI agents

If a rule decides, it is workflow automation. An agent pays off only when the route differs from case to case, and often it takes on just the one step that needs judgement inside a fixed flow.

Which one when?

  • Workflow automation

    Is the route always the same and a rule decides? Then workflow automation. If only one step needs judgement, the fixed flow gets one agent step.

  • AI knowledge systems

    Should it only answer, with evidence, without acting? Then a knowledge system.

  • AI phone assistants

    Does the case come in by phone? Then an AI phone assistant, the same design on the phone.

The way there.

From the first call to a system in operation. Each phase stands on its own; none ties you to the next.

  1. Intro call

    free

    30 minutes. We work out where AI fits for you and whether an AI strategy is worth it.

  2. AI strategy

    €1,990

    AI strategy for mid-sized companies: which of your processes suit AI, what the build costs and brings back, and which ones are not worth the effort.

  3. Pilot & implementation

    from €5,000

    AI pilot at a fixed price: one use case, fixed scope and acceptance criteria in writing up front. Built in your systems, tested with your data.

  4. Operation & improvement

    from 20% a year

    Running AI systems: hosting in Germany, monitoring, model and prompt maintenance, ongoing measurement, and the next use case on the same foundation.

Further reading

All posts

Frequent questions

Common questions about it.

The questions that almost always come up.

What is an AI agent?

An AI agent is a program built on a language model that gets a goal and decides case by case which tools to use. It reads the enquiry, looks things up in ERP and CRM and creates a draft. What it cannot settle, it hands to a person with what it has so far.

How is an AI agent different from an automation?

An automation runs a fixed sequence: the same steps in the same order every time. An agent gets a goal and decides case by case which tools to use in which order. That is right when the cases differ and wrong when they do not. Often the best answer is a fixed flow with a single agent step.

How is an AI agent different from a chatbot?

A chatbot answers in a chat window. An agent looks up what it needs in your systems and writes the result back there: a draft quote into the ERP, a case into the ticketing system. It has tools and permissions; a chatbot has a text box.

Does an AI agent make decisions that commit us?

Only the ones you hand it. For every agent we set what it may do itself and where it hands over. Anything binding that leaves the company or moves money goes through approval by default, and every step is logged.

What happens when the AI agent does not know something?

It hands the case over with what it has instead of guessing. The boundary where it hands over is part of the design, and we measure how often it draws it correctly.

Does an AI agent work in our existing systems?

Yes. An agent with no connection to your ERP, CRM, mailbox or ticketing system is a chat window you retype things into. The integration is where it is decided whether a case holds up. Your systems stay as they are.

When is an AI agent not worth it?

When the task is rare, a script will do. When no expert knowledge is involved, it is an automation. And when a mistake only shows up months later, because then nobody can check whether the boundary sits right.

Where do AI agents run, and what does the EU AI Act say?

Processing in German data centres or inside your own infrastructure, under a data processing agreement under Article 28 GDPR, with no training on your data. Before go-live we classify every agent under the EU AI Act and document purpose, data sources and limits.

The next step

Let us find out where AI pays off for you.

30 minutes, free and without obligation.