Pillar 3 · AI Agents

AI Agents

An AI agent is not a chatbot. It is a goal-oriented system that completes multi-step tasks on its own, combining several sources along the way and delivering a traceable justification for every decision. In other words: an agent does what, until now, only staff with context and judgment could do — within clearly defined rules of engagement.

In growing companies, agents are interesting wherever processes are too complex for classic workflow automation, but too repetitive to permanently tie up experienced staff. Supplier onboarding, quote preparation, contract data extraction, escalation triage — anywhere information from several systems has to be brought together and assessed.

We build agents with three hard boundaries: defined tool permissions, clear escalation rules and a complete audit log. No agent may do anything that has not been approved. No agent runs without oversight. No case disappears into a black box.

What an AI agent can do — and what it is not

An agent can

  • Bring together data from several systems
  • Make decisions with a stated justification
  • Sequence multi-step tasks in the right order
  • Escalate to a person in unclear situations
  • Keep a record of its own actions
  • Learn from corrections without changing the rules of engagement

An agent is not

  • A chatbot with a personality
  • An autonomous decision-maker without oversight
  • A black box without an audit log
  • A substitute for every kind of human judgment
  • A one-click tool with no implementation effort
  • A solution that looks the same in every industry

Concrete use cases in practice

Supplier onboarding

The agent checks master data against official registers, requests credit reports, collects certificates, matches them against internal requirements and only creates the supplier in the ERP once every mandatory criterion is met. In case of deviations, it hands the case to procurement management with a clear justification.

Quote preparation

Incoming inquiries are classified, matched against the product catalog, linked to current stock levels and lead times, and condensed into a finished draft quote. Sales checks, adjusts and approves it — instead of writing every quote from scratch.

Contract data extraction

Clauses, deadlines, obligations and key figures are extracted from incoming contracts, classified and entered into the DMS or contract management system. Critical clauses automatically trigger a notice to the legal department.

Escalation triage in service

Complex service cases are assessed based on contract data, history and severity, routed to the right escalation level, and prepared with all relevant information from various systems. The responsible employee starts with full context.

Master data maintenance and data quality

The agent monitors master data in ERP, CRM and DMS, identifies duplicates, missing mandatory fields and inconsistencies, proposes corrections and carries out approved corrections automatically. Data quality becomes an ongoing discipline instead of a quarterly project.

How we build AI agents

01

Define the task

We describe the agent's goal in one sentence, list the tools it is allowed to use, and define the boundaries of its autonomy. Without these three points, we do not start building.

02

Connect and calibrate the tools

The agent gets access to exactly the APIs, data sources and actions its task requires. Every permission is explicitly documented. No "admin access to everything." We also deliberately calibrate the model parameters for each use case: an agent that reviews contracts needs maximum accuracy, not creativity — an agent that drafts quote text works with different settings.

03

Escalation and audit

We define in which situations the agent hands off to a person, and make sure that every decision lands in the audit log with its inputs, justification and outcome — traceable and auditable.

Typical result pattern

What a properly built AI agent can achieve

This is not a single case study, but a typical pattern from our project work. It summarizes the order of magnitude that is realistic when an agent is deployed for a clearly defined, multi-step process in a company.

Typical pattern · B2B company · Supplier onboarding with an AI agent
Starting point

The purchasing team checks new suppliers manually against registers, credit reports and internal approval rules. Onboarding takes two to three working days, master data gets typed in multiple times, and deviations trigger back-and-forth over several email loops.

Solution

An agent with clearly defined tool permissions checks the commercial register, sanctions-list and credit-rating sources, consolidates the results, pre-fills mandatory fields in the ERP and only escalates borderline cases to procurement management. Every step lands in the audit log with a justification.

Result

Onboarding time typically drops from two to three days to a few hours, master data quality improves measurably, and procurement management only reviews the cases where human judgment is genuinely needed.

~ 80 %
fewer manual steps
2–3 d → h
turnaround time per case
100 %
audit log coverage

Example from a typical project situation. Concrete figures are measured individually in every project.

Related pillars

AI agents rarely stand alone. They access existing systems and are typically combined with the following pillars:

ERP and CRM integration →

So the agent finds the data it needs to work with.

AI process automation →

When the use case is more linear than goal-oriented.

Document automation →

For processing the incoming documents that many agents need.

Managed AI operations →

So the agent runs permanently and under supervision in production.
Guide

AI agent or classic workflow? →

The decision to make before you build: when an agent is the right architectural choice — and when a classic workflow stays the more robust tool.

Frequently asked questions about AI agents

What is the difference between an AI agent and a classic workflow?
A classic workflow follows a fixed path: step one, step two, step three. An AI agent pursues a goal and decides for itself which tools to use in which order. It can query systems, combine information, justify decisions and trigger follow-up steps — within clearly defined permissions and with a traceable audit log.
How do we retain control over an autonomous agent?
Through three mechanisms: first, clearly defined tool permissions (the agent may only carry out certain actions in certain systems); second, escalation rules (for cases outside the rules, the agent hands off to a person); third, a complete audit log with a justification for every decision. You can see at any time what the agent did and why.
What are specific use cases for AI agents in companies?
Especially for multi-step tasks that need context from several systems: supplier onboarding (verifying master data, checking creditworthiness, assembling contracts), quote preparation (needs analysis, product selection, price calculation), contract data extraction (identifying clauses, capturing deadlines, entering them into the DMS), or escalation triage in complex service cases.
Can AI agents make mistakes or hallucinate?
Yes, AI models can produce results that sound plausible but are factually wrong. That is exactly why our agents work with three safeguards: first, clear tool permissions that limit the scope of action; second, defined escalation rules for cases where the agent is uncertain; and third, a complete audit log that documents every input, every decision and every justification. This does not eliminate hallucinations entirely, but it makes them visible, bounded and traceable.
What is the difference between a chatbot and an AI agent?
A chatbot answers individual questions in a conversation. An AI agent completes multi-step tasks with context and its own responsibility — it reads inputs, combines multiple sources, calls systems, makes justified decisions and documents every step. The main difference is not the interface, but the scope of the task: a chatbot answers, an agent gets it done.
When is an AI agent the wrong choice?
If the task is rule-based and well documented, a classic workflow is often better, faster and cheaper. If the consequences of a wrong decision are very high and human sign-off cannot easily be integrated, an agent is likewise not a good choice. AI agents shine where tasks remain multi-step, context-dependent and verifiable through audit log and escalation — not everywhere.

Do you have a multi-step process in mind?

In a short conversation, we assess whether an AI agent pilot makes economic sense — and what the first 60 days would look like in practice.

Request an AI agent pilot