Summary
The best first AI use case is clearly defined, frequently recurring, based on data and measurable in business terms. An AI agent supports the preparatory work; people retain responsibility and approval authority.
THE CENTRAL IDEA
Do not start by asking how AI should transform the entire company. Start by asking which routine tasks are unnecessarily holding back your best employees today.
AI has reached the Mittelstand, but the right starting point remains unclear
Artificial intelligence is no longer a distant prospect for German companies. According to Germany’s Federal Statistical Office, 26 per cent of companies in Germany were already using AI technologies in 2025. At the same time, recent business surveys show that many companies are still looking for a viable path from experimentation to productive use.
Hidden champions and industrial family businesses face a particular tension. They have deep expertise, complex products and established processes. Yet their resources are limited, sensitive customer and engineering data require particular protection, and experienced employees are difficult to replace.
It is tempting to talk immediately about a comprehensive AI strategy, a new platform or a transformation of the entire company. That sounds ambitious, and often expensive. In practice, starting too broadly quickly leads to lengthy planning phases, unclear responsibilities and promises of value that are difficult to measure in everyday operations.
KEY TAKEAWAY
A good start with AI is not a technology project built for some future need. It is a targeted solution to a problem that already costs time, money or quality today.
Why many AI projects start too big
Many companies start with the technology: Which language model is best? Do we need a platform? Should the solution run in the cloud or in our own data centre? These questions matter, but they come too early if the specific work to be improved is not yet clear.
Starting too big usually has four consequences:
- The benefits remain abstract because too many processes are considered at once.
- The organisation initially experiences AI as additional work rather than relief.
- Employees are asked to document data, rules and knowledge before they feel any benefit.
- The investment is difficult to assess against a clear business metric.
A better approach reverses this order: first, select a narrowly defined workflow that creates a burden. Then build a solution whose contribution is visible in everyday work. Only once it works do you move on to the next use case.
The right starting point: a specific task that consumes too much time
Time-consuming tasks are activities that are necessary but tie up a disproportionate amount of valuable working time. They include searching for information, transferring it between systems, comparing documents, performing recurring checks, monitoring deadlines and preparing standard texts.
These tasks are particularly promising for AI agents. An agent can gather information from approved sources, structure a case, identify missing details and prepare a draft. It does not operate independently of the organisation, but within defined permissions, rules and approvals.
Five characteristics of a good first AI use case
- It occurs frequently. The more regularly the task occurs, the sooner measurable benefits emerge.
- It involves substantial routine work. A significant share of the work consists of searching, sorting, transferring, comparing or drafting.
- The result can be checked. An accountable employee can review the proposal professionally and approve it.
- The necessary information is essentially available. It may be scattered, but it exists in accessible, approved systems or documents.
- Success is measurable. Handling time, turnaround time, errors, follow-up questions, completeness or capacity can be compared before and after the pilot.
Not every process is suitable as a first step
A process is unsuitable if it occurs rarely, relies almost entirely on undocumented expert judgement, produces results that cannot be verified or would immediately require far-reaching autonomous decisions. Even a technically spectacular use case is a poor starting point if its business contribution remains unclear.
Why quotation preparation is a particularly good place to start
In many technically focused midsized companies, quotation preparation combines precisely the characteristics that make a useful pilot process. It occurs regularly, connects many sources of information and ties up qualified employees with recurring preparatory work.
A good quotation still requires experience. Customer requirements must be understood, technical solutions assessed, risks identified and commercial decisions made. Yet before this specialist work begins, people often have to search for and gather:
- comparable previous quotations and projects,
- current product data, prices and lead times,
- technical specifications and approved text modules,
- unresolved or contradictory customer requirements,
- contract deviations and required approvals.
A quotation agent can take on this preparation. It reads a customer enquiry, structures the requirements, finds relevant information and prepares a sound basis for further work. The sales engineer reviews the technical solution, adds experience and decides on the final quotation.
A CLEAR DIVISION OF RESPONSIBILITIES
AI handles searching, sorting and preparation. People handle understanding, decisions and responsibility.
How a quotation agent can work in practice
- An enquiry or invitation to tender arrives by email or through the CRM.
- The agent identifies requirements, deadlines, open questions and contradictions.
- It searches approved documents and systems for suitable products, previous quotations and comparable projects.
- It uses current master data and flags information that needs confirmation.
- It prepares a quotation draft, including source references and required approvals.
- The responsible employee reviews, amends and approves the result.
The ERP remains the reliable system for master data, orders, costs and approvals. The AI agent complements it wherever information from different sources needs to be combined and understood in context.
Business benefits must be measurable from the outset
A pilot project should not be assessed using general statements such as “more innovation”. A reliable baseline is established before the start. Afterwards, the company checks whether the agent actually reduces work and improves results.
Useful metrics for the pilot
- Handling time per quotation
- Time spent searching for information and coordinating internally
- Time to the first substantive customer response
- Number of enquiries handled by the existing team
- Completeness of quotation documents
- Number of follow-up questions, corrections and data transfer errors
- Compliance with internal review and approval steps
Even a limited pilot can therefore support a clear investment decision: Does the use case save enough time? Does it improve quality? Do employees accept the support? Can the necessary systems be connected securely? Expansion follows only after this assessment.
Employees are not replaced; their time becomes more valuable
Acceptance of AI is determined in everyday work, not in a presentation. If employees initially experience AI as another documentation obligation, a monitoring tool or an abstract programme for the future, resistance develops. If an agent takes over an unpopular routine task, the benefit becomes immediately tangible.
The aim is not to bypass expertise. On the contrary: qualified employees should gain more time for customers, technical solutions, negotiations, improvements and difficult decisions. Their corrections and approvals also help make verified company knowledge useful for future cases.
KEY TAKEAWAY
Less routine work reduces costs. More time for the real task improves quality, speed and satisfaction.
Secure AI in the company: control is part of the design
For industrial midsized companies, technical AI performance alone is not enough. The solution must fit the company’s protection requirements. Customer enquiries, cost calculations, technical drawings, contract terms and experience are among the business’s most valuable information assets.
A robust solution should therefore meet clear requirements even during the pilot:
- Access only to approved data sources and in accordance with existing user permissions
- Traceable sources for significant statements and suggestions
- Logged processing steps and human approvals
- No use of confidential company data to train public models
- A deployment model appropriate to the data classification; for particularly sensitive information, potentially entirely within the company’s own infrastructure
- Clear rules defining which tasks the agent may prepare and which decisions remain with people
On-premises AI or an isolated private environment can make sense for companies with demanding requirements for data sovereignty, technical confidentiality and independence. What matters, however, is not a slogan but a verifiable security and permissions concept.
Seven steps from a time-consuming task to a productive AI agent
- Identify time-consuming tasks: employees name the activities that regularly waste valuable time today.
- Define the scope: select a clear subprocess rather than an entire department.
- Measure the baseline: record time, quality, errors and follow-up questions before the pilot.
- Clarify data and rules: define sources, permissions, approvals and responsibilities.
- Run a pilot with real cases: under controlled conditions and with human review.
- Evaluate benefits and acceptance: bring together metrics and user feedback.
- Only then expand: improve the process or tackle the next suitable time-consuming task.
From the first agent to a learning organisation
A successful quotation agent is initially a practical aid for a specific task. Yet it also creates a reusable foundation in the background: secure access, connections to ERP, CRM and documents, defined approvals, verified knowledge sources and traceable handling of corrections.
Another agent can build on this, for example for complaints and warranty cases. While the quotation records what was promised to the customer, the complaint shows what became of that promise. Connecting both perspectives enables the company to learn systematically from completed orders.
Further use cases can follow later for contract review, order clarification, supplier enquiries, technical documentation, certificates or quality checks. Each new agent must first justify itself economically. At the same time, it extends the shared, secure foundation for knowledge and processes.
THE LONG-TERM PERSPECTIVE
A company does not become an AI champion through a single large project. It gets there by eliminating one worthwhile time-consuming task after another, while making ever better use of its knowledge.
Checklist: is your company ready for its first AI agent?
- We know a frequent process that ties up qualified employees with routine work.
- We can clearly describe the result the agent should prepare.
- An accountable person can review and approve every result.
- The most important information is available in accessible systems or documents.
- Data access, roles and confidential content can be clearly separated.
- We can compare at least two metrics before and after the pilot.
- The employees affected are involved in selection, testing and improvement.
If five or more statements apply, there is usually enough substance for a limited pilot. If fewer apply, the next step should not be an immediate platform purchase. It is worth clarifying the process, data situation and responsibilities first.
Frequently asked questions about introducing AI in midsized companies
Does a company first need to develop a complete AI strategy?
No. It needs clear boundaries for security, responsibility and objectives. Productive use can nevertheless begin with a narrowly defined use case. The wider strategy develops from real experience.
How expensive is it to get started with AI?
That depends on the process, data sources and integrations. A pilot should be deliberately limited so that its benefits can be tested against manageable costs. Getting started becomes expensive above all when scope and success criteria remain unclear.
Does a midsized company need its own AI team?
Not necessarily. The business function and IT must jointly take responsibility for the process, data access and approvals. Specialist implementation partners can provide the technology and integration; process knowledge and decision-making authority remain within the company.
Can AI work directly with the ERP?
Yes, provided suitable interfaces, permissions and controls are in place. The agent should complement the ERP, using information or initiating prepared steps within defined rules, rather than replace it.
When does on-premises AI make sense?
When particularly sensitive customer, product, development or costing data is processed, when independence requirements are high, or when internal policies rule out external processing. The specific architecture should follow from the data classification.
Which AI use case is suitable after quotation preparation?
Complaints and warranty cases are often a useful next step. They also require extensive information searches and connect the original customer promise with actual experience from delivery, operation and service.
Conclusion: start with relief from routine work, not transformation
Hidden champions do not need to know on day one how their company will work with AI in five years. They need to find a first use case that makes a real difference today.
Quotation preparation is often ideal: it connects customer needs, technical knowledge and commercial decisions. A quotation agent can prepare the routine work without taking responsibility. Its benefits can be measured, employees experience tangible relief and existing systems remain in place.
When that first step works, the next one follows. The result is a series of economically justified improvements, rather than an expensive transformation built for an unspecified future. Almost alongside this, a secure foundation grows that preserves company knowledge, connects processes and strengthens long-term competitiveness.
NEXT STEP
You do not have to start with the entire company. You have to start with the right time-consuming task.
