AI Adoption & Operations

Embedded AI in Mid-Sized Companies: Where AI Belongs in Existing Software

Marin Niehues5 February 202611 min read

Code for data processing

In mid-sized companies, AI belongs in the software your employees already use, at the point in the workflow where a decision is made or a document is created. Mid-sized companies run long-established applications for order processing and financial accounting, and that is where value can be demonstrated fastest. A separate AI interface alongside them creates a break in the workflow. For recurring tasks in the core process it loses out to a function in the familiar screen.

Where embedding touches contracts with software vendors or personal data, involve legal and data protection specialists. This article provides an overview and does not constitute legal advice.

How many mid-sized companies use AI?

According to the German Federal Statistical Office, 26 percent of companies in Germany with 10 or more employees used AI in 2025, compared with 57 percent of companies with 250 or more employees. The gap shows where the Mittelstand stands: smaller companies use AI far less often than large ones. The reasons lie less in interest than in prerequisites, above all skills and data maintenance.

Mid-sized companies bring two advantages to embedding: a manageable software landscape and short decision paths. The bottleneck is skills and data maintenance, and that is exactly where the vendor function helps. It needs no development team of your own and works with the data already maintained in the application.

What embedded AI means in operations

Embedded AI refers to AI functions that work as part of an existing application. The clerk opens the same screen as before and finds a suggestion or summary there, which she checks and accepts. The model in the background remains invisible to her, and only the result appears in the familiar workflow. The article Embedded AI: building AI directly into products and processes describes how embedded AI differs from a standalone assistant.

Standalone AI tools require an extra step: switching application, pasting the context, copying the result back. For open tasks such as research or drafts this is acceptable. For a task that runs a hundred times a day in the same screen, every extra step costs time and reduces usage once the initial curiosity fades.

Embedded AI moves the function to where the work happens. This also changes measurability: the application knows the case and the person handling it, so lead time and correction rate can be recorded per step. That provides the data basis for the business case executive management needs for follow-on investment.

Which ERP and CRM systems already have AI built in?

The major ERP and CRM vendors now ship AI functions as part of their products, including SAP, Microsoft with Dynamics 365, Salesforce, HubSpot and Odoo. Vendors widespread among German mid-sized companies are also building assistants into their solutions. Scope, language support and licensing differ widely and change with every release.

Regardless of vendor, five groups of functions recur:

  • Document capture: invoices, delivery notes and orders are read and pre-filled into screens.
  • Text generation: quotes, customer replies and meeting notes are drafted within the case.
  • Summarization: customer history, tickets and long cases are condensed into a few sentences.
  • Forecasting: demand, payment receipts or churn risk are estimated from the system's data.
  • Natural-language assistant: staff query analyses without building a report.

Check three points for your system before activating a function. Is the function included in your existing license, or does it cost extra per user? Where is the data processed, and does the contract exclude its use for model training? Does it work with your customized screens and fields? If your ERP system runs an older version without these functions, the article Legacy systems and AI describes the route via an interface layer, without replacing the system.

The embedding grid: four questions per use case

Before deciding on a function, a grid of four questions has proven itself. It ranks use cases by their delivery impact and steers attention away from demonstrations that play no role in the workflow.

  • Place in the workflow: at which step do waiting times or rework arise today?
  • Frequency: how often is the step performed, and by how many people?
  • Decision weight: what are the consequences of an error, and who checks the result before it takes effect?
  • Data situation: is the required information available in the application itself?

Use cases with high frequency and good data, whose errors carry little weight, come first. A typical example is pre-filling entry screens from incoming documents, with the clerk seeing every result before accepting it. Cases with high decision weight follow once review procedures are in place and metrics are available.

Three routes to embedding

Three routes are available, differing in speed and controllability.

  • Activate the vendor function: the ERP or CRM vendor supplies the AI function as a module. Getting started is quick, adaptation to your own processes remains limited.
  • Extension via interfaces: an external model is connected to the application and returns results into the existing screen. The effort is higher, and so is controllability.
  • Own component: for a process that makes up the company's competitive advantage, a specific function is developed.

Which route fits follows from how specific the process is and how much control the organization must retain. The article Make or buy: developing, buying or adapting AI covers the underlying logic. In mid-sized companies a sequence has proven itself: the vendor function provides the proof, and extension via interfaces follows where the function reaches its limits.

How much does embedded AI cost for a mid-sized company?

For a mid-sized company, embedded AI costs mainly licenses, implementation effort and operation, with implementation effort lowest for activated vendor functions. Vendor functions are billed as an add-on license per user or by consumption. Extensions via interfaces and own components add development effort and ongoing model costs. External consulting in the German market is typically billed at around 1,000 to 2,500 euros per consultant day, depending on seniority.

Is there public funding for AI in mid-sized companies?

Yes, the federal government, the German states and the EU fund AI in mid-sized companies, mainly through free advisory services and grants for projects with a development component. The Mittelstand-Digital centers of the Federal Ministry for Economic Affairs offer small and medium-sized enterprises free information and support for first AI projects. For projects with genuine research and development character, the Central Innovation Program for SMEs (ZIM) and the research tax allowance come into question. Simply activating a vendor function does not count as development there.

In addition there are state programs, often via the regional development banks, and European Digital Innovation Hubs. Conditions and budgets change regularly. The federal funding database shows the current position per state, and checking before the project starts is essential, because many programs only fund projects that have not yet begun.

What the organization must deliver

With vendor functions and existing interfaces, the technology is solved in weeks. Time is lost in the organization: the use case has no accountable product owner in the business unit, the review of results is unregulated, or nobody measures whether the step actually gets shorter. With legacy systems lacking a modern interface, the connection takes months.

The article From pilot to scale describes the whole path from first use case to regular operation.

Niehues Consulting as partner for embedded AI

Niehues Consulting supports companies in this undertaking from a delivery perspective. Marin Niehues builds organizations that deliver and brings AI into operation. Projects range from VC-funded scale-ups to DAX 40 corporations, with assignments in 20+ countries on six continents.

The approach follows the Niehues Delivery Model with its phases understand, build and roll out. Three of its review questions matter most for embedded AI. In understanding: what is the baseline, and how do we measure success? In building: does the prototype prove the value in the real workflow with real data? In rollout: are users enabled and processes adapted? Where the organization cannot staff governance itself, Niehues Consulting takes it on as delivery manager or product owner. The starting point is a free 30-minute initial consultation via the contact page.

The first steps

A robust start follows a fixed sequence. First, the embedding grid is used to select a single use case that sits in an existing application and is performed frequently. The team then measures the current state of the step over several weeks, so that the later value can be evidenced. A prototype in the real workflow follows with a small group of users, whose feedback sharpens the acceptance criteria. Rollout comes last, with defined operation and training for the business units.

Frequently asked questions

What does embedded AI mean?

Embedded AI means AI functions built into an existing application that deliver suggestions, summaries or pre-filled fields within the familiar workflow. Users work in the same interface as before and check the result before accepting it.

Which use cases suit embedded AI in mid-sized companies?

Suitable are steps with high frequency, low error weight and good data, such as capturing incoming documents or preparing customer replies. What matters is that the required information already sits in the application.

How long does it take to add AI to existing business software?

A vendor function is activated and tested in a prototype within a few weeks. An extension via existing interfaces takes two to three months to regular operation, and legacy systems without an interface add months for access. Most time goes into clarifying responsibilities and measuring the baseline.

Do I need to replace my ERP system to use AI?

No. Current ERP systems come with AI functions, and older ones can be connected via an interface layer or data replication. Replacement only pays off if the system is at the end of its life for other reasons.

What percentage of SMEs use AI?

According to the German Federal Statistical Office, 26 percent of companies with 10 or more employees used AI in 2025, compared with 57 percent of companies with 250 or more employees. Smaller companies therefore lag well behind. Companies cite a lack of expertise as the most common obstacle.

When does developing your own AI pay off?

An own component pays off when the process is specific and makes up a competitive advantage, or when the vendor function does not provide the required control. In all other cases the vendor function leads to measurable value faster.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation