AI Adoption & Operations

Building AI Assistants for Business Functions

Marin Niehues26 February 202611 min read

Marin Niehues speaking with a microphone next to a presentation slide on data classification in front of an audience.

An AI assistant for a business function takes shape in three steps: define a narrow task, build it inside the real workflow, and hand it over to operations with clear responsibilities. This article describes the approach from a delivery perspective, sets out timelines and cost drivers, and names the criteria executives and heads of department use to test each step.

Because internal assistants work with company data, they touch on data protection, the EU AI Act and employee co-determination. This article provides an overview and does not constitute legal advice.

What is an AI assistant for a business function?

An AI assistant for a business function is an AI application that takes over a defined step in a department's workflow, using that department's knowledge and documents. It knows the department's terminology and delivers results in the format the specialists use next. Typical tasks include answering inquiries on the basis of internal policies, preparing quotations and summarizing cases for the next decision.

The difference from a general chat tool such as ChatGPT lies in its tie to the workflow. A general tool answers any question. A business function assistant completes one step in a process that starts with a known input and ends with a verifiable result. The value follows from this tie, and so do the requirements for building and introducing the assistant.

A field study of more than 5,000 customer support agents shows that the value is measurable. An AI assistant that provided suggested responses raised the number of issues resolved per hour by 14 percent on average, and by 34 percent for less experienced agents. The effect arose because the assistant was embedded in a fixed workflow.

Step 1: Narrow the task

The most common reason for missed expectations is a task that is too broad and has no measurable business value. “An assistant for sales” does not yet describe a workflow. “The first reply to technical queries from existing customers” does, with a clear trigger, a defined input and a verifiable result.

Four test questions help select the first task:

  • Frequency: does the task occur often enough for a time saving per case to be noticeable?
  • Input: is the information the assistant is to work with available in digital form?
  • Verifiability: can a specialist assess the result within a few minutes?
  • Ownership: is there a person who decides on the result and approves it?

Each candidate is rated yes or no. Tasks with four yeses go to the front of the queue. In the event of a tie, the head of department decides which workflow is currently the biggest bottleneck.

Step 2: Build inside the real workflow

The prototype is built where the work will later take place. Real cases from recent months serve as test cases, the testers carry out the workflow every day, and the quality benchmark is fixed before the first test. A quality benchmark can be the share of answers a specialist adopts without changes, or the time to a finished result. How quality can be secured through test sets and ongoing monitoring is described in the article Securing the Quality of AI Outputs.

Three building blocks determine the quality of an assistant:

  • Knowledge base: the approved documents from which the assistant answers, with an owner responsible for keeping them current.
  • Instruction: the assistant's role, output format and the limits of its remit, written down and maintained under version control.
  • Feedback: a simple way for specialists to flag incorrect answers, so that individual cases become improvements.

An assistant gains access to company knowledge through retrieval augmented generation (RAG). For each request, the method retrieves the relevant passages from the approved knowledge base and passes them to the language model as its basis. The assistant thus answers from the department's own documents and can cite the source of every statement. The technique is explained in detail in the article on retrieval augmented generation.

The prototype phase ends with a decision. If the assistant meets the agreed benchmark on the test cases, it moves into an accompanied pilot with a larger group. If it falls short, the task is narrowed or the knowledge base is reworked before further investment.

Adoption by the business function is the real acceptance criterion. An assistant that specialists work around in daily practice has missed its purpose, even if the test scores are right. Acceptance therefore lies with those who will later work with the result.

Step 3: Hand over to operations

Between a working prototype and reliable operation lies a distinct piece of work. It begins with clarifying roles. The business function provides the business owner, who prioritizes requirements and approves changes. A technical team, internal or external, runs the application and resolves incidents. A metric, such as the usage rate or the time per case, is reported monthly.

Enabling the specialists determines whether the assistant is adopted, and it is also an obligation: the EU AI Act requires companies that deploy AI to ensure a sufficient level of AI literacy among their staff. Specialists need to know what the assistant is meant for, where its answers must be checked and to whom they report errors.

The technical transition into regular operation follows the requirements for any production application. These include quality monitoring and a procedure for changes to the instruction and knowledge base. A clear route for incident reports completes the picture. The full path from prototype to operation is described in the article From pilot to scale.

Clarify data protection and the works council early

An internal assistant can be operated in compliance with data protection rules when the provider and data flows, including the storage location, have been reviewed and contractually agreed. This includes a data processing agreement, clarity on whether inputs are used to train the model, and a decision on which data classes belong in the assistant.

In Germany, the works council must be involved as soon as an assistant is capable of monitoring the behavior or performance of employees. For AI applications that log user inputs, this is regularly the case. Informing the works council as early as the task definition phase prevents the pilot from having to wait for an agreement. The approach is described in the article Works Council and AI Adoption.

Build your own or use a platform?

For most first assistants, a builder within the existing enterprise platform is sufficient, such as Copilot Studio in a Microsoft environment or custom GPTs within an enterprise license for ChatGPT. A custom build on a model API is worthwhile when at least one of three criteria applies.

The three criteria for a custom build are: the assistant must write data back into business systems; volumes are high enough for usage-based pricing to become cheaper than per-user licenses; or the use case requires quality control that the builder does not offer. The trade-off is covered in detail in the article Make or Buy.

How long does it take to build, and what does it cost?

With a narrowly defined task, a prototype is ready in four to eight weeks, and regular operation follows after a further two to four months of piloting. It takes longer if the knowledge base first needs cleaning up or the works council first needs to be involved.

The costs follow from a simple calculation. The largest item is the person-days for task definition and build as well as for testing and introduction. A first assistant on an existing platform typically requires 15 to 40 person-days of external support. At consultant day rates of around 1,000 to 2,500 euros, as reported in the fee survey of the German consultancy association BDU, this amounts to 15,000 to 100,000 euros. Ongoing license or usage costs come on top, as does the specialists' time for testing and acceptance. A custom build with integration into business systems costs significantly more.

The benefit must be set against this. If the assistant saves ten minutes on each of 2,000 cases a month, that equals around 330 working hours per month. This calculation belongs before the decision to build and should be revisited after three months of operation, this time with measured values.

How Niehues Consulting delivers such initiatives

Niehues Consulting supports business functions from the first task definition through to live operation. The work follows the Niehues Delivery Model with its phases Understand, Build and Introduce. Three test questions matter most for business function assistants. Which business result should change measurably? Does the prototype prove its value in the real workflow with real data? Are users enabled and workflows adapted? First, the workflow, requirements and business case of the business function are analyzed and the tasks are put in order. Next, prototypes are built that prove their value in the real workflow. Finally comes the transfer into operation with internal or external teams, with defined responsibilities and enablement of the specialists.

Marin Niehues steers such initiatives as delivery manager or product owner. An assistant is considered finished only when the business function has adopted it and operation is in place. The entry point is a free 30-minute initial consultation in which the first task is defined: book an appointment.

Frequently asked questions

How long does it take to build an AI assistant for a business function?

With a narrowly defined task, a prototype is ready in four to eight weeks, followed by regular operation after a further two to four months of piloting. It takes longer if the knowledge base first needs cleaning up or the works council needs to be involved.

How much does a custom AI assistant cost?

At 15 to 40 consultant days and market day rates of around 1,000 to 2,500 euros, a first assistant on an existing platform costs 15,000 to 100,000 euros to build and introduce. Ongoing license or usage costs come on top. A custom build with integration into business systems costs more.

Which tasks suit a business function assistant?

Suitable tasks are recurring, have digitally available input and produce a result that a specialist can check quickly. Examples include first replies to inquiries, preparing documents and summarizing cases. Tasks without a defined workflow or a named decision-maker are unsuitable as a starting point.

Who is responsible for an AI assistant in a business function?

Business responsibility lies with the business function, with a person who prioritizes requirements and approves changes. An internal or external team handles technical operation. Both roles are named before the pilot begins.

Should we build our own assistant or use Copilot Studio or custom GPTs?

For the first assistant, the builder of the existing platform is usually sufficient. A custom build is worthwhile when the assistant must write back into business systems, when high volumes make usage-based pricing cheaper, or when dedicated quality control is required.

How does an AI assistant access company knowledge?

Access works through retrieval augmented generation: for each request, relevant passages are retrieved from an approved knowledge base and passed to the model. The assistant answers from these documents and cites the source. Each document needs an owner responsible for keeping it current.

Is an internal AI assistant GDPR compliant?

An internal assistant can be operated in compliance with the GDPR when data processing, storage location and the use of inputs for model training are contractually agreed. The company also defines which data classes belong in the assistant.

Does the works council have to agree?

In Germany, the works council must be involved as soon as an assistant is capable of monitoring employee behavior or performance, which is regularly the case when inputs are logged. Informing it early, during task definition, prevents delays in the pilot.

How do you measure the quality of an AI assistant?

Quality is measured against a fixed set of real test cases, for instance the share of answers specialists adopt without changes. In operation, usage rate and time per case are added and reported monthly.

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