# AI architecture: why a platform beats point solutions

Author: Marin Niehues, Niehues Consulting · Published: 2026-01-05 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/ai-platform-architecture/

A shared AI platform replaces point solutions once several applications need the same data access and models, in practice from three applications onwards. It lowers effort and risk per application. The decision concerns IT as much as the business units, because it determines how quickly an idea becomes a running application and who is accountable for its operation. This article sets out the layers of an AI architecture, the criteria for the right moment, a worked calculation of build cost and the responsibilities that make a platform work in daily use.

## What does an AI platform provide, and where do point solutions end?

An AI platform is a shared technical foundation used by several applications, and point solutions end where the second and third application build the same services again. The platform bundles the services that every AI application requires: access to models, connection to company data, monitoring in operation and management of access rights. The business application itself, such as a customer service assistant or a document review tool, is built on top of it.

A point solution brings these services along anew in every application. It delivers the first result quickly, and for a single initiative that is the right choice. With the second and third application, parallel integrations and licenses appear, and with them parallel sources of error. Four effects then occur regularly:

- Duplicated integration work: each application builds its own connection to the same source systems.
- Inconsistent access rights: the same data is protected differently from one application to the next.
- Missing transparency: usage and costs are spread across project budgets and cannot be consolidated.
- No shared operation: faults are fixed application by application, and the knowledge stays within the respective project team.

These effects originate in the organization. Every point solution has an owner; the sum of all solutions has none. This becomes most visible at the interfaces to grown systems.

**A platform shifts effort from the individual application to a shared service. Reuse begins with the second application, and between the third and the fourth the build has paid for itself.**

## The five layers of an AI architecture

A division into five layers is sufficient for architecture decisions. It shows what is shared and what remains the responsibility of the business application.

**Five layers of an AI platform**
- Data layer: access to source systems, preparation and indexing for search in a vector database, data quality, permissions.
- Model layer: access to language models and other models through an AI gateway, a uniform interface with cost tracking and usage limits, replaceable per use case.
- Orchestration: control of workflows and agents with their tool calls, and handovers between application and people.
- Operations and monitoring: LLMOps with logging and quality measurement, plus cost tracking and fault handling.
- Application layer: the business application with its interface and process connection.

The build follows a fixed sequence. Data access and permissions are settled first, because every application depends on them. The shared model access comes next, followed by operations with measurement and cost tracking. The model layer deserves particular attention: providers and models change faster than business processes. An AI gateway that allows the change to be made in one place protects the investment in every application above it and at the same time creates a central point for cost control and logging. Which models to build, buy or adapt is covered in the article on [make or buy](/blog/build-buy-or-adapt-ai/). Ongoing operation is described in the article on [LLMOps](/blog/llmops-running-ai-applications/).

Whether the platform runs in the cloud, in the company's own data center or in a hybrid setup is a follow-on decision, determined by the protection needs of the data and the available operating capabilities.

## When does an AI platform pay off?

An AI platform pays off from three planned applications that will use the same data sources or models over the next twelve months. With one or two applications, a point solution with a clean interface is the faster route. Four criteria sharpen the decision:

- Number of applications: from three applications with shared data sources, the shared foundation pays for itself.
- Uniformity of requirements: if the same rules on access and traceability apply to all applications, those rules belong in the platform.
- Operational ownership: a team must be able to take on operation permanently. If that team is missing, the role is created first and the platform second.
- Pace of the business units: if business units need new applications in weeks rather than months, the platform shortens the time to launch for every further application.

**Point solution and platform compared**
Criterion | Point solution | Platform | With Niehues Consulting
Time to the first application | Short | Longer to build, short afterwards | Platform grows along the first use cases
Effort per further application | Stays equally high | Falls with every reuse | Building blocks with a named owner from the start
Control of cost and risk | Separate per project | Centrally visible | Metrics anchored in the AI portfolio
Operation | In the project team | In a platform team | Handover into operation with clear roles
Change of model or provider | Rebuild per application | Change in one place | Replaceability as an architectural requirement

The result of this assessment is rarely a pure either-or decision. A proven path starts with the first application that promises the greatest value and builds the shared services along its requirements. The second application already uses them, and the third confirms the design. The platform emerges from delivery.

## What does it cost to build an AI platform?

Building an AI platform costs mainly people: a small platform team over a year is the largest item, with licenses and compute added on top and growing with usage. The figure only becomes meaningful when set against the effort the platform saves each application. The following worked example shows the logic with assumptions that each company replaces with its own values.

**Worked example: build effort and savings**
Build assumption: a platform team of three people works on the shared foundation for half a year. That is around 3 times 110 working days, or 330 person-days.

Savings assumption: without a platform, each application needs around 120 person-days for data connection and model access, permissions and monitoring. With a platform, this share falls to around 20 person-days. The saving is 100 person-days per application.

Result: after three applications, 300 person-days have been saved, after four, 400. The build has therefore paid for itself between the third and the fourth application, and every further application lowers costs permanently.

Test question for the executive team: how many person-days did your existing AI applications need for connection and operation, and how many of them were duplicated work?

To value this in money, the company multiplies the person-days by internal full costs or by the day rate of external staff. According to the fee study of the German consultancy association BDU, external consulting in the German market is typically charged at around 1,000 to 2,500 euros per consultant day. Running costs for model usage and infrastructure belong separately in the operating budget, because they grow with usage and are allocated to each application through the AI gateway.

## Running the platform as a product

A platform delivers value when its users adopt it. This calls for product-style leadership: a product owner and a roadmap, plus a service catalog for the internal customers. Without this accountability, the result is an infrastructure that exists technically and is bypassed by the business units. Four commitments create reliability:

- A named product owner decides on the sequence and scope of platform services.
- A platform team is responsible for operation and has defined response times.
- Business units use the services through a service catalog and raise demand through a fixed channel.
- Metrics on usage and cost flow regularly into management reporting.

The executive team sets the frame: which applications take priority, who decides on exceptions and which budget the platform team is responsible for. How the platform fits into the full path of an AI implementation is set out in the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

## Niehues Consulting as a partner for platform build-up

[Niehues Consulting](/) combines AI management with delivery management. The business case and the selection and sequence of applications come first, and this sequence determines which platform services are needed first. AI prototypes prove the benefit in the real workflow before investment goes into the shared foundation. In the transfer into operation, the consultancy works with internal or external teams and settles processes and responsibilities, along with user enablement. Marin Niehues is accountable for initiatives from the delivery perspective, from VC-funded scale-ups to DAX 40 groups.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Implement. Three of its test questions matter most for the platform decision. From the Build phase: which roles, decision paths and interfaces does operation need? And: is the business case confirmed with real values? The second question replaces the assumptions in the worked example with measured person-days. From the Implement phase: who is accountable for operations from day one? Without this answer the platform remains a project without an owner. The entry point is the [free initial consultation](/#kontakt) of 30 minutes.

**What decision-makers should retain**
A platform pays off from three applications that use the same data and models. It emerges along the first applications, is run like a product and needs an owner with budget and decision rights. Settling this accountability before the third application starts avoids a later rebuild.

**Sources**
- Bundesverband Deutscher Unternehmensberatungen (BDU): Honorare in der Unternehmensberatung, fee study 2022

## Frequently asked questions

### What is the difference between an AI platform and an AI application?

An AI application solves a business task, for example answering customer inquiries. An AI platform provides the shared technical services on which several such applications run, including model access and data connection along with monitoring. The platform is built once and used by every application.

### When does an AI platform pay off?

An AI platform pays off from three planned applications that will use the same data sources or models over the next twelve months. For a single, clearly bounded initiative, a point solution with a clean interface is sufficient. The precondition is a team that takes on operation permanently.

### How much does it cost to build an AI platform?

The largest item is the platform team, in a typical calculation three people over half a year, or around 330 person-days. Licenses and compute come on top and grow with usage. If the platform saves each application around 100 person-days, the build has paid for itself between the third and the fourth application.

### What is an AI gateway?

An AI gateway is a uniform interface between applications and the language models of different providers. It bundles access and cost tracking in one place, as well as usage limits and logging. A change of model thus becomes a change in the gateway, without rebuilding the applications.

### Which layers does an AI architecture have?

An AI architecture can be divided into five layers: data, models, orchestration, operations, application. The first four belong in the shared platform, while the business unit is responsible for the application layer. The build follows the sequence data, model access, operations.

### Who is responsible for an AI platform in a company?

An AI platform is owned by a product owner with decision rights over the sequence and scope of services, supported by a platform team for operation. The executive team determines budget and the priority of applications. Business units raise their demand through a fixed channel.

### Can a platform be built before applications exist?

A build in advance usually misses the actual demand. A more reliable route is the first valuable application, whose requirements shape the shared services. The second application already uses them, and the third confirms the design.
