First use cases, then data architecture, then AI: how mid-sized companies are using AI as a growth driver

by Andrea B. Roch June 05, 2026

First Use Cases, Then Data Architecture, Then AI: How Mid-Sized Companies Are Using AI as a Driver of Growth

23:56

At a glance

Most AI projects do not fail because of the technology, but because of a missing use case, fragmented data and a lack of responsibility.
AI multiplies the architecture it finds: clean data multiplies value creation, while data chaos multiplies the chaos — only faster and more expensively.
Management teams must distinguish between three types of AI: productivity AI (ChatGPT, Claude, Copilot), process-integrated AI (e.g. Breeze AI in HubSpot) and AI agents (ready-to-use or custom-built). Only the latter two are strategic growth drivers.
The requirements increase with each level: productivity AI requires nothing. Process-integrated AI requires a central, consistent data foundation. AI agents additionally require clear processes, access rights and escalation paths.
The non-negotiable foundation for AI is the same as for any modern stack: an ERP plus a customer platform, properly connected.
The sequence is decisive: first the specific business case (a business case with a clear problem, clear opportunity or clear potential), then the data architecture, then AI. Never the other way around.

b4you-erst-use-case-2-1200x700

Why do so many AI projects remain stuck in the pilot stage?

AI projects that start with the technology question instead of the processes in which value is currently being lost usually remain stuck in the pilot stage.

The right approach is the opposite: first identify the pain points in the core processes, derive use cases from them, prioritize them, and only then select the technology.

Step 1: Identify pain points for each core process.

At a meta level, the following four questions apply to each core process:

Where is time being lost?
Where are there media discontinuities between systems?
Where is a lot of manual research or data gathering required?
Where is there tedious work that nobody likes doing?

Step 2: Derive use cases from pain points.

Each pain point, or group of related pain points that point to the same topic, becomes a use case. Each use case is formulated as an area of action, e.g. “speed up the quotation process”, rather than as a technical solution such as “introduce an AI tool for quotations”. Two questions apply to each use case: Which pain point does it address? What business benefit will result if it is solved? Deliberately not part of this step is the question of how the use case will specifically be implemented with AI. The goal is to identify the right topics, not the right tools.

Step 3: Prioritize, then evaluate the technology.

First, the use cases are sorted according to two impact dimensions: business benefit (e.g. strategic value, revenue, market and customer impact, differentiation, reputation) and efficiency gain (e.g. time savings, freed-up capacity, error reduction, cost reduction). Together, the two axes answer the question of whether a use case is worth piloting at all. In a second, separate step, the technology question is addressed: Which platform, which type of AI, which tool could fit the prioritized use case? Only on this basis does a pilot begin on a foundation capable of supporting it.

What does “AI multiplies the architecture it finds” mean?

It means that AI is not an independent system, but rather a layer that operates on the data and processes that already exist. AI does not improve the underlying architecture; it amplifies it. In both directions.

Definition: Process-integrated AI.

AI functionality that is embedded directly into the data model and workflows of an operational system, such as lead scoring, cross-selling predictions or service automation within the customer platform. Unlike productivity AI (stand-alone assistants without access to business processes), process-integrated AI is only as powerful as the data foundation on which it operates.

If a clean, centralized data foundation exists, AI multiplies value creation: it automates workflows, provides reliable predictions and personalizes customer interactions across the entire customer journey. If there is data chaos, no intelligent added value is created; instead, existing problems are amplified more quickly. AI is only as good as the data, processes and sources of knowledge it can access. If these are incomplete, outdated, inconsistent or poorly structured, AI produces incorrect recommendations, unreliable answers or even automates faulty decisions.

For management, this means that AI is not a substitute for clean data management. It makes data quality visible and increases the pressure to address deficiencies. Without a solid data foundation, risks arise in efficiency, decision-making quality, compliance and employee trust.

→ In short: poor data leads to poor AI. Good AI therefore starts not with the tool, but with data quality, governance and clear processes.

Which three types of AI do management teams need to distinguish?

Management teams must clearly distinguish between three categories: productivity AI, process-integrated AI and AI agents. They differ fundamentally in what they do, what they cost and, most importantly, what requirements they place on your data architecture.

Productivity AI

consists of stand-alone assistants such as ChatGPT, Claude or Microsoft Copilot. They support individual employees with research, drafting texts, summarizing content or coding. This AI works immediately, regardless of the state of your tool stack. It does not need access to your business data because it does not know your business processes. Productivity AI is useful — often very useful — and it should be introduced and properly governed. But it is an efficiency tool for individual tasks, not a strategic growth driver. It makes individual work steps faster; it does not change how your company acquires leads, retains customers or realizes cross-selling.

Process-integrated AI

is AI that sits directly within your operational systems and operates on their data model. In a modern customer platform, it performs lead scoring, identifies cross-selling opportunities, prioritizes service tickets and optimizes campaigns. It is an intelligence layer: it provides assessments, predictions, content or recommendations. The actual decision and execution remain with people or existing automations. This AI understands your customers, your pipeline and your business processes, but only if this information is available centrally, consistently and accessibly.

For management teams, this has a concrete consequence whenever a new operational core system such as a CRM or customer platform is purchased: the question is no longer simply whether the system can map the processes, but whether process-integrated AI is already built in and whether all customer data exists within a shared data model (single source of truth). Choosing a system in which marketing, sales and service data live in one place lays the foundation on which AI can later create value. Connecting three or four separate tools through interfaces creates the data chaos that AI will not be able to heal later, but will instead multiply.

AI agents

take a decisive step further: they independently execute multi-step tasks. Instead of making a suggestion, they qualify incoming leads, answer first-line service inquiries, research account profiles before a sales call or create and publish content across multiple steps and systems, with defined escalation to humans when necessary. Agents are the action layer above the intelligence layer.

Here, two variants must be distinguished, as they differ significantly in effort and scope:

Ready-to-use agents come pre-built with the platform, such as the Breeze agents in HubSpot for prospecting, customer service, content or social media. They can be configured within days to weeks, but operate within the logic of the respective provider and on its data model. For standard use cases, they provide a fast lever with manageable effort.

Custom-built agents are developed for a specific company and tailored to its processes, data, systems and requirements. Typical characteristics:

know company-specific processes, procedures and rules
use internal data sources such as CRM, ERP, SharePoint, ticketing systems
or knowledge databases
are aligned with roles, permissions and governance requirements
work according to defined processes and tasks
execute actions or prepare them, rather than merely answering questions

Examples:

Service agent for technical troubleshooting with access to intervention reports
and product data
Sales agent that analyzes sales opportunities, prepares quotations and recommends
next best actions
HR agent for internal policies, onboarding and employee questions
Management agent for market monitoring, KPI analyses or summaries from various systems

Effort: weeks to months rather than days. The investment pays off where the automated process is your USP

The requirements placed on your architecture increase with each level: productivity AI requires nothing. Process-integrated AI requires a clean, centralized data foundation. AI agents additionally require clearly defined processes, properly regulated access rights, defined escalation paths and governance that determines where the agent is allowed to act independently and where human involvement is mandatory. Deploying agentic AI on an untidy stack multiplies the risks of process-integrated AI — now with the authority to take action.

The most common strategic mistake at management level is this: a company introduces productivity AI, sees a genuine benefit and concludes that it is already “doing AI”. It confuses the fast, simple tool with the slower, more demanding lever. Anyone who wants AI as a growth driver means process-integrated AI or agents. And both start not with an AI tool, but with the data architecture.

What makes a data architecture AI-ready?

An architecture becomes AI-ready when customer-relevant data is centralized, consistent and accessible across the entire customer journey, and when the CRM is the source of truth rather than the backup for the Excel graveyard.

The foundation is the same one that every modern company needs anyway: an ERP as the financial and operational source of truth and a customer platform as the customer-facing source of truth, properly connected through open interfaces. Third-party systems are added as needed. If this core does not exist, you do not have an AI question. You have an architecture question that must be solved first.

Three questions are decisive before investing in process-integrated AI:

First: Is your customer data stored centrally? If sales data is stored in Excel, marketing data in a separate tool, service data in a third system and master data in the ERP, no AI can create a comprehensive view.

Second: Can you access every customer interaction across the entire customer journey? AI that is supposed to identify cross-selling opportunities needs the history: what was purchased, asked about, complained about and ordered. If this history is missing, every prediction remains guesswork.

Third: Is your CRM the source of truth, or the backup for the Excel spreadsheets in which the team actually works? AI operates on the CRM. If reality lives alongside it in spreadsheets, AI sees a distorted picture.

This is where the underestimated strength of a modern customer platform such as HubSpot lies: Smart CRM, Marketing Hub, Sales Hub and Service Hub share a common data model, while Breeze AI operates directly on it as an AI layer. What would be a massive data-engineering project in a fragmented stack is standard functionality here. Not because HubSpot has better AI, but because the architecture underneath is right.

Why AI projects fail: three anti-patterns

AI projects typically fail because of three recurring patterns: the AI project without a use case, AI on a data silo and AI that belongs to no one.

Anti-Pattern 1:
AI as an end in itself

The project starts with “We have to do something with AI” rather than with a specific, measurable business case — a business case with a clearly identified problem, opportunity or potential. There is no target state and no metric against which success can be measured. As a result, it is impossible to say at the end whether the project worked. AI without a business case is a solution still looking for its question. Such projects fizzle out once the initial enthusiasm fades.

Anti-Pattern 2:
AI on a data silo.

The company places an AI function on a fragmented data landscape and expects reliable results. The AI only has access to part of the truth because the other half of the data is located in a system it cannot see. The suggestions are plausibly worded but factually incorrect. After the first bad recommendations, sales and service lose confidence and stop using the AI. The tool continues running, the license continues to be paid, but the benefit is zero.

Anti-Pattern 3:
The AI that belongs to no one.

No one has end-to-end responsibility for which AI functions are being used, which data they operate on and who monitors their quality. AI initiatives emerge across individual departments without a shared data foundation or shared governance. The result is the same as with tool anarchy: diversity without architecture. Without clear responsibility, AI does not become a lever, but the next source of uncoordinated growth and complexity.

→ If your AI initiatives currently exhibit one or more of these patterns, they are worth honestly assessing. AI success is not a technology issue. It is an architecture and leadership issue.

What we do differently at business4you

We do not start an AI initiative with a tool demo. We start with an honest look at your architecture, because that is where it is decided whether AI will have an impact in your organization or burn money.

At business4you, we have built our own in-house team of data and IT architecture specialists. We know HubSpot and Breeze AI in depth, and we build reliable interfaces between HubSpot and any ERP or third-party system with an open REST API. Our core expertise lies precisely where AI success or failure is decided: in the clean, centralized data foundation without which process-integrated AI cannot function.

Concretely, this means that we first clarify the use case that AI is supposed to serve with you. We check whether your data architecture can support this use case at all. We tell you openly if the first sensible step is not an AI tool, but organizing your stack — even if that means a smaller engagement in the short term. And we only build on it once the foundation is right.

Strategy, architecture consulting, implementation and development — all in-house from start to finish. No handovers, no gaps in responsibility between consulting and implementation. One client, one team, one result.

Reduce complexity, increase impact. With AI, this applies twice over. Because AI multiplies what it finds, and it is our responsibility to ensure that it multiplies value creation rather than chaos.

Would you like to know whether your architecture is ready for AI and where the first meaningful lever lies?

Book a no-obligation initial consultation. We will listen, look at your data and tool landscape with you, and give you an honest, concrete assessment. No generic AI promises.

Frequently asked questions

How do we introduce AI correctly in mid-sized companies?

Three steps in the right order determine success or a pilot graveyard.

Step 1: Define the business case. What specific, measurable business case (problem, opportunity or potential) should be addressed with AI? Without a clear target state, there is no benchmark for any subsequent decision.

Step 2: Review and organize the data architecture. Only now, and with the specific use case in mind, is it determined whether the data is centralized, consistent and accessible.

Step 3: Establish responsibility. One person or one team takes end-to-end responsibility for which AI operates on which data and how its quality is ensured.

The most common cause of failed AI projects is the reversed order: first the tool, then the use case, and finally the realization that the data is missing.

Which AI use cases are most worthwhile for mid-sized companies?

There is no universal list of the “best” use cases — the right ones emerge from the pain points in your own core processes, not from best-practice templates used by other companies.

Frequently highly profitable areas for getting started in SMEs include: lead qualification and lead scoring in sales, identifying cross-selling opportunities among existing customers, first responses and prioritization in customer service, automated research and quotation briefings before sales calls, and KPI synthesis for management from various systems.

Which of these areas provides the greatest business benefit and efficiency gain for your company only becomes clear after analyzing the pain points in your core processes.

Can we start with AI without consolidating our tool stack first?

That depends on what type of AI you mean. Productivity AI such as ChatGPT or Claude can be introduced immediately, regardless of the state of your stack. It supports individual employees without needing access to your business data.

Process-integrated AI and AI agents, on the other hand, require a centralized, consistent data foundation, while agents additionally require clearly defined processes and access rights. Without this foundation, results remain superficial, incorrect or, in the case of agents, even risky, because an autonomous agent can actively implement incorrect decisions based on poor data.

Anyone who wants AI as a growth driver should invest in the data architecture first.

Why do AI projects in mid-sized companies fail most often?

The most common causes are not technical. AI projects fail because they start without a clearly defined use case, because they encounter fragmented data and because no one takes end-to-end responsibility for the AI.

The technology is rarely the problem. The architecture and leadership are.

What is the success rate of AI projects?

Studies by organizations such as Gartner, McKinsey and MIT have consistently shown for years that 70 to 85 percent of AI projects do not reach productive deployment or fail to deliver measurable business value.

The causes are rarely technological. They almost always lie in the architecture and leadership: no clearly defined use case, a fragmented data foundation, and no end-to-end responsibility for operation and scaling.

Anyone who clarifies these three points before introducing the first AI tool can turn the statistics in their favor.

What is the difference between productivity AI, process-integrated AI and AI agents?

Productivity AI consists of stand-alone assistants such as ChatGPT, Claude or Copilot that accelerate individual tasks such as research or drafting texts. They work without access to your business processes.

Process-integrated AI is embedded in your operational systems, operates on their data model and provides an intelligence layer: lead scoring, predictions and recommendations.

AI agents go one step further and independently execute multi-step tasks (action layer).

With agents, a distinction is made between ready-to-use agents, which come pre-built with the platform (e.g. Breeze agents in HubSpot), and custom-built agents developed for your own workflow.

The requirements placed on your data architecture increase with each level, and so does the strategic leverage.

Do we need our own AI, or is ChatGPT Enterprise enough?

That depends on what you want to achieve strategically.

ChatGPT Enterprise — as well as Claude for Work or Microsoft Copilot for Business — is productivity AI: an assistant for individual employees that accelerates research, drafting texts or summarizing content. It can be introduced quickly, is immediately useful and provides a clear efficiency gain in everyday work.

For strategic growth, however, it is not enough. Anyone who wants to use AI to systematically qualify leads, identify cross-selling opportunities or automatically answer service inquiries needs process-integrated AI or AI agents — in other words, AI that accesses your business data and is embedded in your processes.

Fully developing your own AI is almost never necessary for mid-sized companies. A sensible approach is a combination: ChatGPT Enterprise for productivity, plus an integrated AI solution on your customer platform for growth.

How can we tell whether our architecture is ready for AI?

Three questions provide the answer.

Is your customer data stored centrally?

Can you access every customer interaction across the entire customer journey?

Is your CRM the actual source of truth, or does your team continue to work alongside it in Excel?

If you can answer yes to all three, your architecture is AI-ready. If not, organizing this foundation is the first step.

How much does an AI project cost, and when does it pay off?

The direct costs of AI functions are often the smaller part. With process-integrated AI, the larger cost is usually the groundwork: organizing the data architecture and building the necessary interfaces.

However, this investment pays off independently of AI as well: through better reporting, less manual data maintenance and greater sales efficiency.

In our experience, the most expensive option is the failed AI project: license costs without any benefit, plus the loss of trust teams develop in AI as a whole.

Should we start with a ready-to-use agent or have a custom agent built?

The rule of thumb: standard first, custom-built only where the process is your USP.

Ready-to-use agents (such as the Breeze agents in HubSpot) can be configured within days to weeks and cover standard use cases such as prospecting, first-line service responses or content creation. They are the right choice when the process is a commodity and speed matters.

Custom-built agents are worthwhile where the process makes your company unique and therefore cannot be adequately represented by a standard agent.

The effort is significantly greater. The investment pays off precisely in the USP processes where your competitive advantage lies.

Who should be responsible for AI in mid-sized companies?

The key is that one person or one team takes end-to-end responsibility for which AI operates on which data and how its quality is ensured.

In large corporations, this is a solution or data architect. In mid-sized companies, it is usually the CIO role, a technically skilled member of management or an external architecture partner.

Without this clear responsibility, scattered AI initiatives emerge without a shared data foundation.

Uncontrolled growth instead of leverage.

HubSpot