Artificial intelligence in business is the topic of the moment. Every conference talks about it, every vendor is pitching it, every board is asking for it. Yet there's a huge gap between the enthusiasm and the actual results.
According to a 2024 McKinsey report, fewer than 30% of companies that launch AI projects get them into production with measurable results. The remaining 70% stall at the prototype, the demo or the pilot project that never scales.
The problem isn't the technology. Models are powerful, accessible and improve every month. The problem is how AI gets introduced into organisations: without a clear strategy, without integration into real processes, without governance.
This article is for anyone who has to decide whether and how to invest in AI. Not for people who write code: for business leaders who want to understand what actually works, what to avoid and where to start.
Chatting with a model isn't adopting AI
The first distinction to make is fundamental, and it's often overlooked.
Using ChatGPT to write emails, summarise documents or draft text isn't "adopting AI in your business". It's using a generic tool for generic tasks. Useful, but not transformative.
Adopting AI in a business means building intelligent agents specialised in your own domain, integrated into your existing systems, capable of carrying out complex, measurable tasks.
The difference is substantial:
- A generic chatbot answers generic questions
- A specialised agent knows your processes, accesses your data and integrates with your systems
A specialised agent can run an approval workflow, analyse sales data against historical trends, coordinate a logistics process or triage support requests by real urgency rather than by keyword.
That's the difference between using a model and building an intelligent system. The first is available to everyone. The second is what generates competitive value.
What are companies actually looking for when they say "we want AI"?
When a CEO says "we need to adopt AI", they're rarely looking for artificial intelligence as such. They're looking for something else:
- lower operating costs on repetitive processes
- better, faster decisions
- freeing up people's time for high-value work
- keeping pace with the competition
AI is a means, not an end. And like any means, it only works when applied to the right problem.
The most common mistake is to get it backwards: starting with the technology ("we have AI, where do we put it?") instead of from the problem ("where are we losing the most time and value?").
When you start from the problem, the solution might be an AI agent, but it might equally be traditional automation, a redesigned process or simply an organisational decision. AI isn't always the answer, but when it is, it's a powerful one.
Three mistakes that sink AI projects in business
Most AI projects that fail don't fail for technical reasons. They fail for organisational and strategic ones. Here are the three most common patterns.
1. Isolated tools with no integration
The company buys or builds an AI tool that works in isolation. It isn't connected to existing systems, doesn't access real data and doesn't fit into the daily workflow.
The result: people have to step outside their own workflow to use it, which means that once the initial enthusiasm fades, they stop using it. The tool becomes shelfware: software bought and never used.
2. Prototypes that never reach production
A technical team builds a brilliant prototype. The demo impresses management, but taking it to production requires integration, security, monitoring, error handling, scalability. Nobody planned for any of that, and the prototype stays a prototype.
The gap between a demo and a production system is enormous. A prototype proves something is possible. A production system proves it's sustainable.
3. Dependence on a single person
The AI project depends entirely on one person, often whoever built the original demo. When that person changes role, leaves the company or simply gets pulled onto other priorities, the project grinds to a halt.
This happens when AI is an individual's initiative rather than the organisation's. Without shared ownership, documentation and governance, any project is fragile.
When does AI actually work in business?
Artificial intelligence works in business when it's built for that business's own domain. Not generic: specific.
Here's what that means in practice:
- An agent that knows the product catalogue and can answer customers with accurate, up-to-date, contextual information, not generic replies
- An agent that analyses support tickets and triages tickets by urgency, based on real context rather than static keywords
- An agent that monitors operational KPIs and flags anomalies before they become problems, with explanations, not just alerts
- An agent that manages document workflows, extracting information, checking consistency and proposing actions, cutting hours of manual work
In every one of these cases, the agent isn't a chatbot. It's a system that knows the domain, accesses the right data and produces measurable results.
Specialisation is the critical factor. A generic model can do a bit of everything. A specialised agent does one specific thing very well — and that specific thing is exactly the process where the business is losing the most time or value.
Where to start: the point of greatest value
The most important question isn't "which AI should we use?" but "where do we start?"
The answer is always the same: start at the point with the most friction and the most value.
Identify the process that currently costs the most in time, errors or missed opportunities. The one where people do high-volume repetitive work. The one where a mistake has serious consequences. The one where speed of response makes all the difference.
Then build an agent dedicated to that process. Measure the results, and expand only if it generates value there.
This approach has three advantages:
- Contained risk: you're investing in a single process, not a global transformation
- Fast results: the value is visible in weeks, not months
- Concrete learning: the organisation learns to oversee AI on a real use case before scaling
If the agent doesn't generate value on the highest-impact process, it won't generate value anywhere. Better to find that out straight away, with a modest investment, than after months of development.
Before you build: evaluate the agentic services already on the market
A frequent mistake is to jump straight into building a custom agent without first checking whether a service already solves the same problem.
The market for agentic services is maturing fast. There are vertical platforms offering ready-made agents for specific areas: customer support, document analysis, ticket management, lead qualification, data monitoring. A few concrete examples: tools such as Intercom Fin or Zendesk AI for customer support, or platforms such as Relevance AI and LangChain Cloud for orchestrating agents across business workflows.
Evaluating these options before building makes sense for three reasons:
- Time: a ready-made service can be live in days, not weeks
- Upfront cost: no development investment; you pay as you go
- Validation: it lets you find out whether AI genuinely generates value on that process before investing in a proprietary solution
Ready-made services, however, have clear limits: they don't know the business's specific domain, they don't integrate deeply with internal systems, they can only be customised up to a point, and company data often passes through third-party infrastructure.
The practical rule: if the problem is generic and the process is standard, evaluate an existing service first. If the competitive value lies in specialisation, depth of integration or control over data, build a dedicated agent.
In many cases the most effective path is hybrid: start with an existing service to validate the idea, then build a proprietary solution once the value is confirmed and you need customisation and control.
AI agents versus traditional automation: when do you need which?
Not everything needs to be AI. This is a distinction many vendors would rather not make, but it's essential for investing wisely.
Traditional automation follows fixed rules: "if the customer orders more than 100 units, apply a 10% discount". It works perfectly when the process is predictable, the rules are clear and exceptions are rare.
An AI agent reasons about context: "given this customer, their history, the margin on this product and the volume requested, what discount makes sense?" It's needed when the process requires interpretation, when there are many variables, when the right answer changes with context.
The practical difference:
- Sorting emails into predefined folders → automation
- Reading an email, gauging its urgency and deciding who should respond → AI agent
- Generating a weekly report with fixed data → automation
- Analysing data, spotting anomalies and explaining what's happening → AI agent
The mistake is using AI where automation would do (a waste of resources), or using automation where reasoning is needed (poor results).
The right choice depends on the process. A good technical partner helps make this distinction before a single line of code gets written.
What about your data? Sovereignty and open-source models
This is the question business owners raise most often, and rightly so.
When you use a cloud model from an external provider (OpenAI, Anthropic, Google), company data passes through that provider's servers. For many businesses that's acceptable. For others it isn't.
Regulated sectors such as healthcare, finance and law face strict rules on where data resides and who can access it, but even SMEs in unregulated sectors are starting to ask the same question: sending company data to an external provider means giving up some control.
The good news is that real alternatives exist today.
Open-source models such as Llama, Mistral and Qwen have reached a level of quality comparable to proprietary models for many business use cases, and they can run on in-house hardware: on-premise or on a private cloud.
It's worth knowing the options available, because each model has different strengths:
- Llama (Meta): the most widely used family of open-source models. It excels at reasoning and multilingual text generation. The latest versions compete directly with top-tier proprietary models.
- Mistral (Mistral AI): European models with a particularly good quality-to-size ratio. Ideal when you want strong performance from limited hardware. Strong support for European languages.
- Qwen (Alibaba): excellent at multimodal tasks (text, images, code) and at mathematical reasoning. Offers very compact variants that run even on modest hardware.
- Gemma (Google): lightweight, fast models optimised for specific tasks such as classification and information extraction. Ideal for agents that need to respond in real time.
On the proprietary side, the main options are GPT-4o (OpenAI), Claude (Anthropic) and Gemini (Google): more powerful on complex tasks, but the data passes through their cloud infrastructure.
Running models on your own infrastructure is called data sovereignty: the data stays within the company's own infrastructure, no external provider sees it, and the company keeps full control.
Building AI agents with open-source models is already possible and competitive. It isn't a second-tier option: it's an architectural choice that balances performance, cost and control.
The decision on which approach to adopt — cloud, on-premise or hybrid — should be based on the business's specific requirements, not on a provider's marketing.
How do you know if your business is ready for AI?
Before investing, it's worth making an honest assessment. Not every context is ready for AI, and starting before you're ready is the fastest way to burn through budget and internal credibility.
On the AI projects we run through our AI & Agent Engineering service, the first question we ask isn't "which model should we use?" — it's "does a working manual process already exist?" If the answer is no, if the process is chaotic or made up of constant exceptions, AI won't solve anything: it will just produce chaos faster. We only work where there's a process people understand, with people who know it and where automation replaces repetitive, well-defined work. Ninety per cent of the value of AI in business lies in this: finding the right processes, not the most powerful models.
Here's a practical checklist:
- Available, accessible data: the agent needs data to work. If the data sits in silos, in incompatible formats or simply doesn't exist, the first step is fixing that
- Documented processes: if nobody knows exactly how the process you want to improve actually works, an AI agent can't help. Document it first, then automate
- A team that can oversee the system: someone in the organisation needs to understand what the agent does, how to evaluate its results and when to step in. You don't need a dedicated AI team, but you do need the expertise
- Realistic expectations: AI isn't magic. It doesn't fix organisational problems, doesn't compensate for broken processes and doesn't replace strategic decisions. It works when the problem is clear and the data is there
- A clear position on data sovereignty: the business needs to know where its data can go and where it can't. This decision shapes the architecture, the costs and the timeline
If one of these elements is missing, the advice is simple: fix that first, then invest in AI. It isn't wasted time: it's building the foundations AI will need to work on.
The most common mistakes decision-makers make with AI
Beyond the three structural mistakes already covered, there are decision-making patterns worth recognising.
Confusing speed with urgency. AI moves fast as a technology, but that doesn't mean the business has to rush. Starting well matters more than starting first. A well-planned AI project generates value for years. A rushed one generates costs and frustration.
Delegating everything to the vendor. A technical partner can build the agent, but can't decide which problem to solve, which data matters or how the process really works. Only the people who know the business have those answers. The collaboration between domain expertise and technical expertise is what makes AI projects work.
Chasing the perfect solution. The first agent won't be perfect. It doesn't need to be. It needs to be good enough to generate measurable value and flexible enough to improve over time. The iterative approach — build, measure, improve — is the only one that works in practice.
Ignoring change management. Introducing an AI agent into a process means changing how people work. If that change isn't managed, communicated and supported with training, adoption will be low and the project will fail, regardless of the technical quality. Since 2025 there's also a statutory obligation to provide AI literacy training, which makes preparing the team not just strategic but necessary.
What comes next
AI in business isn't "making ChatGPT available to employees". It's building intelligent agents specialised in your own domain, integrated into your existing architecture, capable of doing complex, measurable things.
The path is clear:
- Identify the point of greatest value: the process with the most friction, cost or opportunity
- Build a specialised agent: not generic, but designed around the business's specific domain
- Measure the results: with concrete metrics, not impressions
- Then expand: only once the value has been validated on the first case
This is the approach that works. Not the fastest one, but the most sustainable.
If you need help working out where to start, evaluating which process has the most potential or building the first specialised agent, this is exactly what QMates does.
We don't sell AI. We build systems that work.