AI can be added almost anywhere today. That makes it just as important to think about where it genuinely belongs — and where it does not.
At Google Cloud Summit 2026, AI featured in almost every session. What caught my attention most, though, were the examples with tangible results and a clear business impact. They all had one thing in common: they started with a specific goal.
I collected the most useful examples together with practical recommendations on what to get right from the start — and what to avoid when introducing AI into a business.
1. Start with the goal, then choose the AI use case
One message came up repeatedly at the summit: a good AI solution starts with a goal. First, you need to know which process you want to improve and how you will measure the impact.
When AI has a clear job to do
METRO uses AI to process orders coming in through multiple channels, including WhatsApp, chat and email. Around 8 million orders are processed automatically, and AI is also used alongside Salesforce by a team of approximately 8,000 people.
O2 uses AI in customer support, where around one third of customer service is automated. AI also evaluates customer interactions, and when it detects signs of dissatisfaction, a human agent follows up within 15 minutes.
In both cases, AI has a clearly defined role: process requests faster, analyse large volumes of communication or bring a person into the process at exactly the right moment.
What this can look like in practice
In an ecommerce platform, portal or internal system, the goal might be to reduce the time support teams spend searching for order information. AI can read the customer’s request, find the relevant order, check its status, prepare a suggested response and pass it to an agent for review.
This is a use case that can be both designed and measured. The business knows which process it wants to improve, what outcome it expects and how it will determine whether AI is actually helping.
What to avoid
AI can accelerate software development, but that does not mean every idea sitting in the backlog should suddenly become a new feature.
More code also creates more work around it: reviews, testing, maintenance, security checks, deployment decisions and responsibility for how the new functionality behaves in production.
The time saved in development can easily reappear elsewhere in the process. A team may produce more, while also having to review, fix, align and maintain more — without creating any real efficiency gain.
What to do instead
Choose a small number of AI use cases with the strongest potential impact. Focus on scenarios where you can clearly define which process is being improved, who benefits, what data is required and how success will be measured.
A useful question to start with is:
Which process do we want to improve, and how will we measure the impact?
2. Every AI tool needs clear rules
Another warning raised at the summit was the so-called Agentic Zoo: a situation where a company rapidly introduces multiple AI solutions without a shared structure or common rules.
One agent helps with content, another retrieves data, a third processes tickets, while others prepare reports or assist with administration. At first glance, that can look like progress. In practice, every additional AI solution brings its own costs, data access requirements, output controls and security implications.
Cost, data and security
That is why AI solutions should be evaluated not only by what they can do, but also by what they will cost at real-world scale. The price of models, tokens and tools can change over time, and a solution that looks inexpensive today may become significantly more costly as usage grows.
It also makes sense to decide which tasks genuinely need AI and which are better handled by conventional code. Precise calculations, straightforward rules and stable system steps often belong in code. AI creates more value when the task involves text, documents, customer requests, ambiguous input or several signals that need to be interpreted together.
What to avoid
Be cautious with internal rankings that reward people simply for using the most AI — sometimes referred to as a token leaderboard. In other words, measuring success by the number of prompts, outputs or tokens a team generates.
More output does not automatically mean better results. A company can end up paying more and reviewing more while getting no closer to meaningful business value. AI should improve the work itself, not turn usage into a competition.
What to do instead
Every AI solution should have clear safeguards for secure, controlled and sustainable use. In practice, this usually means defining four areas:
- Data: which sources AI can access and which information must remain outside its reach.
- Actions: what AI can do independently and where it should only prepare a recommendation for a person to approve.
- Costs: how model, token and tool usage is monitored as volume increases.
- Control: where outputs are logged, who approves them and what happens when the result is incorrect or uncertain.
AI then becomes part of a system with clearly defined boundaries. The company knows where it can work independently, where it should prepare supporting material and where the final decision should remain with a person.
A useful question for this stage is:
What rules does our AI solution need to remain secure, controllable and sustainable?
3. Look for work that AI can scale
One of the strongest examples from the summit involved AI-generated promotional images for eyewear. With a catalogue of around 20,000 active products, organising traditional photo shoots for different visuals and markets is both time-consuming and expensive. The team therefore used AI to create product imagery tailored to individual countries: for Italy, for example, dark-haired models in a more Mediterranean setting; for Nordic markets, blonde models in a distinctly Scandinavian environment. Visitors then automatically saw visuals tailored to their market.
The results demonstrated the value of scale. The entire process, from the initial idea to publishing the image on the website, took around two hours. The cost per image was 200 times lower than with a traditional photo shoot, while testing showed a 34% increase in conversion rate. The website also generated more revenue from the same level of traffic.
What this can look like in your business
The same principle applies to content, product data, reporting, customer responses and administrative work. AI is particularly useful where a team handles a large volume of similar tasks but each output still needs its own context.
What to avoid
AI should not be introduced simply to generate more output. The output needs to solve a specific problem — whether that means improving speed, reducing costs, handling greater volume, personalising experiences or making high-quality information more accessible.
What to do instead
Look for processes where the company handles a high volume of similar work but still needs to tailor the result to a specific product, market, customer or internal rule.
A useful question for this stage is:
Which repetitive tasks could AI help us complete faster, more cost-effectively or at a larger scale?
Conclusion
AI adoption in companies is gradually moving beyond experimentation and into practical use. That is also where one of the biggest risks appears: good ideas can quickly turn into a collection of disconnected initiatives that nobody properly measures, governs or takes through to completion.
A simple framework helps:
- start with a specific goal,
- choose an AI use case with clear value,
- define rules for data, costs and control,
- look for work that AI can scale.
A good place to start is with processes that are already part of everyday operations: customer support, order processing, administration, content, reporting or approvals. These are often where AI can save time, make data easier to work with or prepare better information for people to act on.
Where could AI create value in your business?
Do you have a process you want to make faster, simpler or better connected with your data? We can look at it together and identify where AI can bring real value, what data and rules the solution will need and how its impact can be measured.
If AI is the right approach, we will design a way to integrate it into your digital product or internal process. If a simpler solution would work better, we will tell you.
Frequently asked questions about introducing AI into a business
Where can AI create the most value in a business?
AI has the greatest potential where teams work with large volumes of text, documents, customer requests or data, and individual cases require some degree of context. Typical examples include customer support, order processing, reporting, administration, product data management and content creation.
How should we choose our first AI use case?
Start with a specific process where both the problem and the expected outcome can be clearly defined. A strong AI use case has a clear user, accessible data and a measurable objective — such as reducing processing time, lowering costs or enabling the same team to handle a larger volume of work.
When is conventional code a better choice than AI?
Conventional code is usually the better option for precise calculations, unambiguous rules and stable processes where the system should perform the same step every time. AI becomes more useful when the task involves unstructured content, interpreting meaning, ambiguous input or several signals that need to be considered together.
Does AI need access to all of our company data?
No. AI access should be defined by the specific task it needs to perform. From the start, it is important to decide which sources the solution can use, which information it does not need and which sensitive data should remain outside its reach.
What are safety nets in AI solutions?
Safety nets are rules and technical mechanisms that define the boundaries of an AI solution. They can specify which data AI may access, which actions it can perform independently, when human approval is required, how costs are monitored, how outputs are logged and what happens when a result is incorrect or uncertain.
Does a person need to review every AI output?
It depends on the use case and the level of risk involved. In some scenarios, AI can work independently. In others, it is better used to prepare a recommendation while the final decision remains with a person. The right boundary depends on the potential impact of an error on customers, data or the operation of the system.
How can AI costs be kept under control?
Cost monitoring should begin at the solution design stage. It is important to understand request volumes, model and token usage, the tools involved and how usage is expected to grow. At higher volumes, it is worth regularly reassessing which tasks genuinely need AI and which can be handled more efficiently by conventional code.
How do we know whether AI is actually helping the business?
The impact should be measured against the original objective. Depending on the use case, you might track the time required to complete a task, cost per output, number of processed requests, automation success rate, conversion rate or the amount of work that still needs to be completed by a person after AI has done its part.
Do we need to start with a large AI project?
No. A more practical approach is to start with one specific process where both the potential value and technical feasibility can be tested quickly. If the use case proves successful, the solution can then be expanded or connected with other parts of the system.
