Artificial intelligence is neither a company’s enemy nor its friend. It is a tool. That may sound trite, but it is exactly this basic principle that is easy to forget when AI comes surrounded by promises of automation, savings and scaling work without hiring more people.
ChatGPT, Claude and similar solutions are changing the way companies analyse information, prepare materials, serve customers and organise repetitive tasks. An organisation can’t ignore this. But it should start with the question: what problem do we want to solve, and is AI the right tool for it?
As Paweł Matwiejczuk, founder of Savren Partners, puts it: “AI is neither an enemy nor a friend. It is a tool – like a hammer for a carpenter or Excel for an accountant.” A tool doesn’t set the goal by itself. It doesn’t automatically understand the organisation’s context and it doesn’t take responsibility for the result.
AI doesn’t replace the manager. It changes what is expected of them
AI won’t replace a manager any more than the hammer replaced the carpenter. What it can do is change the set of tasks a manager performs and speed them up. That requires new skills: you need to be able to define the task, provide context, assess the result and spot an error the model may not flag.
The metaphor Paweł uses captures this relationship well: AI is like a very ambitious intern. It wants to help, prepares an answer quickly and often makes a good first impression. But the person giving out the work has to know the subject well enough to know what they expect and how to check the result.
If a company gives AI a task it doesn’t understand itself, it may not be able to judge whether the answer is correct. The problem then isn’t necessarily the model. It may be a lack of knowledge within the organisation, a brief that is too general or insufficient checking of the result.
So the point is not for every employee to become a prompt engineer. It is about being able to break the work down: what needs to be done, based on what information, against what criteria and in what format the result should be produced. The bigger the consequences of an error, the more the knowledge of the person using AI matters.
A study of customer service workers found that productivity improved after a generative AI tool was introduced, although the benefits varied between workers. It is useful evidence that AI can support work in a specific context, but it is no basis for expecting the same effect in every company and every process. The NBER study “Generative AI at Work” shows precisely how much the context of the task and the worker’s experience matter.
Implementing AI is an organisational rollout, not a contest for the best prompt
In a company, implementing AI should be treated like any other technology rollout: start by understanding the process and the needs of the people involved in it. Simply giving employees access to a tool won’t answer questions about data quality, accountability, workflow or rules of use.
Before choosing a solution, it is worth establishing who takes part in the process, who uses it, who approves decisions and what information is needed to do the task correctly. Then you can identify the places where the work is repetitive, predictable and based on available data. Only that is the starting point for checking whether AI can support it.
In practice this may mean mapping stakeholders and visualising the flow of work. Depending on the task, process maps, flow diagrams or UML notation can be useful. The choice of format matters less than the team agreeing on the stages and responsibilities. A start like this is, in essence, a digital transformation programme: diagnosing how things work now, agreeing the target process, choosing the technology and planning the change.
That is exactly why rolling out “AI across the whole company” is rarely a precise enough goal. For one business AI may become an important part of daily operations. For another, the best use will be occasionally preparing a table, tidying up notes or correcting a message. Building advanced infrastructure for a simple task would be an unnecessary cost; using a publicly available tool for a critical process, on the other hand, may not meet the company’s requirements.
A good rollout starts by matching the solution to the need, not by deciding to implement a particular model. The AI risk management framework developed by NIST also organises this work around understanding context, measuring and managing risk, and governance – not around the technology itself.
A faster process isn’t always a better process
If AI speeds up repetitive work, it may be a good candidate for wider use. But that doesn’t mean every speed-up automatically creates value. A wrong answer can be generated faster than a right one, and a weak process can, thanks to automation, simply produce wrong results faster.
So before scaling, it is worth checking not only the time saved but also the quality of the result, the number of corrections, the cost of errors, the load on approvers and the impact on the audience. If AI produces more campaign versions but checking them takes as much time as the work did before, the number of materials generated says little about efficiency.
The same applies to AI agents working in the background. If a company wants to automate repetitive activities, it should first know what the process looks like today and which elements can be safely delegated. It also needs to decide what permissions the agent will have, which data sources it can use, when a result needs approval and what to do when the system hits an exception.
Not every process will be a good candidate. Complex tasks with many hidden dependencies, a high cost of error or frequently changing criteria may need much more oversight. The most sensible approach is to start with limited use, observe the results and only then widen the scope. That isn’t a lack of ambition, but a way of checking whether the assumptions behind the rollout survive contact with real work.
Start with a pilot that has a clear measure
Before launching a pilot, the team should describe how the task is done today and collect a baseline: time taken, number of corrections, typical errors and the time needed for checking. Then choose a small scope and agree who assesses the results and when a person has to take over. After a few weeks you can compare the results with the baseline, and ask employees whether the new solution really makes their work easier.
If the time saving is visible but errors increase, or checking eats up the time gained, the rollout needs improving. When the results are good, expansion can be planned in stages rather than automating every team at once. A pilot should test both the tool’s capabilities and the organisation’s readiness to work in a changed process. It is important that its scope is tied to a real business goal.
Company data needs clear rules
An important part of a rollout is deciding what data employees may pass to the chosen tool. Not every organisation can accept the same rules. What matters is the type of data, the account settings, the terms of service, the workspace configuration, the retention period and how access is controlled.
It is particularly important to distinguish between data used to train models and other aspects of how it is processed. For example, OpenAI states that content from personal accounts may be used to improve models unless the user switches this off, while data from Business, Enterprise and API plans is not used for training by default. Anthropic states that it doesn’t train models on inputs and outputs from its commercial products unless the customer agrees to this or submits materials through a specific feedback route. The specific terms should be checked for the given product and configuration. OpenAI explains its data usage policies.
A commercial plan can be one element of limiting risk, but buying a licence doesn’t replace the organisation’s own policy. The company still needs to know what information may be entered into the system, who is responsible for the decision and whether the solution meets its legal, security and data retention requirements.
Using internal infrastructure or a solution based on RAG – retrieval-augmented generation, i.e. generation supported by retrieving information from a controlled knowledge base – may be justified in selected cases. One example is an organisation that wants to give the model a defined knowledge base with access control. But it isn’t the best choice for every company by default. For simple uses, implementing such an architecture can be expensive and out of proportion to the benefits.
In marketing, it isn’t only the volume of output that counts
Marketing has many potential uses for AI: data analysis, preparing content variants, organising product knowledge, supporting customer service or optimising campaigns. But it is worth distinguishing between producing more materials and producing materials that help achieve a business goal.
Generated content can be useful if it is credible and consistent with the brand context. Otherwise automation increases the scale of the risk. A graphic with an incorrect packshot, text containing unverified information or a message that contradicts the brand guidelines may need fixing later, and in some cases may damage the audience’s trust.
That is why the team should decide where AI can prepare a draft and where expert knowledge and approval are needed. If a model generates content proposals, someone still has to be responsible for the accuracy of the information, the correctness of the message and its consistency with the brand. The more the result matters, the greater the need to control sources and the approval process.
Goal first, technology second
The board should know what problem it wants to solve and what target state it wants to reach. Only then can it judge whether AI is the right tool, whether the process needs changing and who should be responsible for the rollout. Technology shouldn’t define the organisation’s strategy on its own.
As Paweł Matwiejczuk sums it up: “If AI genuinely speeds up our work, implementing it may make sense. But the measure isn’t the number of campaigns or documents. The measure is value: a lower CAC, less time spent on a campaign or less load on the team.” So it is worth starting sensibly with one well-understood process, establishing a baseline and measuring the change. If the solution improves the result without lowering quality or creating new risk, scaling it can be considered. If not, the company has gained valuable information before investing in a broad rollout.
Aligning processes, accountability and technology involves more than choosing a tool. See how Savren approaches a digital transformation programme, and read the case study of a strategy the board approved. You can read about mapping responsibilities in our article on stakeholders.
Facing a similar challenge? Let’s talk about how we can help your organisation.
