When large language models arrived, AI suddenly became accessible to everyone. At the same time, concepts are often confused. Traditional machine learning, which is used for tasks such as fraud detection and pricing based on real-time data, is a separate discipline with its own methods and technologies.
This article focuses on language models and how organisations can use them to create genuine value rather than simply follow the hype.
The chatbot nobody used
Some organisations moved quickly with their first AI projects. Concerned about being left behind, they invested in chatbots and automated solutions, expecting significant efficiency gains. What many underestimated was that deploying AI in a large organisation is very different from using ChatGPT at home. Access rights, data quality, security requirements and what systems are actually permitted to do all have a major impact on the pace of implementation. In regulated sectors such as banking, financial services, insurance, energy and healthcare, these considerations become even more important.
The standard response has often been to invest in training and change management to “bring employees on board”. But consider this: nobody needed a course to start using ChatGPT at home. Professionals taught themselves how to use specialist tools for coding, design and analysis. When employees are not using an internal chatbot, it is rarely because they lack training. More often, it is because the tool simply is not good enough. Shadow AI proves the point. If employees are paying for personal subscriptions out of their own pockets to get work done, the verdict on the internal solution has already been delivered.
The AI journey has two phases
If your organisation has not yet thrown itself into the AI wave, this article is for you. You have the opportunity to avoid costly mistakes and focus directly on what works. However, that requires accepting that the journey consists of two phases: first a modernisation phase, where the organisation becomes AI-ready, followed by the transformation itself, the agentic layer.
Our principles are simple:
- Do not start until the data foundation and pipelines are in place. Everything else depends on this.
- Own your technology stack to maintain control over data flows and scalability.
- Build tools that are tested and deliver value internally before rolling them out to customers.
- Think lean: ship quickly, measure outcomes and iterate.
In practice, this means starting small, building on the right foundation and expanding only when the solution has demonstrated value. Five actions are particularly important.
The recipe: five practical steps
1. Map shadow AI
Conduct an anonymous survey of which AI tools employees use privately for work-related purposes and what they use them for. The responses provide a practical needs assessment. Where employees are already finding their own solutions, the first opportunities for value creation are often hiding. At the same time, the organisation gains greater visibility into where data may currently be at risk.
2. Choose one content area and bring order to it
Not the whole company, just one area. For an insurance company, policy documentation is a natural place to begin. For other organisations, it may be procedures, contracts or product documentation. Gather content from SharePoint, Confluence, email and intranet systems, and ensure the collection process is automated. New policies and procedures are introduced continuously, and models should always work from the current version, not last year's.
3. Build the knowledge base and own it
Everything should be brought together in a knowledge base, a database designed to help models find the most relevant answer rather than simply exact matches. Two lessons learned through experience: think about access management from day one, as not all content should be available to everyone. And build it yourself rather than relying entirely on off-the-shelf solutions such as Microsoft Copilot. This is about controlling the technology stack, the data flow and the ability to scale in the direction you choose rather than the direction your vendor chooses.
4. Launch an internal assistant for one team
Again, not the entire organisation. An adviser assistant for case handlers is a natural place to start. Questions such as “What does the cancellation policy cover?” can be answered in seconds with source references, rather than requiring ten minutes of searching through documents. There is a simple reason for starting internally: the system needs a feedback loop to improve, and mistakes are less costly internally than they are with customers.
5. Measure, learn and expand
Define two or three key metrics before launch: time per case, the proportion of questions the assistant can answer and team satisfaction. Ship quickly, measure, adjust and expand only when results are strong enough to justify moving to the next team or content area. And remember, the benefit is not only time saved. Employees who are freed from repetitive tasks can focus on work that genuinely requires human expertise.
This is not a three-year programme. With the right sequence of actions, the first benefits can be realised within a matter of months.
The next decade will be agentic
If we are allowed to look into the crystal ball, we believe the next decade will be defined by digital agents: systems that perform tasks independently, collaborate with one another and handle repetitive work. Most tools already have interfaces that models can connect to, so the challenge becomes defining clear rules for when and how those tools should be used.
The objective is to put the system into a loop. A customer enquiry arrives, the agent reviews previous conversations, logs and the knowledge base, and either provides an answer itself or routes the case to a human when necessary. The same approach applies to analysis, design and software development. Extend the thinking further and you arrive at software that manages itself, a form of DevOps on autopilot: the system detects a failure, automatically creates a task, a model prioritises and describes it, and a coding agent proposes a fix. A human approves the change.
However, the sequence matters. Agents will never be better than the foundation they stand on. That is why the five steps come first.
Foundation first, agents afterwards
For organisations looking to move beyond AI experiments and towards measurable value, the work begins with the right foundation: high-quality data, strong integrations, a robust knowledge base, internal value creation and only then the introduction of agents. At Itera, we help organisations build this foundation and develop practical AI solutions that fit naturally into everyday work.
If you would like to understand where your organisation should begin, we would be delighted to discuss the actions most likely to create value in the shortest time.
.png)
.png)
.jpg)
