AI delivers real value in day-to-day operations when it takes over specific, repetitive tasks that eat into people's time: sorting requests, reading documents, drafting replies and finding answers in the company's own material. It rarely delivers much when it is treated as a goal in itself or expected to transform the whole business at once.
The AI conversation swings between hype and suspicion, and both miss the point. After building AI into our own products and into client solutions, our conclusion is simple: the technology pays off where the work is well defined, happens often, and a person can easily judge whether the result is right.
The tasks that pay off first
- Document and request processing. Inbound email is sorted by topic and priority, invoices and forms are read and filed, and long documents are summarised.
- Draft replies. Service teams get ready-made drafts based on earlier conversations and the company's knowledge base, and approve or adjust them instead of writing every answer from scratch.
- Data cleanup and entry. Information is moved between systems, standardised and validated.
- Searching your own knowledge. Staff ask in plain language and get answers from the company's manuals, contracts and procedures, with a reference to the source so the answer can be checked.
- Summaries and exceptions. AI reviews time records, website traffic or sales figures and points out what changed, so nobody has to read through the tables every week.
What these tasks have in common is that mistakes are visible and cheap. If a draft reply is unusable, the person sees it at once and writes their own. If a document is misread, it shows up in review. That makes it possible to start with a human approving every result and to reduce oversight as the quality is proven.
Where AI is a weaker first step
AI is a poor fit where a single error can do serious damage and nobody reviews the outcome, such as decisions about people's rights, lending or hiring. In those areas it should at most support a person who carries the responsibility. Nor is it the right tool for exact calculations, accounting or pricing that can be programmed with fixed rules. There, conventional software is both cheaper and more reliable.
The most common reason an AI project disappoints is something else, though: the data is not in one place. If the same information is typed into two systems and nobody knows which figure is right, a language model will not fix it. Integration and simple automation are then the better first investment, with AI coming afterwards. Sometimes an existing product is enough. If the goal is to answer customers faster, a helpdesk with built-in AI drafts may solve it without any custom build.
Language and privacy deserve extra attention
Two things matter more in a smaller language market like Iceland than in many other places. The first is language. General-purpose models now handle Icelandic well, but quality varies between models and tasks. Understanding a request is easier than writing fluent, correct text that goes straight to a customer, and specialised vocabulary in law, healthcare or engineering can cause problems. Quality has to be verified on real examples from your own operations, not on generic benchmarks.
The second is privacy. Data sent to an AI service is processing under GDPR like any other. You need to know which data may go where, where it is processed and stored, whether the provider uses it to train models, and whether a data processing agreement is in place. Business tiers of the larger AI services generally offer processing within Europe and a commitment not to train on your data, but that has to be confirmed case by case. Staff also need to know the rules, because a lot of AI use happens in consumer chat apps that nobody has approved.
What drives the cost
The cost of an AI solution is rarely driven by the model itself. Usage fees for language models are usually a small share of the total at moderate volumes. The heavier items are integrations with the systems the solution reads from and writes to, the state of the data, how much testing and quality evaluation is needed, and how human approval and access control are implemented.
As an illustration, not a quote: a focused pilot on one process with one or two integrations might take 40 to 80 hours, which at our hourly rate of ISK 17,880 excluding VAT comes to roughly ISK 0.7 to 1.4 million. On top of that comes operation: model costs, monitoring and regular quality checks, because models change and what worked in spring needs testing again in autumn.
Start small and measure
The biggest failures we see are grand AI programmes meant to transform everything at once. They run aground on expectations nobody defined. The better path is to pick one well-bounded task, set a measurable baseline such as request turnaround time, and run a small pilot for a few weeks. If the benefit shows up in the numbers, scale it. If not, you have lost little and learned a lot. Before the pilot starts, work through the following.
- Choose one process that is repetitive, takes measurable time, and where a person can easily judge whether the output is correct.
- Measure the current state: how many requests, documents or cases go through each week and how long each one takes.
- Define success in advance, for example shorter turnaround, fewer errors, or the share of drafts sent with little or no editing.
- Decide which data may go into the solution, where it is processed and who is responsible for it.
- Have a person approve every result during the pilot and record when they had to correct or reject it.
- Collect a set of real examples to test quality against, and rerun it every time the model or the instructions change.
- Set a date for deciding whether to scale, adjust or stop.
Common questions
Do we need to train our own model? Almost never. Most operational use cases rely on an existing language model that is given access to your data at the moment of the question, for example through document search. That is cheaper, easier to update, and your data does not become part of the model.
Is it safe to put sensitive data into AI? That depends on the service and the contract. With a business service, processing within the EEA, a data processing agreement and clear access control, personal data can be handled responsibly. With a free chat app on an employee's phone, it cannot.
Will AI replace staff? In the tasks where we see the most benefit, it takes over the repetitive part of the work so people can focus on what requires judgement. In customer service that usually means the same team answers more requests faster, not that the team shrinks.
How soon should we see results? In a well-bounded task, a few weeks should show whether the pilot delivers anything. If nothing shows up by then, the task was probably the wrong choice.
The next step
Start by listing the three to five tasks that take the most staff time each week and are done in much the same way over and over. Rank them by time spent, risk, and whether the data is already accessible. The task at the top of the list is usually the right pilot.
At Novamedia we use AI in our own products every day, for example in Spjallbox, where it drafts replies that staff review, and in Tímabox, which reads receipts from a photo. If you need help finding the right first task, we regularly run these opportunity scans. Discovery and planning is a fixed price of ISK 214,560 excluding VAT for 12 hours, and ends with prioritised recommendations and an indicative cost estimate, not a slide deck. If it turns out an existing product solves the problem, we will say so.