Before investing in AI, it pays to ask whether the real problem is data that does not flow between systems. Often it is. In that case integration and simple automation solve it faster, more cheaply and with less risk, and AI can follow later where it shortens work that is already well defined.
AI is often asked to fix work caused by broken data: the same information entered into two systems, a status nobody trusts, and processes nobody owns. The first step is then rarely a model. It is making the systems talk to each other.
Signs that the problem is in the data
- Staff copy names, ID numbers, dates or amounts from one system into another, often every day.
- The same question, such as the status of an order or stock levels, gets different answers depending on which system you look in.
- A spreadsheet maintained by one person is in practice the company's central system.
- Errors are discovered at month-end close rather than when they happen.
- The process stalls when a particular person is on holiday, because only they know what should happen next.
If several of these apply, AI would most likely just speed up the same manual work or, worse, base suggestions and summaries on the wrong figures. A language model fed inconsistent data produces a convincing answer, but not necessarily a correct one.
Automation that pays off first
- An order, invoice or application moves automatically between the website, accounting and email without anyone retyping it.
- Staff stop copying names, dates and status between systems, because the information is entered once and flows from there.
- An error shows up in one place, with an alert to the person who should act on it, instead of collecting in a spreadsheet.
- The next step in the process happens without someone having to remember it, such as a confirmation to the customer, a reminder or a task assignment.
- Time records, absences and expenses flow straight into payroll and accounting instead of being compiled by hand at the end of the month.
These are not exciting projects, but they are predictable. The rules are known, the result is either right or wrong, and the time saved is easy to measure. For many businesses a handful of well-chosen integrations delivers more than any AI project, and they make the next step, whether that is AI or something else, much cheaper.
How integration actually works
Most modern systems offer an API, a web service that lets other systems read and write data. An integration uses these interfaces to move data between systems. It can be real-time, where a system signals as soon as something happens, for example through a webhook when a new order arrives. It can be a scheduled job that fetches and reconciles data every hour or every night. Or it can be a middleware layer that receives data from many sources and passes it on.
For simple connections between common cloud services, ready-made tools such as Zapier, Make or n8n are often enough, and there is no need for custom work. A custom build pays off when the integration has to handle systems without ready-made connectors, complex rules, high volumes, or data that needs a full audit trail. In every case it needs monitoring. An integration that stops working without anyone noticing is dangerous, because people keep trusting the data.
When AI adds something
AI is useful when the data is already in the right place and the job is to classify, summarise, translate or flag exceptions. It handles what fixed rules struggle with: free text in email, scanned receipts, and the same request phrased in many different ways. There it shortens work that is well defined but hard to program.
It is a poor first investment when nobody knows which number is correct. We therefore start with integration and automation, and add AI where it shortens work that is already defined. A receipt scanner that reads the amount, merchant and VAT from a photo is a good example. It is only useful because expenses already have a place in the system and flow on into accounting. Without that integration it would just be a faster way to create data that someone still has to type in elsewhere.
The right order reduces the risk
The order matters more than the choice of technology. Each step builds on the one before and makes the next one cheaper.
- A clear process. Write down what happens, in what order, and who owns each step. It often turns out that part of the work can simply be dropped.
- Data in one place. Decide which system is the source of truth for each type of information, such as customers, prices and stock, and have the other systems read from it.
- Automation. Connect the systems so data flows on its own and the next steps happen without reminders, with monitoring that alerts you when something fails.
- AI. Add it where what remains is work with free text, documents or judgement calls that fixed rules cannot handle.
AI comes in once the first three are in place. Then its benefit can be measured, because the baseline is clear. And if it adds nothing, it can be switched off without the process falling apart.
Common questions
Isn't automation just old-fashioned AI? No. Conventional automation follows fixed rules that someone wrote and always does the same thing in the same situation. That is a strength: it is predictable, cheap to run and easy to test. AI is more flexible and, for the same reason, less predictable.
Can we connect systems that have no API? Often, yes, for example through scheduled file exports or direct database access, but those connections are more fragile and need closer monitoring. Sometimes replacing the system is wiser than building a brittle integration.
What does an integration cost? It depends on the number of systems, the quality of their interfaces and how complex the rules are. As an illustration, not a quote: a simple connection between two systems with good APIs might take 20 to 40 hours, roughly ISK 360,000 to 715,000 excluding VAT at our hourly rate of ISK 17,880. Monitoring in operation comes on top.
What if our software already has built-in integrations? Then use them first. Many software products ship with ready-made connections to common accounting and payroll systems, and those are almost always cheaper and more reliable than a custom build.
How to start
Pick the process with the most manual work and count how many times the same information is entered. Map where the data comes from, where it goes and where it gets stuck. That picture usually shows right away whether the first step is an integration, an automated workflow, or simply a better procedure.
If you want help with the mapping, Novamedia usually starts this kind of work with Discovery and planning, a fixed price of ISK 214,560 excluding VAT for 12 hours, which ends with prioritised recommendations and an indicative cost estimate. We have connected Icelandic accounting systems, international cloud platforms and custom-built solutions for almost twenty years, and if an off-the-shelf tool is enough, we will tell you.