AI Integration for Businesses: 10 Things That Make the Project Actually Work
TL;DR
AI integration is wiring language models, image recognition, or predictive models directly into the systems you already use — ERP, CRM, email, document archives. Most projects fail not on technology but on scope and adoption. Ten habits make the difference: start with the problem not the technology, pick one process not ten, check the data before calling a vendor, evaluate off-the-shelf first, place AI where work happens, measure the baseline, build modularly, keep a human in the loop, plan for failure, and ship something to production fast.
AI integration is the work of wiring language models, image recognition or predictive models directly into the systems your team already works in — ERP, case management, CRM, email, document archives. The goal is that the workflow does not change. What changes is that half the keystrokes disappear and some of the decisions are made ready-to-go by the machine.
AI integration is the work of wiring language models, image recognition or predictive models directly into the systems your team already works in — ERP, case management, CRM, email, document archives. The goal is that the workflow does not change. What changes is that half the keystrokes disappear and some of the decisions are made ready-to-go by the machine.
On paper it is straightforward. In reality, AI projects are where shelves fill up with pilot demos that never reach production. I have seen it often enough to have an opinion about why. Below are ten things we do when integrating AI in a business, and that seem to separate the projects that succeed from the ones that only produce a nice set of slides.
1. Start with the problem, not the technology
The most common mistake I see is starting from "how can we use AI?". Wrong question. The right one is: "where in the business do we spend the most time on something that feels stupid?" The hour-long task every Monday morning that nobody enjoys. The report that bounces back and forth three times before it is right. The emails that get answered ten times with nearly the same text.
That tells you where AI belongs. The technology is not the point. It is just the tool that solves that specific problem.
2. Pick one process, not ten
Every client has a list of ten things they want AI to help with. That is fine. But the first integration should only solve one of them.
The reason is simple: one successful delivery in production builds internal trust. You also get real numbers to show — time saved, fewer errors, shorter response time. With that in hand, it is far easier to get budget and resources for the next nine. Without it, you are still arguing whether AI is worth it eighteen months later.
3. Check your data before you call a vendor
AI is not magic. It is glue. If your documents are spread across three SharePoints, invoices are in PDFs with handwritten notes, and customer data has been entered by 40 different people over 15 years, then at best AI helps you tidy up — at worst it just makes the mess faster to retrieve.
Before you start an integration: figure out where the data lives, who owns it, and how clean it is. It is fine to do a bit of tidying first. That is always cheaper than tidying mid-project.
4. Do not build anything before you look at off-the-shelf
There are now solid standard tools for the most common AI needs — Microsoft Copilot for office work, specialised SaaS for customer service, or AI features already built into the line-of-business system you are using. Before we build anything custom, we always check whether an existing tool covers 80 percent of the need at a fraction of the cost.
Custom has its place. But most problems are not as unique as you think they are.
5. Put the AI where the work happens
If an employee has to open a new tab, log into yet another tool, and paste text in to get AI help, the tool will not be used. Not after two weeks. Not even after one week.
The best integration is the one you do not notice. It sits inside the Outlook email, inside the form in the line-of-business system, in the Teams chat. Employees keep working as before, and the AI runs in the background — or suggests something inside the interface they are already in.
6. Measure the time you save before you start, not just after
This sounds obvious, and we forget it every time: if you do not know how long a task takes today, you cannot say whether the AI integration was worth the money.
At kickoff we always ask the team to stopwatch the task for a week. Not precisely, but well enough. Then you have an honest number to measure against later. When a consulting firm on Forus was going to automate their reporting, the stopwatch said each project report took 4 hours and 20 minutes on average. After integration: 1 hour and 10 minutes. That is a sentence that sells the next project internally far better than "it is faster now".
7. Build as if the model will be replaced in eighteen months
Language models are getting better and cheaper at a pace that is hard to compare to anything else in IT history. The model you use today is not the one you will use in 2028.
That is why we never bake AI logic directly into the production code. The model should be swappable without touching the rest of the system. The technical term is an abstraction layer. Practically it means you save money every time a new vendor releases a model that is half the price and twice as good. And that happens more often than you expect.
8. Keep a human in the loop — at least at the start
It is tempting to let the AI do the whole job from day one. Do not. It makes mistakes. And those mistakes are often subtle enough that nobody notices until some customer complains.
We almost always start with AI suggesting, and a person clicking "approve" or editing. After a few weeks we look at whether the error rate is low enough to remove the approval step for the simplest cases. Often the answer is yes for 80 percent of cases and no for the remaining 20. That is a healthy balance.
9. Plan for the failures, not just the success
What happens when the AI model is down? When it returns a wrong answer? When it suddenly replies in English because a vendor update changed the behaviour? You need an answer before you launch.
In practice that means logging every response, alerting on anything that deviates from expectation, and a clear fallback — what happens when the AI does not deliver. If an email cannot be classified automatically, it should fall back to a manual queue, not just disappear.
10. Get something into production fast, even if it is ugly
No pilot teaches you what a production user does. A pilot is always pretty, the data set is always hand-picked, and the users are always on their best behaviour.
We have had good results getting a minimal version out to a small group of real users after four to eight weeks. Yes, it is missing features. Yes, the design may be rough. But the feedback you get in one week in production is worth more than six months of internal demos. Build from there.
Systems we integrate with most often
Most modern business systems have an API that makes integration possible. We work regularly with Microsoft 365 and SharePoint, Tripletex, 24SevenOffice, PowerOffice and Visma, Salesforce and HubSpot, Altinn and EHF, and bespoke line-of-business systems in fields from construction to aquaculture. If the system lacks an API, there are usually workable alternatives — RPA (robotic process automation) can click its way through an older interface until a proper integration is in place.
What does AI integration cost?
Short and honest: it varies a lot. A narrow integration — say, an AI that reads incoming email and sorts it — lands at NOK 50,000–150,000 finished in production. An integration that ties several systems together and needs its own logic typically NOK 150,000–400,000. Larger efforts with a fine-tuned model, a dedicated vector database or advanced approval flows can land between NOK 400,000 and 1.2 million.
The most important number is not the price tag but what you get back. If one integration saves two employees half a day per week each, it pays for itself within a year — without making anyone redundant. That is the calculation we look for when sizing projects.
How to get started
Do not start with an AI workshop. Start with a conversation about what steals time in your business right now. Everything else follows.
At WOLLUM we use the first discovery to find one process that is ready for AI integration, and give a concrete proposal with duration, price and expected time saved. If we do not find a good candidate, we say so. We are not afraid to recommend that you wait six months if the data or processes are not ready. It is cheaper than a project that never reaches production.
Does your business have a process that steals time but really follows a recipe? Contact WOLLUM for a no-obligation conversation. We are based on Forus but deliver AI integrations to businesses across Norway.
Frequently asked questions
- Do we have to replace the systems we already have?
- Rarely. The point of integration is that the systems stay. We connect AI to them via API or similar, and the workflow you have today stays unchanged. If an existing system is so old that it cannot be integrated, there are usually options — but we need to talk about it before we promise anything.
- Is it safe to use AI with sensitive data?
- It can be, but it depends on how you do it. We use EU regions with vendors, sign data processing agreements that are actually read, and anonymise where it makes sense. For particularly sensitive data — health records, privileged legal material — there are options that run locally or in Norwegian cloud. We always run a GDPR assessment together with the client during discovery.
- How long does a typical integration take?
- A narrow integration is in production in four to eight weeks. Larger efforts spanning several systems take three to six months. It is slower than many expect because most of the time goes into cleaning data, sorting access rights and testing properly — not into the AI itself.
- What if the model stops behaving as expected?
- It will happen. The vendor updates the model and the behaviour shifts slightly. That is why we log all responses and alert when something deviates from expectation. The system always has a safe fallback too — usually that the task ends up in a manual queue, so no messages get lost.
- Can we start small without committing to a big project?
- Yes. Most of our clients start with one narrow integration at a fixed price. Once it is in production, we evaluate together whether it is worth doing the next one. There is no obligation to continue.
