In February 2024, Klarna published figures that stunned the service world: its smart assistant, built on OpenAI technology, handled 2.3 million service conversations in its first month — two thirds of all inquiries. Work equivalent to 700 full-time agents. Average handling time fell from 11 minutes to under 2, and the company estimated a profit improvement of 40 million dollars a year. And then, a year later, came the part that gets quoted far less often: the CEO publicly admitted that the company had gone too far, that service quality had suffered — and brought human agents back to the contact centre.
Both sides of the Klarna story are exactly why this guide exists. Implementing artificial intelligence in a business works — the numbers are unequivocal — but only when it is done properly: the right problem, real measurement, and people inside the process. Here is what the research and the field really teach us.
What the documented numbers say
A field study by researchers from MIT and Stanford followed 5,179 support agents who were given an AI assistant. The result: an average 14% rise in output — but among the newer, less experienced employees the jump was around 34%, while among the most experienced the impact was almost nil. The practical conclusion: AI does not replace the strong performers — it lifts everyone else to their level faster.
A Harvard experiment with 758 BCG consultants found that consultants who worked with GPT-4 completed 12.2% more tasks, 25% faster, and at 40% higher quality. But the same study also found the other side of the coin: on tasks that fall outside the model’s real capabilities, those who relied on it were 19 percentage points less accurate. In other words — AI in the right place multiplies your strength; AI in the wrong place actively does damage.
Why more than 80% of AI projects fail
A 2024 study by the RAND Corporation, based on interviews with 65 data scientists and engineers, found that more than 80% of AI projects fail — twice the rate of ordinary IT projects. The reasons that came up again and again: management defines the wrong problem, there is no high-quality, available data, the organisation chases the technology instead of a business problem, and the infrastructure is not ready. Notice what is not on the list: the technology itself. The failures are almost always managerial, not technical.
The AI implementation process in a business: 5 stages
Stage 1: Mapping — where time and money are being burned
The opening question is not “how will we use AI” but “which process costs us the most hours per month”. A one-to-two week mapping exercise: which tasks repeat themselves, how long they take, and what an hour costs. A process measured in dozens of hours a month is a candidate; a process measured in a handful of hours — not worth the investment yet.
Stage 2: One pilot, measurable, 30–60 days
You pick one process and one clear success metric: handling time per enquiry, cost per lead, work hours saved. Klarna did not roll its assistant out across everything in a single day — it measured conversations, time and satisfaction against human agents. A pilot without a target number is not a pilot, it is a toy.
Stage 3: Data and infrastructure
Most of the failures on RAND’s list land right here. A model that leans on information scattered across spreadsheets, emails and employees’ heads will deliver results that are just as scattered. Before connecting AI: consolidate the information, clean it, and make sure the system has orderly access to it. This is the least glamorous part — and it is the one that decides the outcome. We expanded on this in the guide to data analysis systems for businesses.
Stage 4: People before machines
The best-documented example is IKEA: the Billie bot handled 47% of contact-centre enquiries over two years — and instead of making people redundant, IKEA retrained 8,500 contact-centre employees as interior design advisers, a role that generates new revenue. Training and reskilling the team are part of the project budget, not an optional extra. An employee who has not been trained will work around the system, and the pilot will fail regardless of how good the technology is.
Stage 5: Measure, then scale
After 30–60 days you check the metric against the target. Did it hit the mark? You expand to the next process using the infrastructure you have already built. Did it fall short? You work out why before you invest another shekel. Gradual expansion is the difference between an organisation that has built a capability and an organisation that has bought a subscription nobody ever opens.
How much does implementing artificial intelligence in a business cost
Realistic ranges as of 2026: off-the-shelf tools (ChatGPT Team, Copilot, Claude) — tens to hundreds of shekels per user per month. Automating a single process with tools such as Make or n8n connected to a language model — a one-off setup of a few thousand shekels. A smart agent or custom development (a service bot built on your own knowledge, a system that handles leads) — tens of thousands of shekels and upwards. The cost everyone forgets to budget for: the implementation and team training hours, which are usually an additional 20%–30% — and without them the rest of the investment goes to waste.
Frequently asked questions
How long does implementing AI take in a small or medium-sized business?
A focused pilot: 30–60 days including definition, setup and running. A broad implementation of 2–3 processes: one to two quarters. Anyone who promises “full implementation in a week” is selling a tool, not a process.
Do you need an in-house technical team?
Not necessarily. Off-the-shelf tools and automations are set up today without writing any code. What you do need: an internal owner responsible for the process and for the measurement, and professional guidance in defining the problem — because that, according to RAND, is where most projects fall down.
Where do you start?
With the process that repeats itself the most times per month and that has available data. In most businesses it is one of three: answering service enquiries, handling leads, or producing documents and reports.
How do you measure the success of an AI implementation?
One business metric defined in advance: hours saved, response time, cost per transaction, the percentage of enquiries closed without a human. “The team is happy” is a nice outcome — not a metric.
So what now?
AI implementation is Rainz’s core service — mapping, a measurable pilot and gradual rollout following exactly the process described here. You are welcome to read about the artificial intelligence implementation service for businesses or to start with the complementary guides.
Related guides
- Artificial intelligence for businesses: the complete guide
- AI sales agents
- Data analysis systems for businesses
- The 88 leading AI tools
- Digital marketing strategy: the complete guide for your business
- Information systems specification: the complete guide for businesses 2026
- Data management and analysis for your business: the complete guide 2026
- The screen is dead. The agent is in: from UX to AX — an opinion piece
- Artificial intelligence for sales: the complete guide
This guide is aimed at small and medium-sized businesses. Organisations with core systems, governance requirements and information security needs will find the process that suits them on the page AI implementation in enterprises.