“We need to increase AI adoption by 30% next year.”
Sounds familiar?
Throughout 2026, we’ve heard variations of this statement in conversations with companies and executives. Sometimes the target is 30%, sometimes it’s 50%, and sometimes there’s no number at all. Just a general feeling that everyone else is doing more with AI and we’d better catch up.
Fair enough. There’s a lot happening, and the opportunities are real. Nobody wants their business to miss an important shift.
But whenever we hear these targets, there’s a question we like to ask:
What exactly are we trying to improve?
More employees using ChatGPT? More prompts sent? More code generated? More AI features shipped?
Suppose we double AI usage across the company, but nothing improves in terms of cost, quality, speed, or customer satisfaction. Have we achieved anything?
Actually, we might have made things worse.
We’ve paid for licences and burned through tokens for no return. We’ve consumed computing resources and energy unnecessarily. Perhaps we’ve also produced another mountain of generic reports, pointless presentations, and AI-generated slop that nobody really needed.
But hey, the AI adoption dashboard looks great! 🎉
To be clear, experimentation is valuable. People need opportunities to try new tools, make mistakes, and understand what’s possible. Usage can even be a useful metric once we know an approach is worthwhile.
The problem begins when adoption becomes the objective instead of a way to achieve an objective. That’s basically FOMO with a budget and a KPI.
Over the course of 2026, we’ve helped several organisations use AI in their day-to-day work, internal processes, and products. These engagements have been quite different, but they’ve reinforced one observation: when people say “AI adoption,” they often mean very different things.
We see at least four distinct types of adoption. They overlap, of course, but each calls for a different approach and different measures of success.
Let’s look at them.
1. Personal AI adoption: start with yourself
This might sound like a strange starting point for an article about business strategy. Bear with us.
Before asking an entire organisation to adopt AI, we think it’s worth becoming a reasonably capable AI user yourself. That applies to CEOs, managers, developers, accountants, and everyone else.
The most useful mindset is to think of AI as a personal assistant: something that can help you organise information, investigate options, clarify difficult subjects, and get through work you’d otherwise postpone.
Not an oracle. Not someone you should trust blindly. Just a surprisingly capable assistant that you can ask for help.
Start with the boring stuff
Think of the tasks you keep putting off because they take too long or are just plain tedious.
Reviewing your health insurance policy, for example. Are there benefits you’re paying for but never using? Could another plan give you similar coverage at a lower price? Or perhaps you’ve been meaning to compare utility providers, review household finances, or understand whether a particular area would be a good place to rent or buy a home.
These questions can affect your life quite a bit, but investigating them properly takes time. And reading insurance terms and conditions after a day of work isn’t exactly a thrilling evening activity.
AI can help you collect the relevant information, understand unfamiliar terms, compare alternatives, and identify questions you should investigate further.
One of our favourite examples is preparing tax returns. Collecting documents, categorising expenses, checking what’s missing, and getting everything organised used to take days, even with reasonably tidy records. Using AI to help with the preparatory work can bring that down to hours. The actual filing and any important tax decisions still need proper checks, of course.
There’s an interesting effect here that goes beyond doing existing tasks faster.
AI can make it worthwhile to do things you previously wouldn’t have bothered doing at all.
Maybe comparing insurance policies could save you a few hundred euros, but you’re not willing to sacrifice a weekend to find out. If much of the investigation takes half an hour instead, suddenly that opportunity looks quite different. And that half hour doesn’t need your full attention: the AI does most of the searching and comparing, and you check in from time to time to review what it found and steer the next step.
Learn something new
Learning about a new topic is another excellent way to start using AI.
Ask AI to explain a difficult concept with examples or diagrams. Ask it to put together a learning plan, suggest books and videos, or create an interactive exercise. Better still, ask it to quiz you and challenge your understanding.
It can be especially helpful when you’re learning something outside your field and don’t yet know which questions to ask.
Of course, it can also be confidently wrong. That means checking important claims, consulting primary sources, and testing your understanding rather than simply accepting a fluent explanation.
This is one of the most valuable lessons at this stage. You start to get a feel for what these tools can and can’t produce, and you learn to treat them as advisors. You still need to build the full picture yourself and make sure the information checks out. When something doesn’t, look for other sources and keep refining your mental model. That mental model stays yours: you’re not delegating critical thinking to the machine.
Common mistakes
The first mistake is using AI only as a slightly more talkative search engine. You type a short question, read the answer, and move on. That misses where it helps most: working through a problem with you over several steps, using your own documents and context, drafting, comparing options, and challenging your reasoning.
The second mistake is asking a couple of vague questions, getting mediocre answers, and dismissing the whole thing. The third is going to the other extreme and believing every answer because it sounds convincing.
Good results often require context, follow-up questions, and a little healthy scepticism. And when working with financial, medical, or other sensitive information, you need to think carefully about what data you share and with which service. For the most sensitive material, models that run locally on your own hardware are an option, but that’s a topic for another article.
Where to start
Pick one task you’ve been postponing. Something real, with a result you can recognise.
Try completing it with AI’s help. Share the relevant context, ask questions, challenge the suggestions, and check the outcome. Then try another task.
You don’t need an elaborate prompt library or an expensive automation platform to begin.
What success looks like
Look at the time you’ve reclaimed, the opportunities you’ve identified, the tasks you’ve finally completed, and whether you’ve learned something you can actually apply.
Some of this won’t translate neatly into euros. That’s fine. Personal value doesn’t always need a spreadsheet.
There’s also a useful side effect: you start developing a feel for what AI does well, when it needs help, and when it’s best avoided. That experience becomes valuable when you bring AI into your professional life.
2. AI as a work assistant: make people more effective
Once you start using AI for personal tasks, similar opportunities tend to become obvious at work.
There’s no shortage of repetitive administration, document review, research, reporting, meeting preparation, and information scattered across different systems. AI can help with all of these, but the best use cases are usually more specific than “everyone should use a chatbot.”
A CEO who needed more time for the actual job
In one engagement, we helped a CEO who was spending a lot of time on the more tedious parts of running a business.
Preparing tenders involved considerable administrative work. Meetings generated decisions to track. Follow-ups needed to be organised and scheduled. Keeping on top of important business activities required constant attention.
We introduced AI-assisted ways of streamlining parts of the tender process, organising meetings, and handling follow-ups more systematically.
The point wasn’t to automate executive judgment. It was to free up attention for the decisions, relationships, and opportunities that really needed it.
That’s a much more useful objective than getting a CEO to use AI twenty times per week.
Better customer support, without replacing the conversation
In another project, we built an internal AI-powered knowledge base for a customer support team.
A common reaction to “AI in customer support” is to build a chatbot that talks directly to customers. Sometimes that’s absolutely the right choice. In this case, we deliberately didn’t do that.
Instead, operators use the system to find relevant information and prepare detailed, accurate, and consistent responses much faster. They still review the material and communicate with customers themselves, often adding the human touch that makes customers feel valued and keeps the relationship genuine rather than transactional.
The team can handle more tickets per operator, with greater confidence and less time spent hunting for answers. Customers get faster, more consistent support, while the business keeps the human interaction it values.
Replacing a person isn’t the only way to improve a process. Sometimes helping that person do a better job is the better investment.
Common mistakes
One frequent mistake is buying licences for everyone before identifying a useful problem to solve. Another is celebrating weekly active users or prompt counts without checking whether work or any relevant business metric has improved.
And there’s an easy trap: automating a process that shouldn’t exist in the first place. Making a pointless report 80% faster is nice, but perhaps it would be better to stop producing it.
Where to start
Pick one team. Look for a recurring task that’s slow, frustrating, or prone to mistakes. Measure how it’s currently done, then trial an AI-assisted alternative with a small group.
Measure the whole process, including the time spent checking and correcting the output. Ask everyone involved for feedback too, from the people doing the work to the people who receive it, including customers where relevant. Faster isn’t better if the experience gets worse. If the experiment works, figure out how to make it repeatable and safe for the rest of the team.
What success looks like
Useful metrics include turnaround time, error and rework rates, cost per completed task, team capacity, and satisfaction among the people receiving the work.
One caveat: ten hours saved don’t automatically mean ten hours’ worth of money saved. The benefit depends on what those hours allow the team to do. Perhaps they can serve more customers, avoid additional hiring, reduce overtime, or spend more time on valuable work.
Measure that outcome, not just the minutes apparently saved by a tool.
3. AI to build products: change how software gets made
This category is particularly relevant to engineering organisations.
In practice, it means engineers work alongside AI coding tools. Some suggest the next few lines as you type. Others, usually called coding agents, take a task described in plain language, read the relevant code, make changes across several files, run the tests, and propose the result for a human to review.
But writing code is only one part of building software. AI can also help with requirements, architecture exploration, prototypes, testing, code reviews, debugging, documentation, and maintenance.
At the far end of the spectrum, there’s a lot of talk about “software factories”: building entire products, from idea to production, without an engineer writing or even reading the code. From a business owner’s perspective, that can sound like the dream. In practice, it’s an extreme option where you give away most of the control over what you ship. It has its merits in some situations, but it’s rarely the best place to start.
More often, the wiser approach is to start small and add more as you learn. Use AI for well-defined tasks, keep engineers in control of the decisions that matter, and expand once you’ve seen the results. You get most of the benefits with much less risk.
The important distinction is that using AI to build a product isn’t the same thing as building AI into the product.
You might use AI extensively in development and ship a product without a single AI feature. That’s a perfectly valid strategy. Building AI features into the product is the fourth type of adoption, and we’ll look at it in the next section.
More code isn’t necessarily more progress
Modern tools can make producing an initial implementation remarkably fast. We remember when writing a few hundred lines of code could take days, sometimes weeks. Today, a coding agent can produce thousands of lines in minutes, and with the latest models, that code can be surprisingly good.
But producing code has never been the whole job.
We still need to understand requirements, figure out what works best for our users and test those assumptions, make design decisions, test behaviour, consider security, and maintain what we build.
If developers can produce changes much faster but the team can’t review or test them effectively, we’ve just moved the bottleneck. Or worse, we’ve found a much faster way to accumulate technical debt. Impressive, but not quite the result we wanted! 😅
The bigger opportunity is to reconsider the development process itself. When it’s cheaper to create a prototype, maybe we should explore more alternatives before committing. When generating tests is easier, maybe we can finally improve coverage around that difficult legacy component. When an agent can handle a well-defined maintenance task, perhaps engineers can spend more time on the parts that need their judgment.
These opportunities depend on the environment, though. Good tests, clear requirements, reproducible builds, security checks, and thoughtful reviews make AI-assisted development far more useful.
Common mistakes
Measuring generated lines of code is an obvious one. So is counting agent-completed tasks without checking whether the resulting changes are correct and maintainable.
Another mistake is looking only at the time taken to produce a first version, ignoring the effort needed to review it, fix defects, and operate it in production.
It’s worth remembering the Pareto principle. Often, 80% of the work is straightforward and takes 20% of the time. The remaining 20% is hard and takes the other 80%. AI is very good at the straightforward part, so a fast first version saves only a slice of the total effort. The real gain is that your team can spend more of its attention on the hard 20%, the part that makes your product work.
And we shouldn’t assume every developer, codebase, or task will benefit equally from the same approach.
Where to start
Find a real engineering bottleneck. Maybe it’s understanding unfamiliar code, maintaining old systems, writing tests, or turning product ideas into testable prototypes.
Select representative tasks and compare the current approach with an AI-assisted one. Include implementation, verification, review, and rework. Try the approach with more than one engineer before drawing conclusions.
What success looks like
Speed is the easiest thing to measure, and the most tempting. But faster delivery is only valuable if quality holds, especially in the parts of the product your customers use. Google’s 2025 DORA report found exactly this pattern: teams with higher AI adoption deliver faster, but they also see more failed changes and rework.
So never track speed on its own. For speed, look at lead time from requirement to production, meaningful delivery throughput, and how quickly you can validate a product idea. For quality, look at change failure rate, defect rate, production incidents, time spent on rework, and issues reported by customers.
Always read the two sets together. Shipping twice as many changes isn’t much of a win if production incidents triple.
Hold customer-facing areas to the highest standard. A rough edge in an internal tool is an inconvenience. A bug in sign-up, checkout, or billing can cost you customers and their trust.
The objective is to deliver better software, not just more code.
4. AI in your products: improve what customers actually get
This is what many people picture when they hear “AI adoption”: adding intelligent features to products and services.
That might mean better search, document understanding, recommendations, conversational interfaces, or automation of a business process. It might also mean an improvement that users never see directly.
And that’s worth remembering: an AI feature doesn’t have to look like AI to be valuable.
Insurance claims, minus a lot of manual work
One of the clearest examples from our work in 2026 involved an insurance claims management platform.
The company relied on substantial manual review of claims. During busy periods, it needed additional reviewers from external providers to handle the volume. That was expensive and slowed down how quickly cases could be processed.
We helped introduce AI into the backend to automate a large portion of the claim checks.
But we didn’t try to automate everything. Cases above a defined value threshold, cases where the AI scoring didn’t reach the required confidence level, and complex cases that needed detailed legal or financial checks still went to a human reviewer.
We also used AI only where it was safe to do so. It mostly handled document understanding: reading the claim documents and extracting the relevant information. The business rules that decided each claim stayed deterministic, written as clear, repeatable logic. That’s what made the outcomes consistent.
The result was a 70% reduction in costs for that business area, much faster claim processing, more consistent handling (similar claims now get consistent outcomes), and a significant improvement in customer satisfaction. Customers now get their claims resolved quickly and more directly, and receive their payouts sooner.
Notice what wasn’t the goal: maximising the percentage of claims handled by AI.
The goal was to process cases faster and at lower cost while keeping human review where it mattered. Customers don’t really care how many models were involved. They care about receiving a timely, reliable service.
Not every product needs a chatbot
“We should add an AI assistant to our app.”
Maybe. But why?
Imagine an energy supplier. If the goal is to help customers check their latest meter reading or download their bill, a chat assistant would turn a two-click task into a conversation. Good UX would do more, or sending the information to customers where it’s convenient for them, such as a monthly email with the bill attached.
If the goal is to help customers understand a complex tariff or work out which plan suits them best, that’s a different story. An assistant that answers questions about their own usage in plain language might be worth testing.
Will customers complete their tasks faster? Will they get better answers? Will they prefer a conversation to the interface they already have?
Sometimes the answer will be yes. Sometimes a good form, a search field, or a conventional piece of software will do the job better, more cheaply, and with fewer surprises.
And if the answer is yes, don’t underestimate the effort. A convincing demo can take an afternoon. A chatbot you can rely on in production is a serious piece of engineering. Accuracy, latency, cost, privacy, security, and failure handling all matter. They matter even more when the software affects financial decisions or customer entitlements.
Plan for that from the start. Evaluate the idea carefully before you commit, budget for proper testing and monitoring, and decide what happens when the assistant gets something wrong, including how customers can reach a real person in real time when they need to.
Common mistakes
Starting with “we need an AI feature” rather than a customer problem is probably the biggest one. Others include evaluating a prototype on a few hand-picked examples, ignoring running costs, and assuming that a confident answer is a correct answer.
A less obvious mistake is pursuing full automation when selective automation would deliver most of the benefit with much less risk.
Where to start
Identify a customer problem that’s expensive, slow, frustrating, or currently difficult to solve well.
Then work out which part of it actually needs AI. AI is useful when the work involves understanding natural language, reading unstructured documents, making plans, or coordinating several steps. It’s not a reason to replace processes that are already deterministic, fast, and cheap. As in the claims example, let AI handle what rules can’t, and keep the rules where they work.
Check that the problem matters. Then build a small prototype and evaluate it on representative cases, including the awkward ones. Compare it with the existing process and with simpler alternatives.
Decide in advance which errors are acceptable, which require review, and which should stop the process entirely.
What success looks like
Measure customer task completion, processing time, satisfaction, cost to serve, revenue or retention where relevant, and the cost per successful AI-assisted operation.
For automation, also track correctness, escalation rates, and the consequences of mistakes. A high automation percentage is useful only if those automated outcomes are good.
AI doesn’t own the outcome. You do.
Before getting into strategy and ROI, there’s one more point worth making: using AI doesn’t transfer responsibility for the result to AI.
Whether we’re writing a report, reviewing a tender, answering a customer, or shipping software, the final result is still ours.
We’ve all seen AI-generated slop by now. Long documents that say almost nothing. Generic marketing copy. Answers that sound authoritative but are wrong. Code that looks reasonable until someone actually needs to maintain it.
The problem isn’t simply that AI was involved. It’s that someone accepted the output without giving it the attention it deserved, and without really understanding the goal or what success should look like. They focused on shipping something, not on getting it right.
Using AI shouldn’t be an excuse to lower our standards or stop thinking. If we wouldn’t be comfortable attaching our name to a piece of work without AI, why should AI’s involvement make it acceptable?
This doesn’t mean manually checking every word or every line of code. The amount and type of verification should reflect the consequences of a mistake. A meeting agenda and an insurance claim decision don’t need the same review process.
We can use automated tests, evaluation datasets, checks against reliable sources, and human review where it matters. In fact, good AI adoption often involves finding better ways to verify work, not just faster ways to produce it.
AI can take work off our plate. It doesn’t take responsibility off our shoulders.
We should treat it as a helpful assistant and an accelerator, using it where it makes sense. It can help us move faster and do better work, but we still need to decide what is worth doing, what a good result looks like, and whether we’ve actually achieved it.
From AI adoption to an actual strategy
So, four different types of AI adoption, with four different sets of goals:
| Type | What we’re trying to improve | Examples of useful metrics |
|---|---|---|
| Personal AI adoption | Everyday effectiveness, decisions, learning | Time reclaimed, useful opportunities found, tasks completed |
| AI as a work assistant | Professional workflows and team performance | Turnaround time, quality, capacity, cost per task |
| AI to build products | Software development and delivery | Lead time, reliability, rework, time to validate ideas |
| AI in products | Customer experience and business operations | Customer satisfaction, cost to serve, successful outcomes, revenue |
Trying to reduce all of this to a company-wide adoption percentage doesn’t tell us very much.
Instead, we recommend starting with five questions.
1. What problem are we trying to solve?
Forget AI for a moment. What’s taking too long? What’s too expensive? Where do mistakes happen? Which useful things are we not doing because they’re currently impractical?
If we can’t articulate the problem, we’re probably not ready to pick a solution.
2. Is AI the right approach?
Not every problem needs a language model or an agent. Sometimes a conventional software solution is simpler and more reliable. Sometimes the process itself needs changing. Sometimes the existing solution is already good enough.
And sometimes AI is the right choice, just not a large language model. A purpose-built machine learning model, such as a classifier or a regression model, can do some jobs better: flagging likely fraudulent transactions, forecasting demand for stock planning, routing support tickets to the right team, or predicting which customers are likely to cancel. For a narrow, well-defined task, these models are often more accurate and consistent, and they can run at a fraction of the cost and time per request. They usually need more engineering effort up front, including good training data. So, once again, evaluate the options in the context of your company and its goals.
Being able to say “we don’t need AI here”, or exactly what kind of AI you do need, is part of having a strategy.
3. What would success look like?
Establish a baseline before changing anything. How long does the task take today? What does it cost? How often does it fail? What do users think about it?
Then define a meaningful improvement. “Use AI to improve productivity” is vague. “Reduce the average time to process a claim while maintaining our quality standards” gives us something we can actually test.
4. What’s the smallest useful experiment?
We don’t need a company-wide rollout to learn whether an idea is promising.
Pick a bounded use case, agree on how it will be evaluated, and test it under realistic conditions. If the idea doesn’t work, that’s useful information too. Discovering that early can save an organisation a substantial investment.
It’s fine to have a grand vision of where you want to end up. But if you get there in small steps and validate each one, you’re more likely to reach the goal. Or you’ll find out early that the approach you had in mind isn’t the best one, before you’ve invested a lot of money and engineering time in it.
5. Is the return worth the cost?
A simplified ROI calculation is:
The difficulty isn’t the arithmetic. It’s being realistic about the inputs.
Costs can include licences, model usage, infrastructure, engineering time, integration, training, reviews, corrections, monitoring, and maintenance. Benefits might include reduced operating costs, increased capacity, faster delivery, improved retention, or entirely new revenue.
Some benefits, particularly personal learning and decision quality, are difficult to price. That’s not a reason to ignore them, but we should be clear about which outcomes are measured and which are estimates.
And let’s not confuse a projected ROI with a demonstrated one. A nice-looking spreadsheet is not proof that something worked.
More AI isn’t always better
Looking back at the examples, there’s a recurring theme.
For the CEO, we didn’t try to automate decision-making. We helped reduce the administration around it.
For customer support, we didn’t put a chatbot in front of customers. We equipped operators to provide better answers.
For insurance claims, we automated routine checks but kept human review for cases that warranted it.
These weren’t attempts to maximise AI usage. They were attempts to improve how work gets done.
That’s an important distinction. Simply adding AI to a poor process might just help us execute a poor process faster. The more interesting opportunity is to reconsider the process now that some of its activities can be done differently.
So the next time someone says, “We need to increase AI adoption by 30%,” perhaps the best response is:
“Why? What exactly are we trying to improve?”
If there’s a good answer, fantastic. Let’s define how we’ll measure the result and find a sensible way to get there.
If there isn’t, perhaps we should figure that out before buying another thousand AI licences.
The goal isn’t to become a company where everyone uses AI. The goal is to become a company that knows when, where, and how AI can genuinely help.
And, just as importantly, when it can’t.
Want to make AI work for your business?
Throughout 2026, we’ve helped organisations move from vague AI ambitions to practical improvements. We’ve worked on making executives more effective, helping support teams serve customers better, and automating business processes with measurable cost reductions.
At Prelude, we combine technology strategy and architecture advice with hands-on implementation. We can help identify worthwhile opportunities, evaluate the expected return, design the solution, and build it.
Whether you’re exploring AI for the first time or have already invested in it and are wondering where the value is, we’d be happy to talk.
Let’s explore what would actually make a difference for your business.