Where to Start With AI: Find the One Bottleneck
A five-step, process-first method any owner can run in about two weeks, with a notebook and honest numbers, before spending anything on AI.
Key takeaways
- Missing know-how is the most common reason European companies give for not using AI. You don't need technical skills to find where AI should go first.
- Start from the process, not the tool. Look for the step where work waits, where one person is the only one who knows how, or where customers go unanswered.
- Put a euro figure on each candidate bottleneck using your own numbers: hours, missed jobs, late payments.
- Pick one bottleneck that happens often, follows mostly clear rules and leaves a data trail. Measure it before you change anything.
- Fix that one, prove the gain, then move to the next. Five half-finished AI tools are worth less than one that works.
Most owners who ask me where to start with AI expect a list of tools. I ask them where work piles up. The best first AI project almost always sits at the one bottleneck that costs the business the most, and you can find it yourself in about two weeks with a notebook and some honest numbers.
Why "where do I start?" is the real barrier
Eurostat's 2025 figures show 20% of EU companies with 10 or more employees using AI. Among small companies it's 17%. Among large ones, 55%.
When Eurostat asked companies that had considered AI and decided against it why, 70% named a lack of relevant expertise. Unclear legal consequences and data protection followed, each just over half.
"Lack of expertise" sounds like a technical problem. In the conversations I have with owners, it usually means something else: they don't know which of their twenty problems AI should touch first, so they either do nothing or buy the tool from the last webinar they saw.
That second path is how you end up with a chatbot nobody uses. I wrote about one of those in Six Figures on an AI Chatbot and Nobody Used It. The tool worked. It just sat next to the problem.
Think like a chemist: find the rate-limiting step
Before AI, I spent years in analytical chemistry and manufacturing. One idea from the lab has stayed with me: every reaction has a rate-limiting step. Speed up any other step and the output stays the same. Speed up that one and everything moves.
A business works the same way. If quotes take four days because only the owner can price them, a faster website won't bring in more work. If the team can't keep up with incoming claims, a nicer CRM won't change the volume.
AI can do a lot, so a list of what it can do won't help you choose. Ask which step limits everything else in your business. That's the step worth automating first.
The 5-step bottleneck method
You can do all five steps without any AI knowledge. You need a notebook or a spreadsheet, your team's honesty and about two weeks.
Step 1: Write down where the time goes for one week
Ask each person, yourself included, to note what they spend time on, in rough half-hour blocks. Don't build a fancy system. A shared sheet with three columns works: task, time, "could someone else do this?".
You're looking for patterns, not precision. After a week you'll usually see two or three tasks that eat far more time than anyone guessed. Email, quotes, scheduling, data entry and chasing payments show up again and again.
Step 2: Look for the places where work waits
Time spent is only half the picture. The other half is where work sits still. Walk through one real job from first contact to payment and note every point where it waits for someone.
Typical waiting points:
- A customer request that sits in an inbox until the owner has time.
- A quote that only one person can price.
- A report or form someone types up by hand from photos, calls or notes.
- Calls that go unanswered after 5 pm or on weekends.
- An invoice that goes out late because the job details live on paper.
A step where work waits and only one person can unblock it is the strongest candidate for a bottleneck.
Step 3: Map how the process really runs
Most AI projects skip this step, and it's where many of them fail. Put the quality manual aside, sit next to the person who does the work and watch.
You'll find the side spreadsheet, the WhatsApp group, the sticky note with the customer's door code. AI built on the official process breaks on the first real case. AI built on the real process holds up. I wrote more about this gap in AI Works in Demos. Here's Why It Fails in Your Business.
Step 4: Put a euro figure on each bottleneck
Now price your top three candidates with your own numbers. Rough is fine:
- Time: hours per week × what those hours cost you. Count owner hours at what they could earn doing billable work.
- Lost revenue: requests you didn't answer in time, missed calls, quotes that went cold. If you don't know the number, count them for two weeks.
- Cash: invoices that go out late or get paid late.
- Risk: what happens if the one person who knows the process is sick for a month.
The bottleneck with the biggest number usually isn't the one that annoys you most. That's useful to know before you spend money.
Step 5: Score each candidate for AI fit
Not every bottleneck is an AI problem. Some need a new hire, a better form or a conversation. Ask four questions about each candidate:
- Does it happen often? Daily beats monthly.
- Does it mostly follow rules? Sorting, extracting, drafting and checking suit AI. Decisions that need years of judgment don't, at least not without a human in the loop.
- Is there a data trail? Emails, photos, forms, call recordings. AI needs something to read.
- Can you measure it today? If you can't measure the "before", you'll never prove the "after".
The best first project scores well on all four and sits high on your euro list.
What this looked like for a truck damage claims client
A leading truck damage specialist in Germany came to us with a clear pain: the team couldn't keep up with incoming damage reports, and that capped how many new clients the company could take on. Every report was a WhatsApp video call with the driver, about 45 minutes each, and only during office hours.
When we mapped the process, the bottleneck was the step where a driver's photos and details had to be collected, at the hour drivers actually have time and in their own language, Vietnamese included. Everything else in their process worked.
So we built TraxAgent for that step only, and left the rest of their process alone. The same team now handles up to three times the volume, around the clock.
Mistakes I see owners make
Starting with the tool. "We should do something with ChatGPT" is a solution looking for a problem. Start with the bottleneck and the tool follows.
Picking the most visible problem. Customer-facing chatbots are visible. The bottleneck is often in the back office: pricing, paperwork, follow-ups.
Automating a broken process. If a step only exists because of an old workaround, remove it before you automate it.
Running five pilots at once. Each one gets a fraction of attention and none gets finished. One working project teaches you more than five demos.
Skipping the baseline. Without a "before" number, every result becomes an opinion.
When to bring in help
You can do steps 1 to 5 yourself, and I'd encourage every owner to try. You know your business better than any consultant.
Outside help earns its fee in two places. The first is step 3: someone who hasn't done the job for ten years sees the workarounds the team no longer notices. The second is what comes after: turning the chosen bottleneck into something that works on real cases, keeps running and gets measured. That part needs someone who builds and runs software.
If you want to know what that costs, I broke it down in How AI Implementation Projects Are Priced.
Start with the step that limits everything else
AI is useful in a small business when it sits on the step that holds everything else back. Finding that step is mostly honest bookkeeping of your own time and your customers' waiting. Do that first, put a number on it, and the question "where do I start with AI?" mostly answers itself.
Frequently asked questions
Start with the step in your process where work waits the longest or costs the most, not with a tool. List where time goes for a week, follow one job from request to payment, and price your top three bottlenecks. The biggest one that happens often and follows clear rules is your first project.
Not to choose the right project. Finding the bottleneck is about knowing your business: where time goes, where customers wait, what only one person knows. Building and running the AI itself does need technical skills, which you can hire for once you know what to build.
Check four things: it happens often, it mostly follows rules, it leaves a data trail such as emails or photos, and you can measure it today. Sorting, extracting, drafting and checking are good fits. Decisions that need deep judgment usually need a person involved.
About two weeks if you do it yourself: one week of noting where time goes, then a few days of mapping one real job and pricing the candidates. The exact length depends on how often your key processes run.
Then you've saved money. Some bottlenecks need a new hire, a simpler form or a change in who approves what. Knowing that before buying a tool is a good outcome.
Want to see what AI can do for you?
Tell us about your business. We'll get back to you within 24 hours.
Schedule a Strategy Call
