Start with the problem, not the tool
A lot of companies come to AI from the wrong direction. Someone sees an impressive demo, leadership asks for "an AI project" and the team spends months looking for a place to fit the technology. What usually comes out is a polished prototype that nobody touches after the presentation.
What works is the opposite. Start from a task that already exists, eats up hours of your team’s time and has a clear right way of being done. Then ask whether an AI model can handle part of it well enough. If the answer is yes, the first project can be small, quick to test and cheap to drop if it does not pan out.
This guide is for small and midsize businesses that want to take that first step sensibly: where AI already delivers, how to choose the first use case, how to measure it and what to watch out for.
Where AI already works well today
Current language models are good at reading, writing, summarizing and classifying text. They can also interpret images and scanned documents. That covers a large share of the admin and customer service work in any business. The use cases below are mature, meaning the technology is stable and the risks are well understood:
- Customer service and FAQs: an assistant answers on WhatsApp, your website or email using your own information (opening hours, policies, delivery times, catalog) and hands the conversation to a person when the question goes off script or the customer asks for one.
- Lead triage and qualification: AI reads the message from someone who came through a form or an ad, works out what they want, asks the missing questions and gives sales a summary with a suggested priority.
- Reading and extracting data from documents: invoices, receipts, contracts and forms become structured fields that go straight into your system instead of being typed by hand.
- Summaries: long support threads, meetings, reports and contracts condensed into the format your team needs, with the key points flagged.
- Classifying orders and tickets: every incoming request gets a category, an urgency level and an owner, so the queue starts out organized.
- Automation between tools: AI bridges email, your CRM, spreadsheets and internal systems, interpreting what comes in on one side to fill in the other.
- AI inside the product: the smart feature becomes part of what you sell. In LessCal, an app we built, users take a photo of their plate and AI estimates the calories in the meal.
How to choose your first use case
Coming up with ideas is easy. Picking the first one is the hard part. A good place to begin is asking your team which tasks they would most like to get rid of. The answer often points straight to the candidates, and the people who will use the solution start out wanting it to work.
Sorting the emails that land in your support inbox is a good first use case. Approving credit for customers is not: the rules are full of exceptions and a mistake is expensive. AI may help with that kind of task later, but it is not where you start. Your first use case should tick four boxes at once:
- Repetitive: the team does the same thing many times a day, almost always the same way.
- High volume: there are enough cases for the gain to show and for you to have real examples to test with.
- Clear rules: you could explain to a new hire, in writing, what a correct answer looks like. If your own team cannot agree on what is correct, AI will not get it right either.
- Cheap mistakes: when AI gets something wrong, and it will now and then, the damage is small and easy to fix.
How to measure results before you expand
Before you switch anything on, write down how the task works today: how long each case takes, how many come in per day, how many come back with errors and how long customers wait for an answer. Without that baseline, any result is just an opinion.
Then define success in a way you can actually check. A few metrics that help:
- Accuracy: in a sample reviewed by a person, how many answers or classifications were correct.
- Time per case: how much time the team spends now, including the time spent reviewing what AI did.
- Hand-off rate: in customer service, how many conversations AI resolves on its own, how many need a person and whether the hand-off happens at the right moment.
- User perception: do customers and staff feel things got better, or just different.
- Cost per case: what you spend on the AI model divided by the number of cases handled.
When to expand and when to stop
Run the pilot long enough to capture the normal ups and downs of the business, quiet days and busy ones. If the numbers are good, expand. If they are halfway there, adjust the instructions and the information AI draws on before deciding it does not work. If they are bad, you found out for very little money, which was the whole point of the pilot.
Agree on these thresholds before the pilot starts. When success is defined after the numbers come in, people tend to find a reason to push ahead anyway, and the project grows without having proven anything.
What to watch out for
Used with care, AI is a reliable tool. Used carelessly, it creates new problems. These points belong in any serious project:
- Personal data and privacy laws (GDPR, LGPD and similar): if AI will read customer conversations, documents or records, you need to know where that data goes, how long it is kept and whether the provider can use it to train models. Send the model only what the task requires.
- Human review where mistakes are costly: answers about prices, contract terms, health or legal matters should go through a person before they reach the customer, at least until quality has been proven.
- Hallucinations: language models sometimes make up an answer with complete confidence. The defense is to restrict AI to your company’s information, ask it to cite the source and instruct it to say "I don’t know" and bring in a person when the information is not there.
- Usage cost per call: on top of development, every message or document processed is billed by the model provider. Per call it looks negligible, but it grows with volume and with the length of the text. Estimate the monthly cost before scaling and set spending limits.
- Vendor lock-in: models change versions, prices and behavior often. A well-built solution isolates the part that talks to the model, so switching providers is an adjustment and not a rewrite.
A step-by-step path
An expensive AI project is almost always one that tried to do everything at once. Moving in stages protects your budget: investment grows alongside confidence, and the decision to move forward is based on data from your own operation, not on a vendor’s promises.
The safer path has four stages, and each one starts only after the previous one has shown results:
- Small pilot: a single use case, with real data and a small group of users. Sometimes AI does not even talk to customers at this stage. It suggests a reply and the team decides whether to use it.
- Measure: compare the pilot numbers with your baseline and listen to the people who used it day to day.
- Integrate with your current tools: once results are confirmed, AI runs inside the WhatsApp, CRM, email or system your team already uses. Nobody has to learn a new tool or copy and paste between screens.
- Scale: increase volume, open it up to more teams or apply the same approach to a second, similar process, while keeping an eye on cost and quality.
Mistakes that make the first project more expensive
Some patterns show up again and again in AI projects that cost more than they should have:
- Starting with the most ambitious idea, like an assistant that knows everything about the company, instead of a well-defined task.
- Not organizing the information AI will rely on. If the manual is out of date, the answers will be too.
- Skipping measurement and deciding based on the impression of whoever saw the demo.
- Letting AI answer customers with no clear way to reach a person.
- Tying everything to a single model provider with no room to switch.
How BG Coding can help
At BG Coding, AI projects start with understanding the process: which tasks take up the most time, where the volume is and which decision AI will support. From there we propose a pilot with real data and a success metric agreed before work begins.
Once the pilot proves itself, we integrate the solution with the tools your company already uses and keep track of cost and quality after it goes live. If you have a repetitive task in mind and want to know whether AI can handle it, tell us how it works today.