Where AI belongs in your company's system
AI handles reading, sorting and drafting well. For calculations, fixed rules and decisions that must always come out the same, ordinary code is usually better and cheaper. A simple way to tell one from the other.
Artificial intelligence. October 22, 2026. 7 min read.
These days almost every request for a new system comes with the question "can we add AI?". The answer is almost always yes. The more useful question is where in the process it saves someone time, and where it only adds cost and uncertainty.
A language model is very good at understanding text written any which way and at producing a reasonable first draft. It gets things wrong now and then, and it gets them wrong with confidence. Those two traits help decide where it fits.
Where AI tends to help
Reading whatever arrives in a mess. An order sent over WhatsApp, an email with an attachment, a photo of an invoice, a contract in PDF. Turning that into organized fields (customer, product, quantity, deadline) is tedious work for people and easy work for a model. A person checks in seconds what would take minutes to type.
Sorting. Tagging a message as a question, a complaint or an order. Sending it to the right queue. Flagging what looks urgent. If the AI gets it wrong, the person working that queue notices and fixes it, and the damage is small.
Writing the first draft. A reply to a customer, a summary of a long conversation, a product description. A person reviews it and sends it. The gain is in not starting from scratch.
Answering frequent questions based on what the company has already written. Business hours, return policy, how to get a copy of a bill. With the official answers as its source, the agent handles the repetitive part and hands off to a person when the conversation goes off script.
Where it tends to get in the way
Calculations. Price, tax, commission, interest. A formula in code is right every time and costs almost nothing to run. A model may get it right 99 times and wrong on the hundredth, and with this kind of math a rare mistake is already serious.
Rules that must always come out the same. "Discounts above 10% need approval" is one line of code. Asking AI to decide that trades a predictable rule for an opinion.
Decisions nobody is watching. Approving credit, canceling a contract, messaging a thousand customers at once. When a mistake is expensive and nobody checks first, AI should suggest and a person should decide.
A quick test
For each step of the process, ask two questions:
- Is the input free text or information that's already organized? Free text points to AI. Organized data points to ordinary code.
- If it goes wrong, will someone notice before the customer does? If yes, AI can act on its own. If not, it suggests and a person approves.
In most projects the result is a mix: the system is built on clear rules, and AI shows up at the edges, where the world comes in messy or where someone needs a draft.
What to ask whoever is going to build it
- At which steps does AI come in, and what happens when it gets something wrong?
- How much does each use of the model cost, and what does that add up to per month at your volume?
- Does your customers' data leave the country? To which provider?
- Can you switch models later without rebuilding the system?
If any of the answers is vague, push back. The right AI in the right place saves your team hours. In the wrong place, it becomes a monthly bill nobody can explain.
