Generative AI for Business: What It Does Well, and Where It Doesn't

Generative AI is worth paying for when it removes one piece of manual work where you can name the person doing it today, count the volume per week, and check the output. Miss any of those three and what gets bought is usually a demo that looks good in a meeting and goes quiet once it is installed.

For an owner, the useful question is which bottleneck in your office consists of reading, sorting, drafting or summarising text, and how cheaply a mistake there can be caught before it costs anything. The wider picture of strategy, tooling and payback sits in the guide to automation and AI for businesses.

Why the question has changed

Owners used to ask whether tools like this really work. Most have now tried one on their own phone, and for personal work the answer is settled.

Where it goes inside a business is still open. Model comparisons and release notes do not answer that. The answer lives in your own process: who currently does the dullest part of the job, how long it takes them, and what happens when they get it wrong. No vendor can answer that for you, and whichever model gets used does not change the answer.

The four shapes of work that generative AI handles well

We build AI where manual work is genuinely the bottleneck, and the work usually takes one of four shapes. All four are text in, text out, and all four are done by hand today.

• Reading documents and pulling the contents out. Supplier invoices, delivery notes or registration forms that arrive as PDFs and photos, then get retyped into a system by an admin.
• Sorting and routing incoming requests. Messages on the business WhatsApp number, email, or complaint forms that someone reads one by one before forwarding to the right person.
• Summarising long material. Meeting notes, one customer's full conversation history, or field reports that have to be condensed before anyone decides anything.
• Answering internal questions from your own documents. A new hire asks about warranty procedure or reseller pricing terms, and the answer already sits in a document nobody opens.

One thing is missing from that list: none of it has the AI deciding anything on its own. All four prepare material so that a person decides faster. If you want worked examples at SME scale rather than criteria, several are set out in practical AI use cases for growing SMEs.

Where it still fails, and why that matters more than the wins

Generative AI output sounds reasonable, and the confident tone is identical when the content is wrong, so mistakes never flag themselves.

Two kinds of work are therefore better left alone. The first is work where a wrong answer is expensive and nobody checks it, which covers replies sent straight to a customer unread, or order statuses changed automatically. The second is work where a number has to be exact. Adding up a bill, counting stock, or matching payments belongs in ordinary code that returns the same answer every time.

Forecasting gets filed here by mistake too. Guessing next month's demand or which customers are about to leave sits outside generative AI, and it depends on data quality first. Those conditions are covered in the guide to predictive analytics for decision making.

The measure we use is whether the process has a place for a person to check, or whether the error is cheap and quickly visible. An admin correcting some of the invoices is still far faster than typing all of them from scratch. An email sent in error to your whole customer list cannot be pulled back.

How to tell a real use case from a demo

Three questions are usually enough, and you can answer them yourself without a vendor.

Who does this work today? If there is no name, the work does not exist yet, and what is being planned is a new feature rather than a saving. What is the volume per week? A few documents will never cover the build cost; hundreds change the arithmetic. What happens when the output is wrong? If the answer is "we find out from the customer", the process is not ready to be automated.

A demo fails all three at once: tidy output on one chosen example, with no person, no volume and no failure case. What decides the result in your office is the ugly examples, the invoice photo that is skewed and dark, the customer message that is half abbreviations, and the procedure document circulating in three versions at the same time.

What this costs to find out

Custom software is expensive, and AI does not make it cheaper. We would rather you got an answer on whether the idea holds up before paying for a build.

Discovery does that work first, over one to two weeks, at one price. What comes out is a written plan and a fixed price for the next stage, and that document stays yours whatever you decide. Every stage after it is agreed before it starts, so the exposure in trying us is one stage.

One possible outcome of Discovery is that an off-the-shelf tool sold by the month already finishes the job. When that is the case we will say so and point you to it, and you still keep the plan.

Why conventional engineering still decides the outcome

In a project like this, the model is the smallest part. The time goes into the ordinary software work around it.

Your documents have to be found first, whether they sit in a shared drive, one person's inbox, or paper in a cabinet. Access has to be decided, because an internal assistant that will read out salary figures to anyone who asks clearly cannot exist. Someone needs a place in the process to check the output before it is used, and a way to reject it. The system also needs to record which answers were corrected, so that after a few months you know how often it gets things wrong.

The last question is who maintains it. An AI feature built with exotic tooling is hard for the next developer to pick up, and the problem only shows up a year later. We build with conventional, proven technology and record the reasoning behind each decision next to the code. The code, the servers and every account are in your name from day one. Whoever inherits it inherits the thinking as well as the files.

Starting with one bottleneck

Pick one process that is manual, high in volume, and cheap per mistake. Usually that is document entry or sorting incoming messages, rather than anything touching money directly or a promise made to a customer.

Build something narrow first, for one document type or one request channel. Run it for a few weeks with the person still checking, see what share needs correcting, then decide whether widening it is worth it. If the correction rate does not fall, you stop after one stage and you know why.

Bring one manual process to the first call

No need to prepare technical vocabulary or a big idea. Just describe the dullest job in your office: who does it, how much of it arrives each week, and what happens when something is wrong. We will work out together whether the shape of the work actually fits. The first call is free, there is no deck, and you are talking to the engineers who would do the work.

If an existing tool already handles it, we will say so. If something does need building, Discovery gives you a written plan and a fixed price, and that document is yours whatever you decide afterwards. Use the contact form, which opens WhatsApp, or email hello@arktik.id.