AI or conventional software?

Most people meet it from the outside. A supplier wants to sell you an AI solution, the board asks what you are doing about AI, or a competitor talks about its new assistant. Then you are sitting with a concrete task (quotes, case handling, documents, customer enquiries) and pressure to choose the technology before anyone has looked properly at the task.

Our experience from 250+ completed projects is simple: a large share of the tasks sold as AI projects today are solved better, more cheaply and more predictably with conventional software. A smaller share really can only be solved with AI. The tool turns the order around and looks at the task first. The questions are not our invention: they build on the criteria that recur in international guidance on choosing between rules and AI, held up against our own projects.

Conventional software is strongest when the rules create the result: the same input gives the same answer, every time, and you can explain why. AI-assisted software lets AI help with reading, finding or suggesting, while people and fixed rules steer the process and carry out the action itself. AI-driven software puts the interpretation or assessment made by AI at the very core of the solution, so its answers drive the action. The difference decides how much control and testing the solution needs before it is given responsibility.

The dividing line is not how modern the technology is, but what the task looks like. Answers from AI are statistical, not guaranteed. Three questions therefore weigh heavily: what does a mistake cost, does the answer have to be explainable (to customers or auditors, for instance), and can you prove that the solution works on your own cases? Many solutions end up as the combination: conventional software as the backbone, and something cleverer in exactly the step where a person used to have to judge. At Leki Aviation the value sits precisely there: fixed rules for what can be put into a formula, and smarter search across a large spare-parts catalogue for everything that cannot.

No. Conventional software still runs by far the most of what companies make their money on: finance systems, stock control, integrations, portals. It is predictable, testable and cheap to run. AI is an extra tool for tasks with judgement and unstructured material, not a replacement for all of it.

For the writing an individual employee does, often yes. But if an answer is to land automatically in your finance system, go out to a customer or trigger an invoice, it needs steering, control and integration around it. That is the difference between a good personal tool and a solution the business can lean on.

Less than most people think, and more than the sales people say. Modern AI solutions are rarely trained on your own cases, but every AI solution has to be testable on representative material. Around 50 cases where you know the right answer, plus a specialist to judge the answers, is enough for a real test. If you do not have them, start by gathering them. A shared folder is enough to get going.

Neither of us, from a distance. But ask the supplier three things: what happens when the solution gets the answer wrong, how is it tested before it goes into operation, and why could the task not be solved with rules? If they have no concrete answer, the recommendation is probably driven by what they sell, not by your task.

Perhaps, and then it is the cheapest choice. AI is arriving as features in the systems many companies already have: the finance system, the office suite, the customer system. Start by trying them out on the task. If they do not reach far enough, you at least know exactly what a bespoke solution has to do beyond them.