Apptologic

Insights · 1 November 2026

AI in a live game: where it helps and where a script is enough

Many small studios start with the wrong question: which AI product should we buy? A better starting point is to list the work that repeats every week, then sort each task by what it actually needs.

Three kinds of work

Arithmetic belongs in a script. Summing revenue by event, comparing two configs, checking that prices follow a rule: these tasks have one correct answer. A plain script gives that answer every time, costs almost nothing to run and is easy to check. Using a language model here adds cost and a new way to be wrong.

Judgment and writing can go to AI. Summarizing player reviews, drafting an event report from finished numbers, proposing offer ideas from past results: these tasks need reading or writing, and a person can quickly tell whether the draft is good. This is where AI tends to save the most time.

Decisions stay with your team. What to release, which players to target, when to change a price. AI can prepare the options, but someone who owns the result should choose.

Give the model clean inputs

AI output is only as useful as what goes in. Don’t send a model raw event logs and ask what happened. Use code to reduce them to a clean table first, then ask for an explanation of that table.

This has two benefits. It costs less to run, and every number in the answer can be traced to a calculation you can check.

Start with one tool

Pick one task that repeats often and hurts when it is late. An after-event report is often a good first choice, because it has a clear format and you can compare the draft with what your analyst would write. A first version might work like this:

  1. A script pulls the event’s numbers and calculates the results.
  2. A model writes a short draft explaining what changed.
  3. An analyst corrects the draft and signs it off.

Keep track of how much the analyst has to fix. That tells you whether the tool is saving time, and which part to improve next.

Know the risks

Be careful before you send player data to any outside service. Check the provider’s terms, and send only the fields the task needs. Also expect models to sound sure when they are wrong, which is why any figure that drives a decision should come from your own calculations.

Good AI tools make your team faster without hiding how a result was reached. If nobody can explain where a number came from, it should not drive a decision.