How to add an AI assistant to a business system without losing control

Adding a chat window that talks to ChatGPT takes an afternoon. Turning it into something people trust with their tasks, invoices or orders takes more. The model has to answer from your data rather than from general knowledge, it must not change anything on its own, and the monthly bill has to stay predictable.

We built an assistant like that into LevelUp, our own management app. It works with Claude, OpenAI and Gemini, reads data across a dozen modules and proposes 17 kinds of actions. These are the decisions that made it reliable. They apply just as well to a CRM, an order system or an online store.

1. Let the model ask for data instead of guessing#

A model that only sees the question will guess. A model that gets the whole database in every prompt becomes slow and expensive, and it gets confused.

What works is the middle ground. The model receives a short summary and a list of data it can ask for: "tasks this week", "expenses by category", "invoices waiting for payment", or one search across everything. When it needs more, it says so in its answer, the system fetches exactly that and asks again. Answers then come with real numbers, and the prompt stays small.

In LevelUp this runs through a strict JSON answer format with a "need" field, so it works the same on every provider, including those without native tool calling.

2. The model proposes, a person approves#

This is the most important rule: the model never writes to the database itself. It returns proposed actions, such as "create a reminder for 09:00 tomorrow" or "move this task to Monday". The system shows each one with Do, Edit and Dismiss.

Each action is a small class of its own. It checks its input, does one thing and knows how to undo it. That gives you:

• Safety. A misunderstood sentence produces a suggestion you dismiss, not a deleted record.
• Undo. Every executed action can be reversed, so approving is not scary.
• Gradual trust. Harmless action types, such as adding a note, can be allowed to run automatically, while anything touching money or customer data always waits.

For an online store this means AI writes product descriptions as drafts, and a person publishes them. In a business system it means the assistant prepares the invoice and a person sends it.

3. Do not tie the system to one AI provider#

Models change every few months, prices change and quality shifts by task. Build one interface with a driver per provider (Anthropic, OpenAI, Google) and the rest of the system never needs to know which one is answering.

On top of that, use tiers: an economy, standard and power model for each provider. A router picks the tier per request. A quick lookup goes to the cheap model, planning a week goes to the strong one. Switching provider or model is then a configuration change, not a rebuild.

4. Put limits on spending before launch#

Every model call has a price, and a loop or a bug can multiply it quickly. Three things keep costs under control:

• Price every call. Store tokens and cost per call, so you know what each feature costs.
• Budgets and a pause switch. A daily or monthly limit, and one button that turns AI off without breaking the rest of the app.
• An automatic breaker. If costs jump or calls keep failing, background AI work stops by itself and waits for a person.

This is also what makes it possible to quote a client a realistic monthly cost.

5. Keep a log of every call#

When the assistant gives a strange answer, you need to see why: what it was asked, which data it requested, which model answered and what it cost. LevelUp has a separate debug screen with every call. Without it, improving prompts is guesswork. With it, most problems are obvious in a minute.

6. Add memory and initiative only after the basics work#

Long-term memory ("remember that I prefer morning workouts") and proactive messages ("this goal is falling behind") make an assistant feel useful rather than reactive. They also multiply the ways it can annoy people or spend money. Add them last, behind the same approval, limits and log as everything else.

What this means for your project#

If you are considering AI in a store or a business system, start with one task where it saves real time, and build these rules in from the first day:

• answers grounded in your data;
• actions that wait for approval and can be undone;
• a provider-neutral interface with model tiers;
• spend limits and a log of every call.

Everything else can grow from there. We do this as AI integration for Laravel systems, PrestaShop and WooCommerce: from AI product descriptions and translations for a catalogue to an assistant inside your own system.