“Our own AI” versus a chatbot: what the difference actually is

When a client says they want “their own AI”, they usually mean one of three things. Sorting out which one, early, saves a lot of money.

Level 1: a chatbot in front of a general model

You take a model like GPT or Claude, give it a system prompt that says “you are the assistant for Acme Ltd”, and put a chat window on the website. It costs very little and takes days to set up.

It is also where most disappointing AI projects live. The model knows nothing about your products, your prices or your policies, so it guesses — politely and confidently. It cannot look anything up, cannot take an action, and every answer is a liability if a customer relies on it.

This level is fine for low-stakes use: FAQ-style questions on a marketing site, internal brainstorming, drafting. If that is the job, we will tell you so and keep it cheap.

Level 2: a model that knows your business

The step that makes AI useful is grounding it in your data. Your documentation, product catalogue, past support tickets, contracts and policies are indexed so that, for every question, the relevant passages are retrieved and the model answers from them, citing where the answer came from. (The technique is called retrieval-augmented generation, but the name matters less than the result.)

The practical differences people notice:

• It stops making things up. When the answer is not in your material, it says so instead of inventing one.
• It can be checked. Every answer links to its source, so your team can see why it said what it said.
• It stays current. Update a document and the assistant knows the new version the same day — no retraining.

Most “custom AI” projects we build are this level, and it is the one with the clearest return.

Level 3: an AI that does things

The next step is letting the assistant act: look up an order in your CRM, book an appointment, create a ticket, answer the phone and transfer when it should. This is where “assistant” becomes “agent”.

It is also where design discipline matters most. Every action needs a permission model (what can it do alone, what needs a human to approve), logging you can audit, and a fallback when it is unsure. Done carefully, it takes real work off your team. Done carelessly, it sends the wrong email to the wrong customer at scale.

Fine-tuning: usually not where to start

People often assume “our own AI” means training a model from scratch or fine-tuning one on their data. For most businesses that is the last step, not the first. Fine-tuning changes how a model writes — its tone, its format, its handling of a specialised vocabulary. It is poor at teaching the model facts, which go stale. Grounding (Level 2) handles facts; fine-tuning is worth it later, when you have thousands of real examples of the output you want.

Where the data lives

One thing we are firm about: your data and your AI’s configuration should live in your cloud account or on your servers, under your control, not inside a vendor’s subscription. If you stop working with us, everything keeps running and you own all of it. That is true of every project we build, AI included.

A quick way to decide

• Questions only, low stakes, public information → Level 1 is enough.
• Questions about your products, policies or customers → Level 2.
• You want it to take actions, or answer calls → Level 3, built on Level 2.
• You have a large set of examples of exactly the output you want → consider fine-tuning on top.

If you are not sure which bucket you are in, a short call usually settles it.