A chat bot GPT setup for your business is not the same thing as the public chat assistant your staff use to draft emails. The language model underneath may be similar, but a business bot has to answer from your facts, stay inside your rules and do something useful with the conversation, such as booking a call or updating a contact. This article answers one question: how do you turn a general GPT-style model into a bot that speaks for your firm without inventing things?
Why a raw GPT model is the wrong starting point
A general model knows a great deal about the world and nothing about your firm. Ask it your fees and it will guess. Ask it whether you cover a region and it will say yes to be helpful. That behaviour is fine for drafting and dangerous on your website.
Three things make the difference. Grounding, which means the model answers from your own knowledge base rather than its training. Instructions, which set tone, scope and what it must refuse. Actions, which let it book, tag or hand off instead of only talking. The levelpartner.ai platform gives an AI agent all three, so you configure behaviour rather than write code.
Before you start
- A short document describing your services, who they are for and who they are not for.
- Your pricing approach, even if the answer is "we quote after a call".
- A list of topics the bot must never discuss, such as specific financial advice.
- The calendar and pipeline the bot should feed.
- Someone senior who will read the first week of transcripts.
How to set it up
- Go to AI Agents > Conversation AI and click + Create Bot.
- Select Prompt Based Bot.
- Choose General Q&A, or Appointment Booking for first calls.
- Continue into the bot editor.
- Configure Bot Settings, including the firm's name.
- Configure Bot Training with your services document and FAQs.
- Configure Bot Goals using prompts, model settings and actions.
In the prompt, set the role, tone and reply length, list the forbidden topics with what to say instead, and tell the agent to answer only from its training and to admit when it does not know. In Bot Goals, enable Appointment Booking with the first-call calendar and use Update Contact Field so names, emails and answers reach the CRM. Start in Suggestive mode so your team approves each reply for a few days, correct the knowledge base from what you see, then move to Auto-Pilot.
Writing instructions that hold
Instructions are where most business bots go wrong. Long, polite paragraphs about being helpful achieve little. Short, specific rules work. "Reply in no more than three sentences." "Never quote a price; say the team quotes after a twenty-minute call and offer a booking." "If someone mentions a complaint, apologise once and hand off to a person."
Write the rules as the people who answer your phones would explain them to a new starter. If a rule needs a paragraph of explanation, the bot will probably misapply it. Split it into two simpler rules.
Keep a running list of every bad reply you find and the rule or knowledge entry that fixed it. After a month that list becomes the best training document your firm has for human staff too.
What it is for in a 10 to 250 person firm
The value is consistency at volume. A firm with flat revenue usually has enough enquiries; it loses them through slow and uneven first replies. A grounded bot gives the same accurate answer at 8am and 11pm, asks the same qualifying questions every time and puts the result in the pipeline. Your senior people stop answering the same five questions and spend time on the conversations that need judgement.
Expect to spend about half an hour a week reviewing transcripts once it is stable, and a couple of hours in the first fortnight.
Suppose a visitor asks "do you work with clients in Scotland". A raw model might say yes and describe services the agency does not offer. Grounded on the agency's own services document, the agent would say the agency works with UK clients remotely, name the two relevant services and offer a call. Now imagine scoring both on a test set of 40 questions. If the grounded agent gives 37 usable answers and the raw model 22, that gap of 15 is the whole case for grounding, and the same test on your own questions shows where you stand.
People assume a cleverer model fixes wrong answers. It rarely does. Wrong answers almost always come from a gap in the knowledge base or a vague instruction. Fix the source before you look at the model.
When not to use a GPT-style bot
Do not use one where a wrong answer has legal or financial consequences and you cannot fence the topic off. Regulated advice, contract interpretation and anything involving a vulnerable customer should go to a person. Mortgage brokers in particular should have the bot gather details and book, never comment on suitability.
Troubleshooting
The bot invents services or prices. It is not restricted to the knowledge base. Add the instruction to answer only from your content and to say so when it lacks an answer.
Replies sound robotic. The instructions describe tone in adjectives. Paste two real replies from your best salesperson as examples of the style you want.
It never offers a booking. The prompt never says when to offer one. Add a rule naming the moment, for example after the visitor confirms budget and timescale.
It hands off too often. The hand-off rule is too broad. Narrow it to complaints, explicit requests for a person and questions outside the knowledge base.
Frequently asked questions
Which model does the platform use?
The platform manages the language model for you, so you do not choose or host one. What you control is the knowledge, the instructions and the actions, which is where almost all of the quality comes from. Usage is billed as used.
Can the bot learn from past conversations by itself?
Not in a way you should rely on. Improvement comes from you reading transcripts and adding better answers to the knowledge base. That keeps you in control of what it says. The knowledge based agent article explains why that design is deliberate.
Is this the same as a conversational AI agent?
It is one form of it. A GPT-style chat bot is the text channel; the same agent can also run on SMS and WhatsApp. Read conversational artificial intelligence for the wider picture and what an AI agent can do for the full list of actions.
Where should I put the bot first?
Start with web chat on your highest-intent pages. The chat bot for website guide walks through the full widget installation.