an agent that knows your business, not business in general.
heysteff is the AI platform for customer support and sales over WhatsApp, Instagram, Messenger, Gmail and Shopify; Steff is the AI agent that runs it. The gap between a generic bot and one that sounds like your brand is not the model: it is what it knows and when it uses it. heysteff's knowledge base is trained on your business and loaded per conversation, not all at once.
Start from a draft that already understands your business
When you create your workspace you describe what you do, and heysteff generates four base files with AI: who the agent is, how it sounds, how it works, and when it has to hand the conversation to a person.
It is not a generic template nor an empty text box: it is a draft written from your own description, editable from day one. You tune it until it sounds like your brand instead of writing it from nothing.
Every file decides when it enters
This is the deep difference from platforms that cram everything into one giant prompt. In heysteff each file picks its moment: the ones defining identity and rules travel every time, and the rest activate only when the conversation calls for them, based on keywords or a pattern you define, evaluated against the latest messages in the thread.
The result is that the agent sees your returns policy when someone asks about returns, not on every greeting. Less noise for the model means sharper answers — and, incidentally, lower cost per conversation, because you are not paying to resend all your documentation on every message.
A cap that prioritises, not one that truncates at random
You can upload as much documentation as you want without the agent drowning in it. There is a limit on how many files enter a single reply, configurable per agent, and when more match than fit, the router prioritises instead of cutting arbitrarily: the base files are never dropped, and of the rest the ones that best match the conversation get in.
Put another way: the agent never loses its identity because it loaded too much documentation.
Escalating to a human is knowledge too
One of the base files defines when the agent should not carry on alone. When the conversation touches one of those topics, heysteff does not just hand the agent context on how to handle it: it also marks the conversation as escalated and pauses the bot, so the following messages land with your team.
It is the necessary flip side of an autonomous agent — knowing where the edge is and stopping there. You can pair it with copilot mode so the AI keeps suggesting while a person drives.
Frequently asked questions
Do I have to write the agent's prompt from scratch?
No. When you create your workspace you describe your business and heysteff generates four base files with AI: who the agent is, how it sounds, how it works and when to escalate to a human. They are editable from day one — the idea is that you adjust a draft that already understands your business, instead of starting at an empty text box.
Is the whole knowledge base sent to the model on every message?
No, and that is the difference. Each file decides when it loads: the ones defining identity and rules travel every time, and the rest activate only when the conversation calls for them, based on keywords or a pattern you define. So the agent sees your returns policy when someone asks about returns, not on every greeting. Less noise, sharper answers and lower cost per conversation.
What if I upload too many files?
There is a cap on how many files enter a single reply, configurable per agent. If more match than fit, the router prioritises: the base files are never dropped, and of the rest the best matches get in. The agent never loses its identity because it loaded too much documentation.
◆ next step
Train Steff on your tone and your rules, and let it load what each conversation needs.
You might also like: copilot mode, API integrations, MCP server, how to train an agent's tone and human escalation rules.