Why most business chatbots disappoint
It’s Friday night and a customer asks whether shipping is included. The bot a contractor set up last spring says yes with full confidence. Shipping stopped being free two months ago. On Monday your team apologizes, someone turns the bot off, and a quarter later the same story plays out with a different tool.
The problem isn’t the AI. It’s that the bot lives outside the business. It doesn’t know the current policy, can’t see the customer record, can’t be limited to what it should touch, and nobody but the contractor understands how it’s configured. AGENTUS fixes that by making each AI employee part of the system instead of an add-on.
How AGENTUS fits into PRANUS
An AI employee in PRANUS is treated like any other team member. It has a seat on the org chart in TEAMUS, permissions that come from that seat, and a clear area of responsibility.
- Customer conversations. Support and sales agents reply inside the CONNECTUS shared inbox: Telegram, Instagram, your website chat and email. Start in suggest mode, where an operator approves every reply.
- Company knowledge. Answers come from the MEMUS knowledge base, with a link to the source document. Permissions are applied before the search, so an agent never sees content it isn’t allowed to.
- Questions for people. When there’s no answer, the question and a summary of the conversation go to the Questions board in CHATUS. A specialist replies, the customer gets the answer, and once approved it becomes a ready answer for next time.
- Tasks and deals. Agents create tasks in TASKUS and update deals in CRMUS, limited to the tools you’ve enabled.
What the AI does, and where it stops
It answers in four tiers. Company profile first (address, hours, billing details), then approved answers, then a MEMUS search, and only then a question to your team. Every approved answer means the agent handles more on its own, at lower cost.
It remembers the right things. Knowledge is stored with its source, a confidence level and a review date. A retired policy isn’t deleted; it gets an end date, so the agent answers by the rule in force today. You can wipe working context or learned facts with a button or on a schedule.
It stays in its lane. Tools are enabled one by one. Writes to other systems follow a fixed path: propose, get a human yes, execute once. Keys are stored separately and the model never sees them. New agents are tested in a sandbox against real past conversations, not live customers.
It spends predictably. Each agent has a daily and monthly budget. When it hits the cap, it stops, drops to a lighter model, or brings in a person, depending on the rule you set.
The AI hub: one gateway to every model
Underneath every agent and service is a single AI hub. Services don’t ask for a specific model; they ask for a kind of job, like “complex reasoning” or “speech to text.” The hub decides which model handles it, with which key, and who pays.
- Fallback chains. If a provider fails, the request quietly moves to the next model in line.
- Limits at every level: person, service, project, team, organization. When a request is refused, the hub tells you whose cap was reached.
- Cost tracking at real prices, plus the option to bring your own API key.
- Local models. The hub works with LM Studio, Ollama, llama.cpp, vLLM and MLX, and supports a local-models-only mode for organizations that don’t want data leaving the building. Hosting choices are covered on the cloud, self-hosted and private deployment page.
How you roll it out
Your first agents are set up by the onboarding agent. It connects your channels, builds a knowledge base from your documents and conversation history, checks it with test questions, and switches on support and sales agents in operator-approval mode. From there you adjust behavior yourself through the agent profile and the instruction wizard. The full rollout is described on the implementation page.
Who it’s for
- Owners and executives who want routine questions off their team’s plate without handing customers to a black box.
- System administrators who need every AI employee, its permissions, spend and change history in one place.
- Companies with strict data rules that want models running on their own servers.
See how an agent answers from your own knowledge base, and when it calls a human, in a PRANUS demo.