Onpolar.ai — An AI Sales Agent You Set Up From a Website Address

Onpolar.ai is Onpolar's own AI sales agent. A business owner types in their website address — or answers a few questions if they have no website — and the agent learns what they sell, answers every lead within minutes, writes to the people on their list during those people's working hours, follows up and books meetings in their calendar. We designed, built and launched it in three weeks, in English, Arabic, Estonian and Russian.

The problem

Most small B2B companies know they should answer leads faster and follow up more. Most AI sales tools assume a sales-ops team to set them up: sequences, enrichment, CRM fields, deliverability. The brief was an agent a non-technical owner could switch on in four steps, with plain words on every screen, that would never embarrass them in front of a buyer.

What we built

Setup in four steps

Add your website or answer a few questions; the AI learns your business; connect email and calendar; go live. The agent reads the site and writes down what it learned — the offer, the best customers, the prices, the tone — and the owner corrects anything in plain words. A few pasted emails teach it the owner's own writing voice. Connecting a mailbox is one sign-in with Google or Microsoft, or mail server details for anything else.

Inbound: every lead answered within minutes

A website chat, contact forms and incoming email get a personal reply within minutes, at night and at weekends too. The agent asks the questions the owner would ask — team size, need, timing — and books a meeting only with buyers who fit. When a conversation needs a person, it hands over with the full history. The first email to an address someone typed into a form or chat waits for the owner's OK, since anyone can type in someone else's.

Outbound: personal, not templated

The owner imports a list. The agent ranks it by how much each company looks like their best customers, reads each company's website and news, and opens every first email with something real about that company. It writes in 45 languages, sends in the recipient's working hours, skips local public holidays, follows up on a schedule and stops the moment someone replies or says no. LinkedIn notes arrive as drafts the owner sends themselves.

Meetings that arrive with a brief

A booking link in every email puts the meeting straight into Google Calendar or Outlook, with a short brief: who they are, what they need and what was said. A short summary closes each day.

Safe for the owner's reputation

Outreach goes out from the owner's own mailboxes — several if they like — which warm up and take turns at sensible daily volumes, after a domain check scores how likely mail is to land. Every outreach email carries an opt-out, opt-outs are enforced in code, approve-first is on by default, and one click pauses everything.

Guarded against prompt injection

Prospects' websites, inbound emails and chat messages are treated as hostile input. The agent is a workflow of narrow AI steps in which code, not the model, decides what happens next. A model with no tools reads untrusted text into short structured fields, and the model that writes only ever sees those fields. Before anything is sent, deterministic checks verify links, addresses, length, language and the owner's never-mention list, and a red-team evaluation runs the guards against real models.

Credits instead of a subscription

No subscription and no automatic charges: prepaid credit bundles from $49, valid for 12 months, paid through Stripe Checkout with Stripe Tax and EU reverse charge. Every AI action has a published price — an email costs 3 credits, a chat answer 1, a booked meeting 2 — and owners get a warning at 50, 75, 90 and 95 percent used.

Built for the US and the Gulf

Four languages from day one, including right-to-left Arabic, prices in US dollars in every language, and US conventions in the English copy. Every workspace carries a data region, so adding a Gulf region later is configuration rather than a migration.

Security where the data lives

PostgreSQL row-level security separates every workspace. The browser never holds a database key: every query runs on the server, through a data-access layer that checks the session itself. Mailbox and calendar tokens are encrypted with AES-GCM and bound to their own row, so a copied secret opens nowhere else. Work that outlives a request — a follow-up days later, mailbox polling — runs on durable workflows and queues, with a watchdog on every scheduled job.

What's next

Finding new companies and decision makers with verified work emails, logging every conversation to the CRM (ERPFlow, HubSpot, Pipedrive and Salesforce), and WhatsApp and Telegram as channels.

The result

A sales agent a business owner sets up from a website address, without a technical team, priced per action rather than per seat, and live at onpolar.ai. Teams that would rather have the agent set up and run for them get the same approach through Onpolar's AI sales agents.