Nivon: Build AI Customer Agents for Every Interaction

Today I'm going to show you how we cut the support tickets that reach a human by 73%.

In 45 days.

Without hiring anyone.

We did it with an AI agent that answers from our own docs. We ran it in production on two of our products. And today we're releasing it as Nivon.AI.

TL;DR: Nivon lets you build an AI agent trained on your site, docs, and past tickets. It answers what your content covers and hands the rest to your team. There's a free plan and no per-seat pricing.

Here's the full story.

Nivon is a platform for building AI agents for your website. It's made by Pimjo.

You connect your content. The agent answers customer questions from that content and nothing else.

(It can also qualify leads before a chat starts.)

And when a question isn't covered?

The agent doesn't make something up. It hands the conversation to your team with the full thread attached.

If you've shipped a product, you know this one. At Pimjo we've shipped more than a dozen. And they all fed the same support queue.

Here's what was in that queue:

Over and over. Across every product.

The frustrating part?

Every answer was already in our docs. Customers still waited hours for a human to paste a link.

So we tried to fix it.

Attempt #1: More people.

We pulled teammates from other departments into the queue.

It kept growing. More people can close more tickets per day. But they can't change how many tickets come in.

Attempt #2: A chatbot.

It answered with full confidence.

It was also wrong. A lot.

And here's what we learned: a wrong answer is worse than no answer. A customer with no reply waits. A customer with a wrong reply acts on it, breaks something, and opens a second ticket.

Attempt #3: A bigger support plan.

Then we looked at the invoice.

Our bill went up from the year before. Our customer count didn't. The difference was seats. More people at the company needed access, and the tool billed us for every one.

That's when it clicked:

Support volume scales with how much product you ship. Support tools charge for how many people you employ.

Those two numbers have nothing to do with each other.

Scripted bots break the moment a question leaves the script.

Help centers go stale the moment you ship a change.

So what's different now?

Accuracy.

Today's models can answer from a fixed set of documents that you control. They don't have to pull from whatever they picked up in training.

Which means an agent can do two things a scripted bot never could:

That second one is the whole game.

Here's the flow:

Step #1: You connect your content.

Website, docs, FAQs, past tickets. That becomes the agent's knowledge base, and its only source.

Step #2: The agent reads the full thread.

Every message, not just the last one. Your customers never have to repeat themselves.

Step #3: It resolves routine questions on its own.

Start to finish. Nobody on your team touches them.

Step #4: It stops when it doesn't know.

No guessing. The agent says it can't confirm the answer.

Step #5: It hands off with context.

The chat becomes a ticket with the full conversation attached. Whoever picks it up already knows what was asked and what the agent said.

Step #6: It shows you the gaps in your docs.

Every handoff is a question your content didn't cover. Add the answer once and the agent handles it from then on.

In other words: your docs get better every time the agent gets stuck.

We weren't going to ask you to put an agent in front of your customers before we'd done it ourselves.

So we ran Nivon on Aymo AI and Meku.dev for 45 days. Here's what happened:

Now, a quick caveat. (Because you should always ask how a number was measured.)

These metrics measure different things.

Most of the 73% drop came from customers using the chat widget instead of filing a ticket at all. The 80% is the share of those widget chats that the agent closed alone.

Also: these are our numbers. They aren't a benchmark.

Yours will depend on how complete your docs are and what your customers ask.

This is the part I think developers will care about most.

What's not ready yet?

Slack, GitHub, and Telegram integrations are in progress. Notion is in review. You can track all of it on the roadmap.

Here's the problem with per-seat pricing:

Teams start limiting who gets an account to keep the bill down. And the people who get cut are usually the ones closest to the customer.

So Nivon charges by support volume instead.

Each plan also caps agents, workspaces, and training pages. The pricing page has the full limits.

And the free plan?

It never expires. No card required.

Fair question. Here's the short version:

Nivon meets GDPR, CCPA, and ISO 27001 level requirements. SOC 2 Type II is in progress, and we'll publish the report when it's done.

More detail is on the data security page. Nope.

Nivon sits on your website and takes the questions that never needed a human. Everything else lands with your team as a ticket, in the tools they already use.

No migration. No new ticketing system to learn.

If your helpdesk works, keep it. Here's the bottom line:

Repeat questions are a documentation problem. They aren't a staffing problem.

Nivon turns the content you already have into an AI customer support agent. The agent answers from your knowledge base. It hands off to a human when it can't. And every handoff tells you which answer to write next.

That's the approach that cut our human-handled tickets by 73% in 45 days.

Want to test it on your own product?

Running customer support across several products or a bigger team? Talk to us and we'll help you scope it.

Every product has one.

Ours was "Where do I find my API key?"

Tell me yours in the comments, and whether your docs already answer it.

Read more on the Nivon blog.