I had to end my $5M ARR business

I own Groove, a 12 year old help desk doing $5M arr.

18 months ago we tried to make it truly agentic. We couldn't do it. So I had to end my $5M arr business.

Everyone in b2b saas is currently sprinting past this realization, so I want to write down what it actually cost us to learn it.

We had every advantage

I know the legacy help desk category from the inside out. I own one. Customers actually like it. We had 12 years of tickets, a team that had been shipping against that codebase for a decade, and profitable revenue funding the entire thing.

If anyone was set up to bolt AI onto an existing help desk and have it work, it was us.

So we did what every incumbent is doing right now. We started retrofitting.

The first six months felt good. Suggested replies, summarization, macros that write themselves, a bot on the front of the queue. It demoed well. Customers liked the demos. Internally we were telling ourselves we were 70% of the way there.

We were not 70% of the way there. We were 70% of the way through the part that's easy.

The three problems were all structural

Once you get past the surface features and try to hand the AI actual responsibility, you hit the same wall from three directions.

The data model is wrong. A ticket in a legacy help desk is a thread of human messages with a status on it. That's the whole object. An agentic system doesn't need a thread, it needs the state of the customer: what they're on, what they've paid, what broke last month, what they asked support 90 days ago, what's happening in the product right now. We didn't have that. We had threads. You can join your way toward it for a while and then you're just building a second product inside the first one.

The workflows are wrong. Every workflow in a 2007 architecture has a person at the center of it clicking a button. Assign. Escalate. Merge. Snooze. Apply macro. Those aren't features you turn off, they're the load-bearing walls. When the AI is the one doing the work, half the system is a set of controls for a job nobody's doing anymore.

The pricing is wrong. We charged per seat. So does everyone else. If the AI works, the customer needs fewer people, which means we sell fewer seats, which means we get paid less for doing more. The industry's answer to this has been to keep the seats and add a per resolution fee on top, so you now pay twice. That's not a pricing model, that's a hedge against your own product.

None of the three are fixable with a feature. I kept looking for the version where they were.

AI on legacy architecture is still a chatbot, just a more expensive one.

The month we stopped

18 months later, we stopped.

That's a clean sentence to write now. At the time it was me sitting with the fact that I'd spent a year and a half and a meaningful amount of my own money proving that the obvious path didn't work, and then having to say that out loud to a team that had been shipping hard the entire time.

Most people I told thought I was crazy. A few said it more politely than that. The consensus was that I was chasing a shiny object and abandoning a real business with real customers for something that didn't exist yet.

Fair read, honestly. Groove was flat at $5M arr and had been for 3 years, but flat at $5M is still $5M.

I want to be clear that I'm still proud of Groove and it's still a good help desk. But a good help desk and an agentic system are different animals, and no amount of engineering effort turns one into the other. I've got 18 months of scar tissue on that, which is more than most vendors selling you "AI powered" anything can say.

What we did instead

We started over from zero and created a completely new co. around it: Helply.

Before writing a line of the new product I ran 50+ requirements calls with support teams. Not sales calls. Just asking people what their day actually looks like and where the current tools make them do stupid work. Three things came up so often they became the product.

Then 12 months building what is essentially an ai native zendesk. Agentic from day one, not retrofitted. And pricing that matches: $0 per seat, $0.50 per resolved ticket. We only bill for tickets where something actually got done. Spam, notifications and no-touch tickets don't get charged, which sounds like a small thing and is the reason the pricing page survives being read closely.

Groove funds all of it. 85% of R&D, 100% of marketing. We'll run both forever. Groove is in managed decline on purpose, which is a strange thing to say about a company you built, and a much better outcome than pretending it's about to grow again.

8 months since launching, Helply is at $1M arr growing 7% a month. Most of it comes from people leaving the exact platforms still selling the retrofit.

I think we've validated it but you never know.

The part I still don't know

The market is stacked with well funded players who are further along on distribution than we are. Most days I think we can compete and that it isn't a zero sum game. Other days I'm sitting there going how the fuck are we actually going to compete with these goliaths.

Different goals though when you own 88.02% of the thing. $10M arr and $100M arr are not the same target and the smaller one goes a long way.

Buyers want to pay for resolutions now. A per seat ticketing system architecturally cannot sell that. Most of my own category hasn't noticed yet, or has noticed and is hoping the invoice sorts it out.

So if you're bolting AI onto a 2007 architecture right now, in any category, I'd genuinely love to be wrong. I just spent a year and a half proving to myself that I'm not.

100% confident the product competes with the goliaths. Distribution is what keeps me up at night.

Lets see.