Lucas Giordano, CTO and founder of nottelabs, is also a former AI researcher.
getty
We never really decided to launch our newest product; it started as an internal tool.
My team built it to run our own workflows: answering marketplace reviews, triaging the inbox against what we’re actually shipping, firing off tasks from Slack the way you’d message a coworker. We used it every day for two months. When we decided to launch, we didn’t have much left to build. Our own usage had written the roadmap.
Most companies treat dogfooding as a quality assurance strategy: use the product internally, catch the bugs before customers do. That’s real, but I’ve discovered that it’s actually the least valuable thing it does.
What Dogfooding Actually Buys You
I’ve found that dogfooding buys you three main things (here they are in ascending order of importance):
1. It compresses the feedback loop. A bug reported by a customer takes days to reproduce and reach the person who can fix it. A bug the engineer hits at 11 a.m. in their own workflow takes minutes.
2. It reorders your roadmap by friction rather than by opinion. Feature prioritization in most companies is an argument between the loudest customer, the biggest logo and the person with the strongest view in the room. When your team lives inside the product, the annoying thing gets fixed because it’s annoying.
3. It produces conviction, which is something I underrated for a long time. People can tell the difference between a founder describing a product and a founder describing a tool they use. It comes through on sales calls, in hiring conversations, in the way you write about the thing. You stop reaching for the words marketing gave you and start describing what you did last Tuesday. Conviction is a distribution channel, and it’s one a well-funded competitor cannot buy.
The AI Window, And Why It’s Misleading
Software companies building AI tools are living in an unusual moment. Coding assistants are written using coding assistants. Agent platforms are operated by teams whose entire job is operating agents.
Anthropic’s engineers built Claude Code as an internal tool before it ever became a product. When your ideal customer is a software engineer, and your company is full of software engineers, the dogfooding loop is free.
That’s why AI developer tools iterate at a pace that looks impossible from the outside. The feedback loop has no external dependency in it.
The loop isn’t free everywhere, though. If you build financial software for CFOs, or scheduling software for nurses, your engineers are not your users, and you can’t mandate your way there.
Where It Breaks
Even when you are your own user, dogfooding fails in predictable ways:
• You become an expert user. You never see the onboarding again. You know the workaround, so you stop noticing the thing that requires one. Your documentation stays bad because nobody who works there reads it.
• Your sample size is one. Wanting your own product proves that it’s good. It does not prove that a market exists. The Y Combinator doctrine is make something people want, and it often gets shortened, in founder conversation, to make something you want.
• And then there’s dogfooding theater. You’re technically using the product, but not the one your customers use. Your team’s requests skip the rate limits. An engineer’s bug jumps the support queue because they can just message the person who owns that service. In this case, you have not been using your product; you’ve been using a better one that only you have access to.
Making The Loop Real
The fixes are mostly about removing your own privileges. Run on the same instance your customers do: same version, same limits, same region. File internal bugs through the public support flow instead of messaging the person who owns the service.
Every exemption you grant your own team is a way of opting out of your own product, and the exemptions are always granted for good reasons, which is why nobody notices.
The harder problem is that you’ve become an expert user. You can’t unlearn the workarounds. What you can do is watch someone else hit them. Watch a new hire set it up for the first time, sit there and say nothing. Where they get stuck is where your customers get stuck, and the impulse to lean over and help is the thing to resist.
Then ask the people on your team who aren’t using it why. Someone always isn’t. Engineers have strong opinions and will tell you exactly what’s wrong with it, which is more than any customer will do before they churn.
What It Can’t Tell You
The best software still gets built by people who need it. That hasn’t changed, and the current wave of AI tooling makes the point loudly enough.
Dogfooding is a compass, not a map. It tells you whether you’re pointed at something real. But it can make you blind to every user who isn’t you.
When you spend all day inside your own product, you stop spending time inside anyone else’s, and you find out what shipped six months late. Therefore, pair it with sustained contact with customers who look nothing like your team, and with the work of people building the same thing you are.
And if you find you don’t want to use the thing you’re building, don’t explain that away. Sit with it.
Forbes Business Council is the foremost growth and networking organization for business owners and leaders. Do I qualify?
