We use Avery NXR to build Avery NXR

Most software companies have one product. They use their product internally. The Slack team uses Slack. Notion's company runs on Notion. It's a common pattern.

We dogfood Avery NXR more aggressively than most companies dogfood their own products. By "more aggressively" I mean: the company essentially runs on Avery NXR agents, not on traditional SaaS subscriptions.

Here's what that actually looks like, and what we've learned by doing it.

The Avery Software stack

Our company runs on:

For agents: Avery NXR (obviously). We're our own customer.

For non-agent tooling: → GitHub for code → Linear for issues → Notion for docs → Slack for communication → Gmail / Google Workspace → Stripe for billing → Vercel for deployment

That's mostly it. Notable absences: → No standalone meeting transcription tool (Sophia agent handles it) → No standalone competitor monitoring tool (Yuki agent handles it) → No standalone CRM AI features (custom agents handle it) → No standalone customer support AI (Priya agent handles it) → No standalone content drafting tool (custom agents draft, humans edit) → No standalone resume screening tool (Marcus agent handles it) → No standalone analytics AI features (custom agents pull from databases)

These are categories where most companies our size pay for AI-flavored SaaS subscriptions. We don't, because we run agents that do the same work.

What we actually save

Estimated AI-flavored SaaS subscriptions we'd be paying without Avery NXR agents:

→ Meeting transcription with AI follow-ups (Otter Business or equivalent): ~$2,400/year → Competitor monitoring (Crayon or equivalent): ~$1,200/year → Sales pipeline AI tool: ~$3,000-6,000/year → Customer support AI (Intercom Fin or equivalent): ~$3,600-7,200/year → Content drafting tool (Jasper or equivalent): ~$1,200-2,400/year → Resume screening AI (ATS add-on): ~$1,200/year → Analytics AI features (various): ~$2,400/year

Total avoided: ~$15,000-22,800/year.

Plus we save on what would have been our cloud-LLM bill if we'd built these agents on a competing platform. Estimated: ~$30,000-50,000/year given our agent volume.

Combined annual savings from dogfooding our own product: ~$45,000-75,000/year for a small team.

But the dollar savings is the smallest interesting part of dogfooding.

What we learn by being our own customer

Bugs surface faster. We hit bugs in production scenarios that our test suite couldn't simulate. Every bug we hit in our own use is a bug fixed before customers hit it.

Features get prioritized correctly. When we work in our product, we feel the friction points. Features that solve real-feeling pain get prioritized over features that sound impressive in product reviews.

We can't ship things we don't use. If we ship a feature and don't use it ourselves within 2 weeks, we know we've lost the plot. The "do we actually use this" check is one of the best product filters we have.

We see what's MISSING. Our daily use reveals workflows we wish we had agents for. Those become customer-facing feature requests.

We become expert users. When customers ask us "how do I configure X," we know because we've configured it ourselves. Support is faster + more useful.

What we use Avery NXR for internally

The full list of agents running inside Avery Software at any given time, roughly:

→ Anna (daily news digest) — 15 sources, AI industry focus → Sophia (meeting follow-ups) — runs after every sales call, every customer call, every internal meeting → Marcus (resume screening) — active when hiring → Priya (customer support triage) — runs on incoming support emails → Carlos (sales pipeline digest) — daily morning summary → Yuki (competitor monitoring) — weekly competitive intelligence → Liam (server health) — monitors our infrastructure → Content idea generator (custom) — daily suggestions for blog posts and threads (this thread came from one) → Newsletter drafter (custom) — drafts our weekly customer newsletter → Inbound lead qualifier (custom) — processes founder@ inbox → Documentation reviewer (custom) — flags doc updates needed based on product changes → Release notes drafter (custom) — drafts release notes from merged PRs → Customer health digest (custom) — flags accounts showing signs of churn risk

That's 13 agents. We add 1-2 per month as use cases emerge.

What dogfooding has changed about our product

Specific product changes that came directly from being our own customer:

Audit ledger improvements. When we needed to investigate something an agent did and the ledger wasn't queryable enough, we improved the ledger.

Template library. When we built a 4th custom agent that felt similar to ones we'd built before, we noticed it was template-worthy. Some of our pre-loaded templates came from internal use first.

Connector additions. When we needed to connect to a service that wasn't in our connector library, we added it. Many of our 63 connectors started as "we need this for our internal use" before becoming customer-facing.

Cost projection calculator. When we explained to a customer why their cloud-LLM bill was projecting to grow faster than they expected, we built the math into a calculator. That calculator is now on our website.

Onboarding flow. When new Avery Software employees got onboarded, we noticed which parts of agent configuration were confusing for them. We rebuilt those parts of the onboarding flow.

Multi-user features. When our team grew past 3-4 people and we needed agents to be visible/shareable across users, we built the multi-user features in Pro tier.

These are all features that exist because we're our own customer. Without dogfooding, they would have been built later, worse, or not at all.

The discipline dogfooding enforces

Dogfooding creates a discipline that we think is worth describing:

We can't ship something we don't trust ourselves. If we wouldn't deploy a feature in our own production environment, we don't ship it to customers.

We can't market a workflow we don't run. Marketing claims about Avery NXR are constrained by what we actually use. We don't market features that sound good but we haven't proven internally.

We can't recommend an architectural pattern we don't follow. When we say "local-first is the right architecture for operational AI," we follow that ourselves. No central cloud version. Local on our laptops + private deployment.

We can't shortcut quality. If a feature would degrade our own experience, we don't ship it. The "would this break my day?" check is honest because it would actually break our day.

What we'd recommend to other software companies

If you're building software for businesses, dogfood it. Not casually. Aggressively.

If your product is supposed to absorb a workflow, absorb that workflow inside your company. If you can't run your own product on your product, that's a signal worth listening to.

If you ARE running your product on your product and learning from it, you have a competitive advantage other companies don't have access to — the ability to see your product through customer eyes every day.

Most companies dogfood occasionally. Few dogfood aggressively. The aggressive version is harder. It's also where the compounding learning happens.

The bigger principle

Avery NXR exists because we wanted local-first AI agents for our own work and couldn't find a tool that did them right. We built it for ourselves first.

The fact that our own company runs on the product is the most honest signal we can give about whether we believe in it. We do. So we use it. Daily. For everything we can.

If you want to know what Avery NXR is — look at how Avery Software operates. The agents we run are the same ones available to you. The architectural choices we made are the same architecture you'd inherit. The cost savings we get are the cost savings you'd get.

That's not a marketing claim. It's just how the company works.

→ avery.software — Free Desktop tier. The agent platform built by people who use it every day for their own work.