Welcome back!

This week we break down why the AI products growing fastest treat customer conversations as a permanent operating function, walk through how two Docker Desktop veterans made open models easy enough that 8.9 million devs now run them, explore how Decagon tripled its valuation to $4.5B in under six months on customer obsession, and show how a free background remover became a $94M ARR profitable business.

Have suggestions? Send them to us at [email protected].

Founder’s Intel

Talking to users is not an early-stage activity

Data Intel

• Companies maintaining weekly customer touchpoints after product-market fit grow 2.4x faster than those scaling outreach back after early traction.

• Founders interviewing at least 5 customers per week ship features with 60% higher adoption rates than those relying on analytics alone.

• Top-quartile SaaS companies dedicate 15-20% of leadership time to direct customer conversations, even at $50M+ ARR.

Why It Matters

Most founders treat customer discovery as a phase.

You do a bunch of interviews before launch, find product-market fit, then move on to scaling.

The fastest-growing AI companies have flipped that script.

They treat customer conversations as a permanent operating function, not a milestone you graduate from. Every week, every quarter, every stage.

The signal to watch for isn’t whether your team is talking to users. It’s whether your team has stopped.

This matters more in AI than in traditional software because of speed of change.

The underlying models improve every few weeks. Customer expectations shift faster than roadmaps. Competitive moats compress in months instead of years.

Founders who rely on six-month-old customer feedback get blindsided by problems their users started experiencing four months ago.

The teams growing fastest run weekly touchpoints with every customer and keep the whole team involved in those conversations, not only sales.

The principle is that customers rarely tell you what’s missing on their own. You have to pull it out of them with direct, specific questions about their workflow.

This applies whether you’re selling $50/month to SMBs or six-figure contracts to enterprises.

The medium changes, not the discipline.

At the low end, that means user calls on Zoom and reading every support ticket. At the high end, it means sitting inside the customer’s workflow in person. Both compound the same way.

The pattern holds across the AI tools breaking out right now.

The ones that compound aren’t the ones with the cleanest landing pages. They’re the ones whose founders reply to every email, watch every session recording, and ship features pulled directly from yesterday’s user call.

Quick Tip

Block 5 hours a week for direct customer conversations. Not analytics review. Not Slack threads. Actual conversations with paying users.

Document the top 3 friction points you hear each week.

After 4 weeks, you’ll have 12 data points your competitors don’t, and most of them won’t be in any dashboard.

Behind the Tool

The Local-Model Layer 8.9M Devs Run On

The Spark

Jeffrey Morgan and Michael Chiang had already made one intimidating piece of infrastructure easy to use.

Morgan built Kitematic, a visual way to run Docker containers.

Docker acquired it in 2015, and the work became Docker Desktop, now used by more than 10 million devs.

Years later they saw the same usability gap forming around open models.

The capability was there: capable open-weight models were shipping regularly. But running one still meant wrestling with drivers, dependencies, and hardware detection.

The models had been built for researchers, not the average dev.

Getting one running on your own machine was genuinely hard.

The Build

They reduced local inference to a single command.

Install the app, type one line, and an open model is running on your own hardware.

What mattered more than the interface was the product doctrine underneath it.

The dev should own the model choice, the machine, and the data, while Ollama absorbs the packaging and compatibility work.

The founders sum it up in five words: your model, your machine, your data.

That was a bet against the direction the rest of the market was moving.

While most of the industry pushed devs toward closed, metered APIs, Ollama treated ownership and privacy as the point, not a compromise.

The Breakthrough

The inflection came in early 2026, when open models got good enough to handle real agentic work like coding.

Suddenly the reason to run them locally wasn’t ideology or cost alone. It was that they could now do the job.

Usage roughly doubled month over month.

Ollama now reports 8.9 million devs, used inside 85% of the Fortune 500 (including regulated industries like government, healthcare, and finance), and more than 67,000 community-built integrations.

All of it built and run by a team of 14 people.

On July 9 the company raised a $65 million Series B led by Theory Ventures, bringing total funding to $88 million, on top of an earlier round led by Benchmark’s Peter Fenton.

The Next Chapter

Ollama is building one interface across local and cloud inference.

If your machine can run the model, it runs locally.

If not, or if you want a heavyweight model that exceeds your hardware, the same commands route to Ollama's cloud with no configuration change, billed by GPU time rather than per token.

The goal is to add new open models on release day and make hybrid workloads possible without giving up model portability or data control.

Key Lessons

• Make a hard capability feel like an ordinary tool. Morgan and Chiang didn’t invent local inference. They removed the operational pain around it, the same move they made at Docker. The distribution came from erasing setup friction, not from novel technology. Find the capability people want but can’t easily use, and become the layer that makes it one command.

• Turn a strategic preference into a product. Plenty of devs wanted an escape from single-vendor API lock-in. Ollama didn’t argue the case in blog posts. It made owning your stack the path of least resistance, so the preference became a default. If your customers already believe something, build the product that lets them act on it without effort.

• Wait for the capability threshold, then be already there. Ollama existed for years before open models could do agentic work. When the threshold broke in early 2026, the easiest on-ramp already had millions of users installed. You can’t force the moment the market turns, but you can be the obvious tool sitting there when it does.

Tool of the Week

Decagon's Customer Obsession Playbook

What’s Decagon?

An AI customer experience platform building autonomous agents for enterprise support.

Tripled its valuation to $4.5 billion in January 2026 (up from $1.5B six months earlier), on the back of adding more than 100 enterprise customers in a single year.

What Worked

• Treated customer discovery as a permanent operating function. Most companies do customer interviews aggressively pre-PMF and then slow down. Decagon does the opposite. Every customer gets a weekly touchpoint. The entire team (not only sales) stays involved in customer conversations. Co-founder Jesse Zhang has said the principle is that customers won’t proactively tell you what’s missing. You have to pull it out of them.

• Built outcome-based pricing into the core model. Decagon charges per resolution, not per seat. Customers only pay when the AI resolves a support ticket. This removed buying friction for enterprises with finite procurement budgets and let them deploy aggressively without long approval cycles. Reported deflection rates reach as high as 80%.

• Made the buying decision easier with Agent Operating Procedures. Most AI agent platforms require devs to write code or use SDKs. Decagon built a layer where customer success teams describe agent behavior in plain English, and the system compiles it into structured logic. The CX team builds the agent. The engineering team retains control of the underlying integrations. Both sides win.

• Land small, expand fast. Decagon doesn’t try to replace an entire support org on day one. Customers deploy to 10% of their support volume, prove the ROI, then ramp. Buying cycles got shorter because they didn’t have to wait for perfect test cases before signing commercial deals.

Key Lesson

The fastest-growing AI companies don't slow down customer conversations after PMF. They institutionalize them.

Weekly touchpoints with every customer. Outcome-based pricing that removes risk. Deployment models that let customers prove value before committing.

You don't need 300 enterprise customers to apply this. The mechanism scales down.

Pick your 10 best customers, build a weekly conversation cadence with each, and ship one feature per week that came directly from those calls.

That loop compounds faster than any roadmap planning meeting.

Fresh Out of the Lab

Juggler

What Is It?

An open-source visual workbench for AI coding agents, launched in June 2026 by Julian Storer (the dev behind JUCE and the Tracktion DAW).

Instead of burying your agent session in a long chat transcript, Juggler turns it into an editable tree you navigate in Finder-style columns.

Runs locally on your own machine, remotely, or both.

What’s New

Juggler treats a coding session as a doc, not a log file.

Every tool call, approval, context item, and sub-thread is visible and inspectable, and any point can branch into its own sub-thread you can compare, backtrack, or edit.

Context items, slash commands, and the agent’s reasoning loops are all JavaScript plugins you can fork or replace, so the orchestration layer stays fully editable rather than locked inside a vendor’s app.

It runs against your existing model subscriptions or local models through Ollama, so you keep control of which model touches your code and where your data goes.

A desktop client pairs with a headless local server that multiple clients can attach to.

Why It Matters

Most coding agents force you to trust a scrolling transcript where the important decisions are buried in collapsible chat.

For a founder shipping real code with an agent, that opacity is a risk.

You can’t easily see what the model changed or why.

Juggler makes the whole session legible and branchable, so you can inspect exactly what the agent is doing, isolate risky work on a branch, and own the workflow instead of renting another closed tool.

Founder’s Edge

This Week’s Builder Toolkit

• Dev Tool: Resend is an API-first email platform replacing SendGrid and Mailgun with cleaner docs and a developer experience built to work on the first try.

• Free Dataset: The State of AI Report 2025 by Nathan Benaich and Air Street Capital is the most comprehensive free industry snapshot, covering research breakthroughs, commercial deployments, geopolitical shifts, and safety. 200+ pages, refreshed yearly, treated as required reading by most operators in the space.

• No-Code App: Stack AI lets non-technical founders build AI agents and workflows with a visual interface, with built-in support for RAG, function calling, and multi-step automation across 100+ integrations.

• Productivity Hack: Sunsama is a daily planner that pulls tasks from Notion, Linear, Asana, Trello, GitHub, and your calendar into a single timeboxed view. Founders use it to plan tomorrow at the end of today, which removes morning friction and protects deep work time.

• Learning Resource: Learn Context Engineering by Scrimba is a hands-on course on deciding what an LLM should load, keep, compress, or drop at each step. You’ll build a chat app that doesn’t crash or forget as you go, editing real code in the browser rather than watching slides.

Note: If you’ve found a tool that’s sped up your build process, hit reply and share it. We’ll feature the best submissions in a future issue.

AI Founder’s Journal

The Free Tool Worth $94M ARR

In 2019, Matthieu Rouif quit his product manager job at GoPro with a small idea.

He’d seen friends struggle to take clean product photos for the secondhand resale apps they were using. Vinted, eBay, Depop, Le Bon Coin.

Every listing needed a clean background, and every seller spent fifteen minutes faking one with editing apps that weren’t built for the job.

He partnered with Eliot Andres, a ML engineer, and the two of them built an MVP in 14 days. A single-feature app.

Upload a product photo. Get back a clean, background-removed image you could drop into any listing.

They launched on the App Store in February 2020 under the name Photoroom.

The product didn’t try to be a full photo editor. It didn’t try to be an AI art generator.

It did one thing only. It removed backgrounds from product photos better than anything else on mobile.

That focus was the moat.

Resellers found it. Etsy sellers found it. Shopify merchants found it.

Each one had the same problem (their listings looked unprofessional next to the polished ones) and Photoroom gave them a fix that took five seconds instead of fifteen minutes.

Word spread inside ecommerce Facebook groups, Reddit threads, and reseller Discord servers.

By August 2020, six months after launch, Photoroom had 300,000 monthly active users and $1M in ARR.

By December that year, they were at $2M ARR and growing 50% month over month.

Here's the part most founders miss.

The free background remover was the entire funnel.

Users came in to remove a background. The app was free.

After two or three uses, the in-app upsell pitched the broader Photoroom Pro experience:

AI-generated backgrounds, batch editing, brand templates, instant shadows, magic retouch.

The conversion ran on its own because users had already felt the speed advantage on the free utility.

The Trojan horse worked because Photoroom never broke the trust of the free user.

The background remover stayed free, stayed fast, stayed unlimited for casual use.

The paid plan added features sellers wanted as their listing volume grew, not as a gate on the original feature.

The Results

Photoroom hit $50M ARR in late 2023, four years after launch.

By the end of 2024, they were at $94M ARR, up 89% year over year, and profitable.

They’ve since crossed 150 million app downloads.

Their API now powers product image generation for Shopify, Netflix, and Warner Bros.

They raised $43M at a $500M valuation in February 2024 from Balderton and Aglae Ventures, despite already being cash-flow profitable.

Here's what made it work:

• Solve one painful, specific problem better than anyone else. Photoroom didn’t try to be Photoshop. It didn’t try to be Canva. It picked one workflow (clean product backgrounds for online sellers) and got obsessively good at it. The narrowness of the target audience became the marketing advantage. Every seller who tried it told three other sellers.

• Use the free utility as the SEO and word-of-mouth engine. The background remover at photoroom.com ranks for thousands of high-intent search terms (remove background free, transparent background maker, product photo editor). Every search becomes a conversion funnel. The cost of acquiring the free user is essentially zero, and the upsell mechanics inside the app convert the high-intent ones into paid subscribers.

• Make the paid product feel like an extension, not a tax. Free users could remove unlimited backgrounds. The paid tier added speed, batch features, and AI-generated backdrops sellers couldn’t get elsewhere. Sellers upgrading didn’t feel they were paying to keep using a feature they already had. They were paying for new capability that fit their growing volume.

• Stay profitable, then raise. Photoroom broke even one year after launch and stayed profitable. They didn’t raise their Series B until they’d four years of operating data and a real product moat. That bargaining position translated into a $500M valuation on $43M raised, with founders retaining most of the equity.

Key Takeaway

The free utility tool is the highest-leverage acquisition channel a small AI founder has access to right now.

You don’t need to build a full SaaS product to start. Build one narrow tool that solves one painful workflow better than anything else on the internet.

Ship it free.

Let SEO and word-of-mouth fill the top of the funnel. Then upsell users into the broader paid product once they trust the speed and quality of the free version.

The mechanism is simple. The discipline is in resisting feature creep on the free utility long enough for the funnel to compound.

— Matthieu Rouif, Photoroom co-founder and CEO

Weekly Challenge

One Experiment. One Week. One Win.

The Goal

Run 5 customer conversations this week and ship one feature, fix, or change based on what you hear by the end of the week.

How It Works

1. Pick 5 customers you haven’t talked to in the past 30 days. A mix of high-volume users, low-volume users, and recent signups.

2. Send each one a short message. Ask for 15 minutes. No agenda beyond understanding their workflow and what’s frustrating them right now.

3. After the calls, identify the one pattern that came up more than once. Build the fix, ship the change, or write the help doc.

4. Send a follow-up to those 5 customers letting them know what shipped.

Why It Works

Most founders rely on dashboards and Slack threads to understand their users.

Both lag reality by weeks.

Direct conversations surface friction no analytics tool will catch, and shipping a fix within days proves to those customers their voice moves the product.

Spotlight

Share what you heard and what you shipped in the TAAFT community by end of week. We'll showcase the fastest customer-to-feature turnaround in an upcoming newsletter.

AI Market Watch

Deals, Discoveries, and Demand

Megadeals

• CuspAI – $450M Series B (AI for materials discovery).

• Fireworks AI – $1.505B Series D (platform to train and serve custom models).

• Spectro Cloud – $100M+ Series D (AI infrastructure management software).

• Chai Discovery – $400M Series C (AI-driven molecular design for drug discovery).

• Walden Robotics – $300M launch out of stealth at a $1.1B valuation (general-purpose robots for manufacturing and logistics).

Top Research

• Metacognition in LLMs (first broad review of whether models know what they know)

• When is Routing Meaningful? (fewer than ten differentiated models capture most of a router’s value)

• Self-Improvements in Modern Agentic Systems (survey of agents that convert experience into lasting capability gains)

• Tracing Agentic Failure from the Flow of Success (finds the faulty step by training only on successful runs, 200-5000x faster)

• Safety and Alignment in an Era of Long-Horizon Models (a long-running model exploited its sandbox to slip past safety scanners)

Search Trends

Top “best AI for…” searches:

1. Video Generator

2. Image Generator

3. Image Editing

4. Story Writer

5. Text to 3D

Thanks for reading!

Got a project in motion? Hit reply and share what you’re working on with us.

Keep an eye on your inbox. We’ll be back soon with more founder stories, fresh tools, and data worth your time.

Signing off,
— AI Empires