Hey - It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 3 most important things:

• Picnic, a startup building robots that made pizza, has shut down — learn why below

• 24 tips for giving S-tier demos

• Microsoft Build 2026 was this week — all important announcements below

A huge thanks to today’s sponsor, HubSpot. Download their free AI guide to grow your business today.

Demystify AI for Your Business in Just 4 Steps ^AD

Overwhelmed by AI? HubSpot's free guide cuts through the noise. Get the ultimate crash course for non-technical entrepreneurs who want to harness AI's power—without getting lost in the jargon.

"AI for Business Builders" delivers:

• A 4-part roadmap to AI mastery

• Jargon-free explanations of large language models

• Practical prompt engineering tips you can use today

• Real-world examples of AI boosting businesses like yours

Arm yourself with the knowledge to make informed AI investments and skyrocket your startup's growth.

Download Your Free Guide Now

This Week In Startups

🔗 Resources

Startups become what they hunt

24 tips for giving S-tier demos

Small Businesses are the Next Frontier for AI

Your startup didn't work out. Now what? * 

📰 News

OpenAI launches new Codex tools for white-collar work

Anthropic files to go public

Claude Opus 4.8 is out and is SOTA

Meta rolls out a new AI creator assistant on Facebook

💸 Fundraising

OpenRouter, an AI model exchange and inference routing platform, raises $113M Series B funding

Reactor, a real-time generative video platform for developers, raises $59M Series A funding

Catena Labs, an AI-native banking infrastructure startup, raises $30M Series A funding

Saris, an AI workflow automation platform for banks and credit unions, raises $28.8M Series A funding

* sponsored

Fail(St)ory

Robot Aquarium

Picnic built a robot that could sauce, cheese, and toppings at assembly-line speed.

It could sit inside a commercial kitchen and help one worker turn out 100+ pizzas an hour. That made it easy to like and easy to demo. 

The problem was turning a pizza-making robot into something enough operators needed every month.

What Was Picnic:

Picnic built the Picnic Pizza Station, a robotic pizza assembly system for commercial kitchens. It applied sauce, cheese, pepperoni, and other toppings onto dough, then handed the pizza off to the rest of the kitchen workflow. 

Pizza was a smart first target. The work is repetitive, the ingredients are standardized, and rush-hour demand can be painful to staff. 

The company claimed one employee could produce roughly 100+ customized 12-inch pizzas per hour with the system, depending on setup and size.

That was the pitch: fewer hands on the line, more predictable output, less waste from inconsistent topping.

Picnic sold the machine through a leasing model. When it reached the market in 2021, pricing started around $3,500 per month.

That made the product feel closer to an operating expense than a large capital purchase, but the customer still had to believe the machine would earn its place every month.

The strongest use cases were dense-volume environments: stadiums, campuses, military sites, large venues, and high-throughput foodservice operators. Those are the kinds of places where pizza behaves more like production.

A normal restaurant is messier: some hours are slow, menu changes matter, and kitchen space is tight. Staff still need to load ingredients, monitor the machine, clean it, recover from jams, and trust it during the exact moments when failure is most expensive.

Picnic had real customer interest. It partnered with Ethan Stowell Restaurants, tested with Domino’s Pizza Enterprises in Berlin, listed customers like Aramark, Chartwells, and Compass Group, and pushed into venues and institutional foodservice.

It also carried the cost structure of a hardware company. Machines had to be manufactured, shipped, installed, maintained, monitored, and supported. In 2024, Picnic raised $5M to scale production and meet demand across North America.

By May 2026, the company had become insolvent. Its assets and IP were sold to an undisclosed buyer through liquidation. 

One customer described being left with a $250,000 “robot aquarium,” which says a lot about the risk operators take when their automation vendor disappears.

The Numbers:

• 🍕 Founded: 2016 in Seattle

• ⚙️ Throughput claim: roughly 100+ pizzas per hour

• 💸 Lease pricing: started around $3,500/month in 2021

• 📈 Funding: roughly $50M-$53M raised

• 👥 Peak team: about 100 employees by 2023

• 🏟️ Customers and pilots: MOTO Pizza, Domino’s Pizza Enterprises, Aramark, Chartwells, Compass Group, military and venue installs

Reasons for Failure: 

• The ROI depended on dense, predictable pizza volume: Picnic made the most sense when pizza production was concentrated: stadium rushes, campuses, bases, large venues, and very busy shops. In lower-volume restaurants, a $3,500/month lease had to compete with slow hours, uneven staffing needs, and limited kitchen space. 

• The robot automated only one visible part of the job: Topping is important, but it is only one step in a pizza operation. People still had to prep ingredients, manage dough, bake, cut, box, serve, clean, and fix problems when the workflow broke. For operators, the buying decision was really about total kitchen reliability, not demo throughput. A robot that needs monitoring and support can still save labor, but the savings need to survive real shifts.

• Hardware burn met a colder funding market. Former CEO Clayton Wood said Picnic was “caught in the squeeze” between the free-money period and the harsher post-2022 market. That matters for a company selling physical systems. Picnic needed capital for manufacturing, deployment, maintenance, and sales cycles that move slower than software. 

• Food robotics had become a scarred category. Picnic was selling into a market where Zume, Basil Street, Piestro, and other restaurant automation startups had already failed. That gave buyers a fair reason to ask what would happen if the vendor disappeared and the robot became expensive furniture.

Why It Matters: 

• Sell the customer the full outcome, because automating the most visible task may still leave them with the annoying parts.

• Your best use case is only a market if there are enough buyers with the same pain, budget, and workflow.

• A buyer’s fear of being stuck with dead equipment can matter as much as your ROI math.

Trend

Microsoft Build 2026

Microsoft used Build 2026 to show how much of the enterprise agent market it can absorb by default.

It already controls the work context, identity layer, data platform, dev workflow, runtime, and endpoint. Now it is wiring those pieces together so agents can use company files, calendars, permissions, business data, and approval rules without customers stitching the stack themselves.

For founders, generic workplace agents just got harder to sell.

Why it Matters

• Enterprise context becomes distribution. Microsoft IQ gives agents company context from files, people, workflows, data, and the web. Startups have to rebuild that through integrations; Microsoft already sits where the work happens.

• Generic coordination agents get squeezed. Scout handles meetings, prep, stalled work, calendars, and deliverables across Microsoft 365. Startups need a narrower workflow, unique data, or a buyer Microsoft will not serve well.

• Microsoft wants model control. Copilot leaned heavily on OpenAI. The new MAI models give Microsoft more control over cost, latency, routing, and product behavior.

Microsoft IQ makes company context the product

Microsoft IQ is Microsoft’s name for the context layer underneath its agents.

Think of it as the system that tells an agent what it needs to know before it acts: who works on what, which files matter, how the business defines key terms, what happened in meetings, and what outside information is current.

It has four pieces:

• Work IQ: context from Microsoft 365: files, chats, meetings, people, calendars, and workflows.

• Foundry IQ: internal company knowledge for agents built in Microsoft Foundry.

• Fabric IQ: business meaning for company data, like how customers, products, invoices, and metrics connect.

• Web IQ: fresh external information from web pages, news, images, and video.

The idea is that companies define this context once, then reuse it across Copilot, Foundry, Copilot Studio, and Microsoft 365. That makes agents less dependent on long prompts and one-off integrations.

Scout is Microsoft’s always-on work agent

Scout is Microsoft’s first Autopilot: an agent that can run in the background instead of waiting for a prompt every time.

It connects to Microsoft 365 tools like Teams, Outlook, OneDrive, SharePoint, calendar, email, contacts, and browser context. The job is coordination work: scheduling meetings, preparing users, flagging stalled decisions, blocking focus time, and noticing upcoming deliverables.

Under the hood, Scout is built on OpenClaw, Microsoft’s open-source agent runtime for autonomous workflows.

The important part is control: Scout gets its own governed Entra identity, so its actions are visible and attributable. Companies can limit what it can access, require approval for sensitive actions, and apply existing Purview policies when it handles data.

GitHub Copilot is getting a new app

The new GitHub Copilot app is a desktop app for running coding agents as separate work sessions.

Instead of asking Copilot for code in the editor, a developer can spin up agent sessions from GitHub Issues, Pull Requests, or previous sessions. Each session gets its own branch, files, conversation history, and task state through git worktrees.

The practical change is parallel work. One agent can investigate a bug, another can draft a feature, and another can respond to review comments, without mixing files or branches.

The developer still decides what gets merged. Copilot is taking on more of the execution loop, while the human keeps the review, testing, and approval role.

Microsoft is reducing its OpenAI dependency

Microsoft also announced seven in-house MAI models for reasoning, coding, image generation, transcription, and voice.

The main one is MAI-Thinking-1, a 35B-active-parameter reasoning model for multi-step tasks, long-context work, and code generation. It is still limited to early partners.

This is important because Copilot and Microsoft’s AI products have relied heavily on OpenAI models. The MAI models give Microsoft more control over cost, latency, routing, and product packaging. Simple tasks can run on cheaper Microsoft models, while harder tasks can still use frontier models when needed.

Help Me Improve Failory

That's all for today’s edition.

Cheers,

Nico