Altostratus AI: Building a Proactive AI Travel Disruption Service

Summary

Altostratus AI was a proactive travel disruption service that detected flight cancellations, texted travelers immediately, and offered to call the airline on their behalf to find the next available option. I built it as a solo founder, owning product strategy, UX, architecture, AI agent design, and go-to-market testing. The product validated strong traveler demand, but the business ultimately required enterprise distribution through Travel Management Companies (TMCs) or airlines. As a solo founder without venture backing or an enterprise sales team, I chose to wind it down rather than pursue a capital-intensive path I did not believe fit the likely exit economics. (Agent view demo, designed for desktop)

The Opportunity

In short: build for where the market is going.

Proactive AI traveler support is inevitable for the travel management industry because of the cost savings and the ability to provide superior traveler experiences.

Two things drive success for travel management companies (TMCs):

1. Clients want to save money.
2. Travelers want a better business travel experience.

There is an endless list of features that travel management companies focus on, but these two goals sit underneath almost every product request. Clients want lower travel costs and a better traveler experience. If a travel management company can provide those things in a clearly superior way, it has a meaningful advantage when competing for enterprise clients.

By automating operations in a way that was previously impossible, TMCs can simultaneously become far leaner organizations while providing every traveler with VIP-level support during their trips. This will translate directly into more competitive bid processes between TMCs: When one TMC is half the cost with 10x better service, it will be difficult for any TMC still relying on their global operations teams to compete.

This may look like a new era enabled by AI, but it's really a continuation of a decades-long trend of software replacing operations. It happened first in leisure travel on the booking side, and now we're going to see software replace service operations as well. You don't call your local travel agent when you book a family vacation. Soon, it will feel just as outdated to wait on hold yourself when your AI agent can do it for you.

The Product

Altostratus AI tracked upcoming flights. When a disruption occurred, it proactively texted the traveler and offered to help.

The pitch to the traveler was simple:

Sorry to see that your flight was canceled. Do you want me to call the airline and get you on the next available flight?

The product was intentionally not positioned to travelers as "an AI tool." I was not selling AI. I was selling a solution to a painful problem. Customers do not want AI. They want their trip fixed. AI was simply the technology that made the service possible at scale.

With the proactive customer service model, Altostratus AI could present the solution at the exact moment of pain. The airline sends a notification saying your flight was cancelled, and seconds later Altostratus AI texts you and offers to help.

The product did not require the traveler to remember to open an app, search for a support number, or understand what to do next. It reached them at the moment when help was most valuable.

There are cases where an airline automatically rebooks you on a reasonable flight, in which case you don't need any assistance. Unfortunately, that's not always the case. Airlines hold an obligation to get you to the destination, but not to do it quickly. If you don't have any connecting flights, and the guy next to you has a connecting flight to India, they're going to prioritize him over you since his connecting ticket is harder to rebook. My wife and I experienced this flying back from Japan. We landed in Seattle and saw the text, "Your flight to Portland at 12:30 PM has been canceled. Your new flight departs at 11:35 PM." There are flights from Seattle to Portland every hour. By standing in line and advocating for ourselves, we were able to get bumped to a 5pm flight and make it home at a reasonable time. I wish Altostratus existed at the time to make that day easier.

What I Built

I built the MVP as a solo founder, covering the full product and technical stack:

• Web application and API built with Next.js
• Supabase backend for auth, database, and real-time state
• Twilio SMS for traveler communication
• Twilio Voice for outbound airline calls
• OpenAI Realtime Voice for AI phone conversations
• Flight tracking logic to monitor upcoming trips
• Internal context management to keep traveler state, flight state, and conversation state synchronized
• Multiple AI loops to determine when questions and answers needed to be relayed between the traveler and airline conversations, operating in real time with low enough latency for phone calls
• Human monitoring and takeover workflows
• Safeguards to reduce hallucination risk in a high-stakes customer service scenario

The hardest product challenge was not getting an AI model to talk. It was designing a reliable operating model around the AI: what it was allowed to do, what it was not allowed to do, how it should handle uncertainty, when it should escalate, and how a human could intervene if the model approached a risky edge case. Before making any travel decision or when encountering unknown info, the voice AI that is talking to the airline would say, "hold on, let me go ask the traveler about that." It would then text the traveler with the question, wait for a response, and relay that response once it was received.

Because the product operated in a high-stakes travel scenario, trust mattered. A hallucinated flight option, missed rebooking option, or unauthorized traveler decision can materially harm a customer. Altostratus AI therefore needed to behave less like a chatbot and replicate the actual behavior of a human travel agent.

Product Principles

1. Solve the customer problem, not the AI demo

The customer-facing promise was not "talk to an AI agent." It was "I'll help get your trip back on track."

Everyone has experienced the pain of a flight cancellation. The promise was simple: I'll make it suck less. There is no need to wait on hold with the airline or stand in a long customer service line. Altostratus AI could deal with the airline while the traveler did something else.

2. Be proactive at the moment of pain

Most travel support is reactive. The traveler has to identify the problem, find the support channel, wait in line, and explain the situation.

Altostratus AI reversed that pattern. It detected the disruption and reached out first. This was also the trigger event for an API integration with any potential enterprise customers. Travel companies are already tracking your flight. With very few code changes, you could integrate proactive rebooking assistance powered by Altostratus AI.

3. Keep the traveler in control

The traveler needed to know what Altostratus AI was doing and have confidence that the system would not make unwanted decisions on their behalf. AI must know the limits of its knowledge, and ask the traveler if anything is unclear. The traveler could also text Altostratus at any point during the process to request updates or changes.

4. Design for human takeover

The goal was not to pretend AI could resolve every edge case perfectly. The goal was to let AI handle repeatable work while allowing a human to monitor and take over when necessary. In a fully rolled-out product, reducing human takeover events becomes a key KPI, much like developing a self driving car.

If the AI failed, I would personally be on the phone with an airline to get a trip back on track. That handoff is seamless to the traveler of course, but it's essential for fine-tuning any AI system that offers to solve a problem. It has to work 100% of the time.

(This turned into a minor logistical challenge as a solo-founder, as I found myself tracking my customers' flights at all hours of the day and night for months on end.)

5. Delete the UI

From the beginning, this was built around the idea that a helpful AI system will be invisible until you need it. This is an AI service designed to help you at the exact moment your stress levels just spiked to their highest point in the last month. If I asked users to log in to an app, enter their flight details, and then be able to access AI customer support, the friction would be too high to ever get a single user.

Additionally, it was always designed for API-level integrations with enterprise clients. As mentioned in point 2 above, your travel company is already tracking your flight. Let their system determine when the flight cancellation happened, and then initiate the travel rebooking process. From the traveler's point of view, this is just their TMC reaching out via text and being genuinely helpful at their exact moment of pain.

The Plan

Historically, I've had success gaining support for original product ideas inside large organizations when there is a strong proof of concept. This is how I turned my side project of a hotel rate recommendation feature into BCD Travel's top development priority. A working prototype is more persuasive than any Excel model or PowerPoint deck.

With Altostratus AI, my plan was straightforward:

• Build a working travel disruption protection service powered by AI
• Sell it to the public as a proof of concept
• Share this working solution with travel management insiders who understood the problem but had not yet settled on a solution

A working product mattered because no one else could offer that "wow, this is the future" moment of proactive customer service solving a painful problem in your own life for a low cost. Every company wants to create the next great travel experience, and no one else was offering live, proactive service like this.

Validation

After six months of building, I launched the platform to real travelers in mid-2024.

With simply sharing this on a few forums and social media posts, about 50 people signed up, including many with whom I had no personal connection. Around 100-150 flights were tracked. This strong initial response was super helpful because I could still manage this number of bookings solo while getting valuable feedback from early users. The pain point was easy to understand because almost every frequent traveler has experienced a cancellation, long hold time, or customer service line during a disruption. Airline customer service sometimes feels like talking to a robot. Now you could make them talk to your own robot.

Individual travelers liked the product and understood the value quickly. The service was easy to explain because the pain was familiar: when your flight is canceled, you want someone to fix it while you do literally anything else. Even after I later wound down the company, I had friends asking if it was still running because their flights were cancelled and they wanted help.

No one cared if it was AI or not. They had a painful problem and they wanted it to go away.

That said, the consumer business model had a structural challenge: flight cancellations are common enough to be painful, but not common enough for a consumer business to work without large-scale distribution. Based on industry averages of how often flights are cancelled, I needed to track about 60 flights on average to encounter one cancellation where Altostratus AI might be able to help and earn a small fee. Doing the math, the consumer product was unlikely to become profitable without significant scale.

The stronger business model was always enterprise distribution: partner with a large company that already had thousands of bookings and wanted disruption protection for their travelers.

That scale would have done two things:

1. Created enough disruption volume to make the economics work
2. Generated enough live operational data to improve the AI workflow and reduce the need for human intervention

Additionally, an API-level integration with a TMC would enable this to be a truly invisible AI product to travelers until the moment they need it. The consumer product still required travelers to track their flights with the Altostratus website in advance. An API integration with a company that has already tracked your flights eliminates that last piece of the UI. You wouldn't even need a user account.

Why I Wound It Down

The core problem was not traveler demand. The core problem was distribution.

As a one-person company, I did not fit the profile of a software vendor for a large enterprise. A large company expects procurement readiness, security reviews, implementation resources, account management, and forward-deployed engineers who can integrate the product into the client's systems.

Altostratus AI was designed for an easy API integration, but that was not enough. My pitch was less compelling than rival startups that appeared more stable because they had larger teams, more funding, and more enterprise implementation capacity.

I later came across this X post from Paul Graham:

Excellent advice. I wish I had seen it earlier.

With competitors gaining more traction in the market, and with the enterprise path requiring a level of capital and team size I didn't believe was going to lead to a good outcome as a founder, I decided to wind down Altostratus AI at the end of 2025.

On Raising Venture Capital

Altostratus AI had a strategic tension: the business likely needed enterprise sales resources to succeed, but I did not believe the likely outcome justified a large venture-backed path.

The product was valuable, but the acquisition ceiling seemed tied to how difficult it would be for a large travel company to build the same capability internally. As AI-powered software development improves, that build-versus-buy threshold moves lower every year.

The likely acquisition outcome for this type of company was modest. A small acquisition could be life changing for a solo-founder who raised very little money and retained most of the company. The same exact acquisition could be disappointing for a venture-backed company with multiple founders, a large team, and investors expecting large returns.

Additionally, the math doesn't work for VCs if you're honest about your likely acquisition size. They don't want to invest $10M for a startup that will get acquired for $12M. You would need to tell them you're building a $200M+ company, and I simply don't believe any startups building these products will see those returns.

What I Would Do Differently

If I were considering starting Altostratus AI again, I would find other entrepreneurs who decided to raise venture capital and work for them instead. This is a viable business, but the reward of being a founder of a company like this is unlikely to be commensurate to the risk. The most likely financial outcome for anyone working on this business is the salary they took along the way.

What I Learned

Altostratus AI made me sharper as a product builder because I had to own the entire process: market thesis, customer pain, UX, technical architecture, AI behavior, safety, sales, support, and the decision to shut it down.

The biggest lessons:

• Startups should sell to startups.
• Enterprise trust is a product requirement. Enterprise buyers don't just evaluate the product, but also the vendor's integration risk, support models, procurement readiness, and operational continuity plans.
• It's hard to make an impressive AI demo. Something 6 months old already looks outdated.
• Making any AI automation product is much like designing self-driving car technology. Like self-driving cars, that last 1% of automation takes 99% of the work. After the initial release, your core KPI becomes reducing instances of human intervention.

Following Altostratus AI, I applied some of these lessons to Canopy Commons, a network for small business owners to find new retail markets. The contrast was immediate. For that business, it did not matter that I was a one-person company. I could message small business owners directly, they would sign up, and I could get real customer feedback and usage immediately without going through any enterprise procurement process.

Closing

Altostratus AI did not become the company I hoped it would become, but it remains one of the clearest examples of the kind of product work I like most: finding a painful operational problem, building a working product quickly, using AI as a driver of efficiency rather than a marketing term, and making the right decisions when the market reality becomes clear.

It made me better at product strategy, AI product design, enterprise workflow thinking, and founder-level tradeoff decisions. Those are the skills I want to keep applying in product roles where ambiguous problems, technical products, and real customer pain all meet.