How Milo works

Last updated November 11, 2026

AeroMilo is Contrail's flight-delay helper, and Milo is the bird who answers you. The idea fits in one line: a little bird told you. And the early bird gets the first word. He's built on insight from flight attendants, pilots and air traffic controllers around the world. Other trackers tell you what's happening. Milo tells you what to do. This page lays out how he does it: every source he reads, the rules he follows, where AI comes in and where he can be wrong.

What he does that trackers don't

He makes his own call. A tracker repeats the time the airline posts. Milo works out when your flight can realistically leave and shows both times side by side when they differ.
He follows the plane. Most delays start with the aircraft that is supposed to fly you. Milo tracks that plane, the flight before it and its last three days.
He tells you the next move. Backup flights, what to say at the desk, what you may be owed and whether a rebooking the airline offers is any good.
He keeps score. Every prediction is logged against what happened, in public.
He learned from insiders. Milo's rules come from conversations with flight attendants, pilots and air traffic controllers about how delays really happen and what travelers should do.
He's free as a bird. No cost and nothing to download.

Where his information comes from

Milo reads these sources directly. Flight predictions come only from flight, FAA and weather data. News is used for context and advice and never changes a forecast time. He does not use social media or other trackers.

FlightAware. Flight schedules, live status, gates, the incoming plane, tail numbers and each plane's recent flights. A lookup is reused for two minutes.
FAA airport status. Ground stops, delay programs, closures and arrival or departure delays at each airport. Refreshed every two minutes.
FAA daily operations plan. What air traffic managers expect later in the day, such as a ground stop that is "possible after 3 p.m." Refreshed every five minutes.
Aviation Weather Center. Current conditions and the official airport forecast at both ends of your trip, plus active thunderstorm warnings, including ones lying across your route.
National Weather Service. Warnings and watches for winter storms, ice, hurricanes, high wind, severe thunderstorms and dense fog. US airports only for now.
TSA. Typical security wait times by airport, day and hour. These are historical averages, not a live count of the line.
Google Maps. Live traffic and transit times, only when you ask when to leave.
Eurocontrol. Europe's air traffic network manager publishes each country's on-time rates and air traffic delay causes a day later. Milo shows the latest day on European airport cards as background. It is not live and never changes a forecast.
EU and UK passenger rules. The fixed compensation amounts and care rules for delays and cancellations, known as EU261 and UK261, as they stood in November 2026.
US Department of Transportation. The refund rules, the tarmac delay rules and what each airline has promised in writing for delays and cancellations.
Airline logos. Loaded from a public logo image service. They are shown for recognition only.
Contrail's reporting. Milo reads the headlines and summaries Contrail publishes, refreshed every 10 minutes. When a recent story mentions your airline or airport, he links to it under the card. In conversation he treats Contrail as his own newsroom and says "Contrail reports".
The web. When you ask about something outside the flight data, such as a strike, a travel waiver or an airline policy, and Contrail has not covered it, the AI can search the web. It is told to prefer airline travel alert pages, the FAA, the DOT, the TSA and major news outlets, to name its source and to say "reportedly" for anything that is not from an official page. Sources appear as buttons under the answer. This is new and may not be switched on for every answer.
Contrail's editor. Short notes posted by Contrail's newsroom when something outside the data matters, such as a staffing shortage or a strike. Milo says when a note comes from the editor.

Bird's Eye View: how he judges a flight

Bird's Eye View is the name of the prediction model behind Milo. It is Contrail's own: the signals it reads, how it weighs them and the odds it produces. For a flight that has not left, it checks each signal below. Each one can raise the risk to "Some risk" or "High risk". The highest one wins.

The delay the airline already shows. 15 minutes or more raises the risk. 45 or more makes it high.
The incoming plane. He takes when your plane is due in, adds the time it needs at the gate and compares that with your departure time. If the result is 15 minutes or more late, he says so.
Time at the gate. He assumes about 25 minutes for a regional jet, 35 for a standard jet and 70 for a wide-body jet. These are working estimates, not airline figures.
One plane further back. If your plane has not left for you yet, he checks the flight before it too. A delay there can reach you hours later.
A plane that just arrived. A plane at the gate is not the same as a plane that is ready. If it pulled in too late to turn around, he counts that.
Plane swaps. He remembers which aircraft was assigned to your flight. If the airline changes it, he checks whether the new one can make the posted time.
The plane's last three days. Cancellations, diversions and delays that began at the gate after the plane had arrived on time. Repeated trouble with one aircraft may point to the plane itself. Milo cannot see maintenance records, so he says "may".
The FAA at both airports. A ground stop or closure at your destination, a delay program and long arrival or departure delays all raise the risk.
The FAA's plan for later. Trouble the FAA expects at either airport later in the day.
Weather. Storms, low visibility, strong gusts or snow in the forecast around when you leave or land, official warnings at either airport and thunderstorms across the route.
How the airport is running. The FAA's plan for later in the day when it covers your flight, how many planes an hour the airport is landing against its normal rate, and whether planes are being deiced.
The live flow of traffic. How far apart planes headed to your airport are being spaced, delay programs that have been proposed but not started, and runways that are closed. This comes from a live government air traffic feed that Milo listens to around the clock. It covers airline flights only, it is for information and not for flying or dispatching, and it can drop out without warning. When it does, Milo carries on with his other sources.
The flight's track record. How often this flight number ran 15 minutes or more late or was canceled on this route in recent days. Late 60% of the time or more raises the risk.
FAA takeoff slots (experimental). When the FAA is metering an airport, it assigns affected flights a takeoff time. Milo tries to look up your flight's slot, which can show a delay before the airline posts it. This is new and may not return anything yet.
A flight that is overdue. Ten minutes past its time and still at the gate with no new time posted.

Delay odds

Every flight that has not left gets odds from Bird's Eye View: the chance it leaves 15 minutes or more late, the chance it leaves an hour or more late and the chance it is canceled. The odds move as the day does, and the card shows where they opened, how far they have moved and what is moving them. Nobody is betting on anything. It is Milo's own estimate, shown the way a price is shown.

What it was built from. Two sets of the Department of Transportation's on-time records, and nothing older than 2023.

The national record. Every US airport and airline in the records from July 2023 to June 2025: about 15 million flights, 362 airports and 21 airlines. From it Milo knows how much more or less often each airport and each airline runs late or cancels than average, and how the time of year changes things. That covers small airports as well as big ones.
Flight-by-flight records. About 435,000 individual departures in 2023, each with its scheduled and actual times, its airline, its aircraft and the weather that hour. These teach Milo the things a monthly total cannot show: time of day, a late incoming plane, a backed-up airport and weather.

What the records showed.

Where you fly matters. Among the busiest airports, about 15% of Salt Lake City's flights ran late and about 25% of San Francisco's did.
So does who flies you. By airline the late share ran from about 15% to 29%.
And when. About 14% of flights ran late in October and November and 28% in July.
Time of day matters most early on. About 9% of 6 a.m. departures left late. By 8 p.m. it was 33%.
A flight's recent habits carry over. Flights late on fewer than 1 in 10 recent days ran late 11% of the time. Flights late on 6 in 10 recent days ran late 49% of the time.
A late plane stays late. When the same aircraft had already left 45 minutes to 2 hours late earlier that day, its next flight left late 71% of the time. When it had been on time, 19%.
A backed-up airport drags everyone. When fewer than 1 in 10 recent departures at the airport were late, a flight ran late 10% of the time. When half or more were late, 66%.
Weather. Heavy rain doubled the late rate, from 20% to 42%.
Low clouds and visibility. Milo reads the forecast cloud height and visibility at both airports for the hour you leave and land, sorted into the same bands pilots and controllers use. The lower the band, the fewer planes an airport can handle an hour, and the more the odds rise.

How it was tested. Milo was given only the 2022 national record and the first nine months of 2023, then asked for odds on 104,245 flights from October to December 2023 that it had never seen. It ranked late flights above on-time ones 69% of the time, where a coin flip scores 50%. Adding the national record is what lifted that from 66% hours ahead. Its odds ran high: the tenth of flights it rated riskiest averaged 50% and 40% left late, in months that were calmer than usual. That gap is the kind of thing Milo's own results will correct.

What is judgment, not data. The records cannot tell Milo how much an FAA takeoff slot, an official weather warning, a plane with a rough week or a delay the airline has already posted should move the odds. The same goes for the FAA's plan to slow an airport later in the day, an airport landing fewer planes than its normal rate, deicing, spacing between planes and a closed runway. The records hold visibility but not cloud height, so low cloud counts the same as the matching visibility band, and conditions at the airport you are flying to count half as much as at the one you leave from. Those are set by hand. A posted delay puts the odds at 97%. Milo's own math on a late incoming plane puts them between 72% and 90%.

Limits. The national record ends in June 2025 and counts late arrivals, which Milo uses as a stand-in for late departures at the same airport. The flight-by-flight records all leave from New York's three airports, so time of day, weather and late-plane effects are assumed to work the same way elsewhere. For Canada and Europe, Milo ran his own study: 141,592 departures from 96 airports on 12 days between December 2025 and October 2026, bought from FlightAware. It gives 61 airports and 45 airlines outside the US a starting track record. Twelve days is a small sample, so those numbers are cut down hard. Where the study and the 15-million-flight US record cover the same airports, they agree only loosely, and the study's figures are scaled to match: an airport counts about a third of what the study alone would say and an airline about half. The study's cancellation counts were not reliable, so cancel odds outside the US still start at the national average. The odds only update when someone checks the flight. From November 11, 2026, every flight's opening odds are logged against what happened, and the accuracy page will show whether a "30%" really comes out late 30 times in 100.

His forecast time

Milo turns those signals into one expected departure time: the latest time any of them points to, rounded to five minutes. When that is at least 15 minutes later than what the airline shows, the card puts the two side by side as "The airline" and "Milo".

He always says "likely" and never "will". Odds are an estimate, not a promise. Airlines swap planes and make up time, and a forecast can miss.

Where AI comes in

The rules above are plain arithmetic. They do not use AI and they give the same answer every time. AI is used in four places:

Conversation. When you ask a question in your own words, an AI model reads it, looks up what it needs with the same tools described on this page and answers in Milo's voice.
"My read". On a flight that has not left, the AI gets the same signals as the rules and gives a second opinion in a sentence or two. It can lean up to 30 minutes either way and has to say why. It does not change the forecast time in the box.
Outside events. The AI reads Contrail's headlines and can search the web, as described above. Web results can be wrong or out of date, which is why Milo names the source.
Reading a confirmation. If you paste an airline confirmation, the AI pulls out the flight numbers and dates.

The AI model is made by OpenAI or Anthropic. It can misread a question or state something wrong, which is why the cards show the underlying data and every delay answer ends with "confirm with your airline".

Texting Milo

Milo can also answer by text message. A text goes through the same AI and the same tools as the website, with one difference: there are no cards in a text, so he writes the key facts out in a few sentences. He can also text you first. Link your phone from your account, or send him a flight number and reply WATCH, and he texts you when that flight changes. Reply STOP and he stops.

What is new and still being proven

These parts were added recently and have not yet been checked against many real flights: looking two planes back, the three-day aircraft history, plane-swap detection, delay odds, "My read", the FAA takeoff-slot lookup, Contrail story matching, web search, Europe coverage and texting. Treat them as early signals.

Where he can be wrong

He is strongest in the US. There he has the FAA's live delay feed. In Canada and Europe he follows your plane and reads the airport weather, but he has no live feed of air traffic delay programs, security lines or official weather warnings. Flights elsewhere may show limited information.
Live data reaches about two days ahead. For a flight further out, he shows how that flight number usually runs. That is not the airline's schedule for your day.
He cannot see inside the airline. Crew scheduling, maintenance decisions and spare aircraft are invisible to him.
Connection advice is general. His guidance on layover times uses general estimates for each airport, not each airline's official minimums.
Europe and UK compensation. For flights covered by EU or UK passenger rules, Milo estimates the cash you may be owed from the flight's planned distance. The law uses the straight-line distance and airlines can refuse when the cause was outside their control, so treat his figure as a starting point. He decides which rules apply from the departure airport and the airline's code, which can be wrong on flights operated by a partner airline.
Rights and refunds. What he says about refunds and airline promises is general information, not legal advice. The airline promises come from the DOT's published list as of December 2024 and airlines change them.

His track record

Milo logs each prediction and what happened to the flight. You can see the results, good and bad, on the accuracy page.

Your information

What Milo remembers about you, who else sees it and how to erase it is on the privacy page.

Who makes Milo

Milo is made by Contrail, an aviation news site. He is free to use. Airlines and airports do not pay to appear or to change what he says. Milo links to Contrail's stories because Contrail makes him.