SFO Opens $300M Automated Air Cargo Terminal for E-commerce Growth

The Analysis

Three hundred million dollars, opened on 17 June, built around automated storage and retrieval and high-throughput systems, and aimed at e-commerce volumes. That is what the report says about the new air cargo terminal at San Francisco International Airport, and I read it the way I read every terminal story, which is from the paperwork side. A building is not what changes my week. What changes my week is which record the machine believes, because in an automated facility the system's record is the physical truth, and if the record is wrong the goods are wrong, and no amount of concrete fixes that.

Here is the difference in plain terms. In a conventional shed, if a handler wants to find your ten cartons, somebody walks over and looks. In a facility with automated storage and retrieval, the machine decides which location a unit sits in, and getting it back out is a request you file. Two things follow from that, and both of them are document problems. The system only knows what was keyed in at induction, and the only way to lay eyes on the goods, for an exam, a sample or a photograph, is to get the system to produce them, which is an appointment in a queue.

So below I take one cross-border e-commerce consignment and walk it through: master air waybill, house air waybills, advance data, entry, HS code and duty treatment, exam appointment, release. I show where it jams and why a shorter dwell time leaves you less room to fix a mismatch rather than more. One calculation, with the assumptions written out. And one handover: Dr. Ingrid Voss has walked through the scope of this facility's customs mandate; here I only deal with what happens to the shipment.

Start by being exact about what was reported, because everything below depends on not adding to it. San Francisco International Airport unveiled a new $300 million air cargo terminal on 17 June. It is designed with automated storage, retrieval and high-throughput systems, and it is meant to handle growing e-commerce volumes. It is expected to reduce dwell times and improve customs clearance efficiency. It reflects a broader trend of airports investing in cargo infrastructure to capture express and cross-border e-commerce demand. That is the whole of the report. No published dwell-time figure, no throughput tonnage, no clearance-time target, no numbers about how many cartons an hour the automation moves. Every number that follows that is not in the report is mine, and I will say so each time I use one.

Now the change that matters to you. In a conventional terminal, the physical world and the paper world can disagree for a while without anybody noticing, because a human being reconciles them by walking over and counting. In an automated facility the reconciliation happens at induction, by a scanner and a weight check, and from that moment the system's count is the truth that every downstream step works from. If the piece count keyed at induction is 118 and your entry says 120, the machine will happily store a phantom and happily fail to produce it later, and the time you would once have spent hunting for a carton you now spend arguing about a number. This is why I say the alignment problem moved earlier, not away. It did not get easier because the building got newer.

Walk one consignment through and you can see where the jams are. Overseas consolidation: cartons are packed and palletised, and the master air waybill is issued by the carrier or the forwarder while house air waybills are issued underneath it by the forwarder or by the sellers whose goods are in the box. Advance cargo data goes in before or at departure. The flight lands, the unit load devices are broken down, cartons are inducted into the automated store, and the system records a piece count and a weight. The entry goes in: start with the HS code, then look at duty treatment, because the classification is what everything downstream hangs from. Then one of two things happens. Release, and the goods are retrieved and trucked. Or an exam, and now the goods have to be produced from the machine into an exam area, which is an appointment.

The ownership of the data at the consolidation end is worth a paragraph of its own, because that is where most of my mismatches are born. The master is produced by the carrier or the forwarder. The houses underneath it are frequently produced by the sellers whose goods are in the box, or by a platform that aggregates them, which means the party cutting the master is often not the party that knows how many cartons are actually on the pallet. Then add a repack at the consolidation warehouse, which happens constantly in e-commerce because somebody decides to combine two parcels into one carton to save on the last mile: the count changes after the house has already been cut.

So the rule I would write down is simple and unglamorous. Nobody cuts a master until the last repack is finished, and any repack after a house is cut requires that house to be re-cut the same day. One sentence, and on my own experience it removes most of the three per cent.

Four alignment points, and I would put a person against each one. Master against house: the master piece count has to equal the sum of the house piece counts, and the same for weight, and if one house is wrong the whole consignment holds, because the machine does not release a part of a unit it cannot account for. Declared pieces against scanned pieces: the automated store counts what it scanned, so a one-carton difference between the entry and the scan triggers a reconciliation before anything moves. HS code against what is actually in the box: on e-commerce, one carton frequently holds twenty SKUs, and the question of which code the lot is entered under has to be settled before departure, not at the airport. And exam appointment against the time you have left: which brings me to the arithmetic.

Here is the calculation, and the assumptions go first because the report does not give me the two numbers I need. Assume you clear 300 house air waybills a week at SFO, consolidated into 20 masters. Assume your current mismatch rate between master and house piece counts is 3%, which is a baseline you should replace with an extract from your own system, since that extract is the only honest input here. Three per cent of 300 is nine mismatches a week. Assume each correction, with complete paperwork, takes four working hours: supplement the documents, amend the entry, wait for the re-check. Assume the report's direction on dwell time lands as a fall from 36 hours to 20 hours. That 36 and that 20 are my assumptions and not the report's figures; the report only says dwell times are expected to fall. So the window in which you can fix a mismatch shrinks from 36 hours to 20 hours, a loss of 16 hours.

Put the two together and you can see why this particular change bites. Nine mismatches at four hours each is 36 hours of correction work, which just about fits inside the old 36-hour window. That is why the process works today, and it works with nothing to spare. In a 20-hour window the same nine corrections need 36 hours and you have 20, so you are 16 hours short every week, and the shortage shows up as held freight rather than as a problem anybody scheduled. There are only two ways out. Cut the mismatch rate, or move the correction earlier so that less of it has to happen inside the window. Both are worth pricing.

Cutting it first. Assume each house is worth $2,500, and assume a held consignment costs you $60 a day in storage, customer service and refund preparation. Both are my assumptions. Nine mismatches a week at $60 is $540 a week, and over 52 weeks that is $28,080 a year. Drive the mismatch rate to 0.5% and you have 1.5 mismatches a week, which is $90 a week and $4,680 a year. The difference is about $23,400 a year, and you obtain it with a reconciliation step rather than with a facility.

Moving it earlier, which is the part I would actually do first. Assume you run a master-versus-house reconciliation 24 hours before departure and assume it catches 70% of the mismatches before the aircraft leaves. Both assumptions, and the 70% is the one to test in your first month rather than believe. That leaves 2.7 mismatches a week arriving at the airport: 2.7 times four hours is 10.8 hours of correction work against a 20-hour window, and the window now has room in it. Put a trigger on it as well, because a window you do not measure is a window you will overrun: at 80% of the window consumed, escalate the same day. On a 20-hour window that is hour 16, and on my numbers you would hit it only in a bad week.

Timing is the other half of the same problem, and it is worth writing as a sequence, because the sequence is where people slip. The house data has to exist before the master can be right. The master has to be right before the advance filing can be right. The advance filing has to be right before the aircraft leaves. Run it backwards from departure and the discipline shows itself: assume advance data is due four hours before wheels up, assume your reconciliation takes two hours, assume consolidation closes six hours before departure, and you have four hours of slack, which you should spend rather than celebrate. Assume nothing about cut-offs at any particular station. Every port I have worked has its own, and the published one is not always the enforced one. Ask what the enforced cut-off is, get it in writing, and then set your internal deadline a full working shift earlier than that.

The $300 million belongs in your thinking too, though not as a cost you pay. Assume twenty years of depreciation, which is my assumption and not the airport's accounting: $300 million over twenty years is $15 million a year, about $288,000 a week of fixed cost that has to be carried by throughput. A facility carrying that has a structural reason to keep freight moving, and lower dwell is therefore a design target rather than a favour to shippers. Read the report's sentence in that light. Dwell times falling is the facility doing what it was built to do. The consequence for you is that slow freight will be the freight the system wants out, and the cost of being slow lands on whoever did not do their paperwork, which is not a legal conclusion, just an observation about which way the queues run.

On the e-commerce declaration question I will go as far as I am qualified to and no further. One carton holding twenty SKUs is a classification decision before it is anything else, and my working rule is unchanged: start with the HS code, then look at duty treatment. What I would insist on is consistency. Pick a treatment for a given SKU combination, write it down, and use the same treatment every time, because when a question is raised months later, the speed of the answer depends on whether your own history agrees with itself. What the correct treatment is for any particular combination is exactly the kind of thing I will not improvise. Take it to your broker or your compliance adviser and get it in writing before the goods move.

Then get the exam process in writing, because it changed shape. Goods in an automated store are not in a location you can describe to a driver; they are in a system location, and producing them is a transaction with a queue behind it. Ask the terminal and your handler for three numbers: how much notice an exam appointment needs, how long retrieval to the exam area takes, and how long the return to store takes after the exam. I would not accept a verbal answer on any of the three. As working targets I would push for an appointment obtainable inside 24 hours and retrieval inside four hours, and if the actual figures are worse, that is a planning input, not a complaint. Ask one more question while you are there: what happens to pieces that exceed the automation's handling limits on size or weight. Anything that cannot ride the automated path is handled some other way, and that other way has its own clock.

Who owns what, then, because none of this happens without a name. The pre-departure reconciliation belongs to whoever produces the house air waybills, and it happens 24 hours before the flight, not after the truck leaves. The advance data belongs to whoever files it, and it has to agree with the master to the piece. The classification belongs to your broker, and it belongs to them before departure rather than at the airport. The exam appointment belongs to whoever is on site, and that person should be booking it the moment a hold appears rather than the next morning. And the measurement belongs to whoever signs the invoice, who should be looking at a mismatch count and a window-consumption figure every week rather than at a monthly accrual.

One more thing about the trend the report names, because it changes what you should build rather than what you should do this week. Airports are putting money into cargo infrastructure to compete for express and cross-border e-commerce demand, and this facility is one instance of that pattern rather than a one-off. So the procedures you settle on here should be written to travel. If your exam-appointment playbook, your piece-count reconciliation and your unit conversion rule live in one person's head at one station, the next automated facility you meet costs you the same learning curve all over again. Write them once, as a single page naming the four alignment points, the two thresholds and the three exam numbers, and hand the same page to the next station. I have watched companies pay for the same lesson at three airports in a row because nobody wrote the first one down.

And measure the thing, because a process you do not count is a process you will lose. Three numbers, once a week, on one page: mismatches caught before departure against mismatches found after arrival, average hours to close a correction, and the worst window consumption of the week. The first tells you whether the pre-departure reconciliation is earning its keep. The second tells you whether my four-hour assumption is honest, and it should be tested in the first month rather than believed. The third tells you how close you came to the trigger. Put the review on the same day each week and give it a name. If the pre-departure step works, the first number climbs and the second stays flat, and if it does not, you change the assumption rather than the story.

There is one operational question I would settle before you need the answer, and it is about the part of the consignment that is not held. When an exam lands on one house inside a master of twenty, what happens to the other nineteen? Sometimes everything waits, sometimes the clean part can be split out and released, and which of those applies at a given station is a question with a procedure behind it, because splitting a master means producing a new set of house data and a new entry line, which is exactly the kind of work that has to happen inside the same shrinking window.

Ask the question now and get the answer in writing. If splitting is available, price it: on my numbers, one split is another four hours of correction work, so a week with three exams is twelve hours of split work sitting inside a twenty-hour window alongside everything else, and that is the week the process breaks. If splitting is not available, then one exam on one house holds all twenty, and your mismatch rate matters even more than the arithmetic above suggests. Either answer changes what you do, which is why I would rather have it in June than in the middle of a peak week.

The ways this goes wrong, all of which I have watched happen. A house amendment made without touching the master, so the two disagree by one carton and the whole consignment holds. Piece counts in the systems of two parties counted at different levels of packing, so both are right and neither matches. Weight units that do not agree, pounds on one document and kilos on another, with nobody owning the conversion. Exam appointments counted in working days while the clock on the freight runs in hours. Returned and refunded e-commerce parcels consolidated into the same lot as inbound ones, which turns one clean entry into two messy ones. And a facility handling limit nobody read, discovered at the moment a carton that cannot ride the machine has to be re-routed by hand.

Where I leave it: this consignment can be saved, but not this afternoon if you start after lunch. The new terminal does not give you more time. It gives you a faster machine and a shorter window, and the shorter window is the part you have to plan for. Do the reconciliation before the aircraft leaves rather than after it lands, write the two thresholds into the way you work, and get the exam numbers on paper while nobody needs them. The rest of the story, the part about what this facility is authorised to do and where the regulatory boundary sits, belongs to a different desk. What I care about is that on the day your twenty masters land, nine of them do not turn into nine arguments in a twenty-hour window.

↑ Back to the key points

— By Derek Xu