The 2027 Truth Report.

Why the failures are at the top

A transparency report where the good numbers come first is a marketing document with a table in it.

The credibility of everything below depends on you believing we would have published a bad number in the same place as a good one. So the bad ones are printed first, in full, with the dates.

Everything here is measured from production systems rather than from a survey. Where a figure is an estimate we say so. Where we do not have a figure we say that instead of producing one.

Promises we did not keep

Five entries. Two are engineering failures, two are planning failures, and one is a claim we should never have made in the first place.

Failure one, we promised 200 destinations and delivered 145

We said 200 plus destinations at launch. We ended the year with 145 live.

The gap is 55 destinations, and it is not a rounding error or a definitional trick about what counts as a destination. We counted the destinations our wholesale agreements would eventually reach, and then said that number out loud before the agreements were signed.

Thirty one of the missing 55 are in Africa and the Pacific, where a single carrier holds the only viable network and the commercial negotiation ran past our schedule. Fourteen are in markets with regulatory registration requirements we underestimated. The remaining ten we simply deprioritised.

What we should have said was 145 live and 200 contracted. We are saying it that way from now on, and the number on the homepage is the live one.

Failure two, we missed the support target in two months

Our published target is a human reply in under 60 seconds, 24 hours a day. Across the year we hit it on 91.2 percent of conversations, with a median of 38 seconds.

In two months we missed badly. In July, 74.1 percent were under 60 seconds. In December, 79.6 percent.

July was a capacity failure. European summer travel volume rose faster than our roster, and we were short by roughly four people for five weeks. That is a planning error and it was entirely predictable.

December was worse, because it was self inflicted. We deployed a change to the assistant handoff on 18 December that silently dropped the escalation keyword in three languages. It took us 31 hours to notice. During those 31 hours, travellers asking for a human in German, Portuguese and Korean were answered by software that would not step aside.

The fix shipped on 20 December. The monitoring that should have caught it in minutes shipped on 4 January.

Failure three, nine countries below the local market median

We publish measured throughput per country, and we compare it to the local market median. In 9 of the 145 destinations we came in below it.

Below the local market median means a traveller on a local prepaid SIM would have had a better connection than one on Orbislo. That is the definition we use and it is the harshest one available.

Six of the nine share a cause. We hold a single carrier agreement in those markets and it is not the strongest network in the country. In Nepal and Myanmar we are on the second network. In Papua New Guinea, Vanuatu, Tonga and Samoa there is effectively one usable network and we are on a resold path with an extra hop.

The other three are our own fault. In Ethiopia and Madagascar our carrier priority list was stale for most of the year. In Guyana we were routing through a partner platform that added latency we did not measure until August.

The table below is the whole list. It is also the list we hand to journalists when they ask for our worst results.

Countries where our measured speed came in below the local market median

Nine countries out of 145. In the other 136 we met or exceeded the local market median. We publish the losing nine here and on each of their own coverage pages, because a traveller going to Nepal deserves the number before they buy rather than after.

Failure four, the automatic refund missed 214 orders

We promise that a provisioning failure refunds itself inside 60 seconds with nobody asked to file anything. For six weeks, between 3 September and 15 October, it did not.

A change to the provisioning state machine introduced a terminal state called abandoned, which the refund job did not recognise as a failure. Orders that reached it were left charged and unrefunded.

214 orders were affected, worth $4,911. Nobody was told, because the whole point of an automatic refund is that nobody has to notice.

We found it because a customer wrote in about a $20.89 charge on 14 October. We refunded all 214 on 16 October, added $10 in wallet credit to each account as an apology, and wrote to every affected traveller.

The refund job now fails closed. Any terminal state it does not recognise is treated as a failure and refunded, and an unrecognised state raises an alert.

The numbers that went well

Provisioning success rate

98.7 percent of orders provisioned successfully on the first attempt. After automatic retry and carrier failover, 99.94 percent succeeded.

That leaves 0.06 percent, which is 1 order in 1,667, that never provisioned at all. Every one of those was refunded automatically, apart from the 214 described above.

The first attempt figure is the one worth watching, because a retry costs a traveller between 40 and 90 seconds at exactly the moment they are least patient. It improved from 97.9 percent in January to 99.2 percent in December.

The largest single cause of a first attempt failure was a device that had reached its stored profile limit, which is a phone problem rather than a carrier problem, and which we now detect before taking payment.

Median install time

41 seconds, measured from the tap on the install button to the profile attaching to a network.

The ninetieth percentile was 96 seconds, and the ninety ninth was 4 minutes 12 seconds. The long tail is almost entirely iOS devices that require a restart before a newly downloaded profile will attach.

Android was faster than iOS by 6 seconds at the median, which surprised us and reverses the previous year.

The slowest measured installs are in countries where the initial attach has to traverse a distant home network element. In Fiji the median install was 2 minutes 51 seconds, which is not acceptable and is on the 2027 roadmap.

Refunds

2.4 percent of orders were refunded during the year. That number is higher than the category norm, and we think a low refund rate is usually a sign of a hard refund process rather than a good product.

Of the 2.4 percent, 0.9 percent were provisioning failures, which are our fault. 0.8 percent were travellers who bought the wrong plan, most often a country pack for a country they were transiting rather than visiting.

0.5 percent were device incompatibility discovered after purchase, which is a failure of our device check and is falling as that check gets stricter.

The remaining 0.2 percent were goodwill refunds where support judged the traveller had a fair point and the policy did not cover it. We do not require agents to escalate those, and we do not track a budget for them.

Support response time

Median time to a human reply was 38 seconds. 91.2 percent of conversations reached a human inside 60 seconds.

The assistant resolved 74.3 percent of conversations without a handoff, which is the figure that makes the 60 second target affordable at all.

Median time to full resolution, not just first reply, was 11 minutes. For anything requiring a carrier side change it was 4 hours 20 minutes, and that number has not improved all year.

Satisfaction was 4.6 out of 5 across 41,208 rated conversations. The lowest rated category was device compatibility, at 3.9, where the honest answer is often that a traveller's phone cannot do what they need.

How this report is compiled

Every figure comes from production telemetry, queried on 8 January 2027 for the period from 1 January to 31 December 2026.

Throughput figures come from consented measurements on real traveller traffic, not synthetic speed tests, because a speed test burns a traveller's data to produce a number that flatters us.

Local market medians for the country comparison come from third party mobile measurement published for the same period.

No figure in this report has been rounded in our favour. Where a number could be presented two ways, we used the less flattering one, and the queries that produced them are kept so next year's report can be compared to this one.

Questions about this report