---
title: "AI booking import benchmark: 57 test documents scored"
description: "Viaondo’s own test of turning booking confirmations into an itinerary: 57 fictional documents, 11 trips, 10 languages. Full tables and every miss we can count."
url: "https://viaondo.com/blog/booking-import-benchmark"
language: "en"
updated: "2026-11-17"
---

# We ran 57 fictional travel documents through our AI import: what it got right and what it missed

> Viaondo’s own test of turning booking confirmations into an itinerary: 57 fictional documents, 11 trips, 10 languages. Full tables and every miss we can count.

Canonical: https://viaondo.com/blog/booking-import-benchmark

Last updated: 2026-11-17

- By Pablo Bravo: https://viaondo.com/blog/author/pablo-bravo
- Updated 17 November 2026
- 13 min read
- Product: https://viaondo.com/blog/category/product

## Key takeaways

- On 57 fictional documents from 11 trips in 10 languages, Viaondo’s AI import found all 104 expected bookings and added 3 extra rows that a person would have to check or remove.
- It matched 525 of 527 plan steps and 82 of 83 fixed-time steps, so 1 of the 11 trips failed our script’s criteria. 272 of 335 places got an exact map pin, and 63 did not.
- This is a best case, not a blind test: Viaondo built these documents and tuned the import on them, and 19 of that day’s 52 runs failed our script, nearly all before the last fix. Expect worse on documents it has never seen.
- AI output varies by run: on the same import code, the same cruise documents passed with 67 of 68 steps in one run and failed with 66 of 68 in another.
- The import makes a draft. Nothing reaches the client until the agent reviews it and publishes.

## Findings you can cite

- In Viaondo’s own best-case test, on 57 fictional documents the import was tuned on, its AI import found all 104 expected bookings and added 3 extra rows a person would have to check. ([Link to this finding](https://viaondo.com/blog/booking-import-benchmark#finding-bookings-found))
- In Viaondo’s own best-case test, on 57 fictional documents the import was tuned on, its AI import matched 525 of 527 plan steps (99.6%). ([Link to this finding](https://viaondo.com/blog/booking-import-benchmark#finding-steps-matched))
- In Viaondo’s own best-case test, on 57 fictional documents the import was tuned on, its AI import matched 82 of 83 fixed-time steps (98.8%), the ones that can’t move, such as a port arrival. ([Link to this finding](https://viaondo.com/blog/booking-import-benchmark#finding-fixed-time-steps))
- In Viaondo’s own best-case test, on 57 fictional documents the import was tuned on, 272 of 335 places (81.2%) got an exact map pin and 63 did not. ([Link to this finding](https://viaondo.com/blog/booking-import-benchmark#finding-exact-map-pins))
- In Viaondo’s own test, 19 of 52 completed AI import runs in one day (36.5%) failed its scoring script, including 1 of the 14 runs on the final import code, so results vary from run to run. ([Link to this finding](https://viaondo.com/blog/booking-import-benchmark#finding-failed-runs))

How to cite: Viaondo, AI booking import benchmark, measured on 5 November 2026 (own test on fictional documents), https://viaondo.com/blog/booking-import-benchmark

**Short answer:** you can turn a booking confirmation into an itinerary automatically, as a draft that someone still has to check. We ran 57 fictional documents from 11 trips through Viaondo’s AI import on 5 November 2026. It found all 104 expected bookings and matched 525 of 527 plan steps. It also added 3 rows the answer key didn’t have, showed one cruise disembarkation at 08:00 where the documents print 09:00, and left 63 of 335 places without an exact map pin.

Before you read those numbers, read what they are:

- **These are fictional test documents built by Viaondo, not customer data.**
- **The same documents were used to tune the import.** This is a best case and not a blind test. Expect worse on documents it has never seen.
- **We found a bug in our scoring script, and fixing it improved the result.** The correction, further down, gives the numbers from before and after.
- **It is Viaondo’s own measurement.** [We make the product](https://viaondo.com/about), and no outside party ran or checked this.
- **These results describe the import as of 5 November 2026.**
- **Nothing goes to the client until the agent reviews and publishes.** The import fills a draft.

## Can booking confirmation PDFs become a client itinerary automatically?

Yes, as a draft to review, and several tools offer it. [Viaondo’s booking import](https://viaondo.com/product/booking-import) reads PDFs, Word files, screenshots and emails. [mTrip](https://viaondo.com/compare/mtrip) [describes an import wizard](https://www.mtrip.com/en/resources/ai-itineraries-practical-guide/) for GDS, mid-office and supplier documents, and [TripCreator](https://viaondo.com/compare/tripcreator) [takes a link or a PDF](https://tripcreator.com/blog/import-with-ai).

[mTrip’s own guide](https://www.mtrip.com/en/resources/ai-itineraries-practical-guide/) puts the limit well: “AI does not replace the agent’s review.” We have no accuracy data for any other tool, so this post compares nothing. It only reports ours.

## What we tested: 11 trips, 57 documents, 10 languages

The test set is 11 trips with 57 documents: 37 text PDFs, 14 emails, 3 phone screenshots, 2 Word files and 1 WhatsApp chat export. The trips include a cruise, a safari dated only by day numbers, a group tour with a rooming list and a congress with a schedule change. None of the 57 is scanned paper or a raw GDS export, so this test says nothing about those.

Each trip has an answer key: the bookings a careful person would extract, once each, and the plan’s days, steps and places. A script compares the trip the AI made with that key. A trip passes only if every expected booking is found once, at least 95% of steps and every fixed-time step match, and the other checks stay clean: places, descriptions, repeated steps, and the trip’s language and logic. We used the latest completed run of each trip, whatever its score. All 11 of those runs used the same import code.

## Results per trip: bookings and steps

On these test documents, which the import was tuned on, Viaondo found all 104 expected bookings, added 3 extras and missed 2 of 527 steps. One of those steps had a fixed time, so the cruise failed the script’s criteria, and the other 10 trips passed. Two trips have no plan in steps, so they have no step score. On the Balkans tour, a program in prose, all 7 days kept their words and no step was invented. Fixed-time steps are the ones that can’t move, such as a port arrival or an all-aboard time.

| Trip | Language | Documents | Bookings found | Extra bookings | Steps | Fixed-time steps |
| --- | --- | --- | --- | --- | --- | --- |
| Balkans group tour | Turkish | 5 | 7/7 | 0 | n/a | n/a |
| Baltic ferry and rail | Swedish | 5 | 8/8 | 0 | 52/52 | 9/9 |
| Barcelona congress | German | 6 | 12/12 | 0 | 48/48 | 10/10 |
| Kenya safari | Dutch | 5 | 14/14 | 0 | 65/65 | 11/11 |
| Lisbon and Algarve family | Portuguese (Brazil) | 6 | 9/9 | 0 | 72/72 | 7/7 |
| Maldives honeymoon | Polish | 5 | 9/9 | 0 | n/a | n/a |
| Mediterranean cruise | Italian | 4 | 9/9 | 2 | 66/68 | 16/17 |
| Pacific Northwest road trip | English (US) | 5 | 13/13 | 0 | 71/71 | 9/9 |
| Paris and Loire family | French | 6 | 10/10 | 0 | 55/55 | 9/9 |
| Rome city break | Spanish | 5 | 7/7 | 0 | 49/49 | 6/6 |
| Scotland by rail | English (UK) | 5 | 6/6 | 1 | 47/47 | 5/5 |
| **Total** | **10 languages** | **57** | **104/104** | **3** | **525/527** | **82/83** |

## Results per trip: places, descriptions and notes

In the same test, every place the plans link appeared on its day, but only 272 of the 335 places in the trips got an exact pin. “On its day” means the place the document names is listed on the right day with its Google Maps link. “Exact pin” means the import found exact coordinates for it. These are two different tests, with different totals. A description is wrong when a day loses the text its document prints, or gets unmarked AI text the document doesn’t print.

| Trip | Places on their day | Exact pin | Descriptions wrong | To check notes | Run (UTC, 5 Nov) |
| --- | --- | --- | --- | --- | --- |
| Balkans group tour | n/a | 5/8 | n/a | 0 | 11:28 |
| Baltic ferry and rail | 30/30 | 20/34 | 0 | 1 | 11:28 |
| Barcelona congress | 22/22 | 11/17 | 0 | 2 | 11:28 |
| Kenya safari | 34/34 | 21/25 | 0 | 3 | 11:59 |
| Lisbon and Algarve family | 45/45 | 36/40 | 0 | 1 | 11:59 |
| Maldives honeymoon | n/a | 4/5 | n/a | 2 | 11:59 |
| Mediterranean cruise | 32/32 | 38/47 | 0 | 2 | 16:42 |
| Pacific Northwest road trip | 53/53 | 39/49 | 0 | 4 | 17:04 |
| Paris and Loire family | 34/34 | 35/38 | 0 | 0 | 17:20 |
| Rome city break | 36/36 | 34/36 | 0 | 0 | 16:10 |
| Scotland by rail | 34/34 | 29/36 | 0 | 1 | 16:25 |
| **Total** | **320/320** | **272/335** | **0** | **16** | |

The arithmetic, in the order of the tables:

```text
Documents   5+5+6+5+6+5+4+5+6+5+5             = 57
Bookings    7+8+12+14+9+9+9+13+10+7+6         = 104 of 104
Extras      0+0+0+0+0+0+2+0+0+0+1             = 3
Steps       52+48+65+72+66+71+55+49+47        = 525
  of        52+48+65+72+68+71+55+49+47        = 527 (99.6%)
Fixed-time  9+10+11+7+16+9+9+6+5              = 82
  of        9+10+11+7+17+9+9+6+5              = 83
On its day  30+22+34+45+32+53+34+36+34        = 320 of 320
Exact pin   5+20+11+21+36+4+38+39+35+34+29    = 272
  of        8+34+17+25+40+5+47+49+38+36+36    = 335 (81.2%)
To check    0+1+2+3+1+2+2+4+0+0+1             = 16
```

### A correction that favors us

Our first scoring of these same runs showed 91 of 91 bookings and 16 extras. When we checked each extra against the documents, we found a bug in the script: it treated separate bookings in the answer key as one when they shared a reference and a date, or a reference alone for hotels, or a date and a similar title. So a safari’s stays on one DMC file counted as one booking. 13 of the 16 extras were real bookings in the key, found in the trip, that the script had never scored.

We fixed the script and rescored all 52 saved runs. No import was rerun for this, and the fix changed no answer key. The fix improves our numbers, so we’re giving you both versions.

## What the score doesn’t check

A trip can score well here with a wrong detail inside it. A booking counts as found when its type and date match, along with its flight number, or its time (within 20 minutes) and either its reference or enough of its title. A wrong reference or a wrong end time on a found booking is logged but doesn’t lower the score. Seats and passenger names aren’t scored.

A step counts when its start is within 5 minutes and about a third of its title’s words match, with looser rules for arrivals and hotel check-ins. A pin counts when the place has coordinates marked exact, and the test doesn’t check they are in the right spot. Extra bookings don’t fail a trip either: Scotland passed with one.

Two of the 104 are bookings a later document cancels, a flight in Barcelona and a massage in the Maldives, and those count as right when they are gone or marked canceled. The 104 also include 4 cruise port calls and items still pending. So “104 of 104” means the other 102 expected bookings were found, and it doesn’t mean every field of every booking was right.

Here is what that looks like. We noticed these while checking the extras, and we didn’t audit every field, so there are probably more:

- **Kenya:** four stays show a 15:00 check-in that no document prints. The program gives about 22:30, 12:00, 14:00 and 11:30.
- **Kenya:** an overland transfer shows one hour, and the DMC prints about five. Other transfers in the test get the same one-hour end time.
- **Balkans:** the Lapad Bay stay keeps the rooming list’s reference and shows 15:00, and the hotel’s own confirmation gives another reference and 14:00.

An agent should still check times and references against the source.

## What changed in the import after this test

On 7 November 2026 we shipped import changes related to details like the ones above. We don’t know yet whether they fix any miss on this page: only a rerun will tell. Every figure on this page comes from the import code of 5 November, before those changes, and as of 11 November 2026 we haven’t rerun the test on the new code. When we do, this page gets the new numbers and keeps the old ones beside them.

- A stay with no printed check-in or check-out time no longer shows the usual time as if a document printed it. Travelers see the day only, and the agent sees the stay marked as a usual time, not printed.
- A transfer, tour or dinner with no printed end no longer gets an invented end time.
- When two readings of the same stay disagree, the printed time replaces the usual one, and the hotel’s own reference replaces a group reference shared by several stays.
- A fixed step that falls later inside a longer booking keeps its own time, unless the titles agree.

## The misses, and what kinds they are

**3 extra rows, in 2 of the 11 trips.** That is about one for every 35 expected bookings. We read each against the documents, and none is invented. Two are on the cruise: the disembarkation and the embarkation slot, both real items that the answer key treats as plan steps. The third is in Scotland: “Highland Sleeper departs”, which repeats the sleeper train already booked in that trip.

**2 missed steps out of 527, both on the cruise.** The documents print “09:00 Sbarco” (disembarkation) for this family’s blue luggage tags. The import folded that step into the extra disembarkation row and gave it 08:00, the ship’s arrival time. The row’s note still says 09:00, but its headline time is wrong, and a client would see it unless the agent fixed it. That missed fixed time is why the cruise fails. The other miss is an 08:40 meeting point at Milano Centrale: its advice moved onto the train booking and its time was lost.

**63 places without an exact pin, out of 335.** Baltic (14), Pacific Northwest (10), the cruise (9), Scotland (7) and Barcelona (6) hold 46 of them. In the [traveler app](https://viaondo.com/product/traveler-app), a place with only its city’s location gets no pin on the map and opens the maps app by its name. We read all 63 against the documents, which is a judgment and not a measurement. 29 look like real places the import could have found, 12 of them in the Baltic trip, so those misses are the import’s. 13 have invented names and no real location to find. The other 21 are unclear, and 19 of those are records with no name, which inflate the total.

**16 notes under To check.** These aren’t errors. They are open questions the import left for the agent, such as a pickup time a voucher doesn’t give. The test counts them and doesn’t score whether they were the right questions.

## Same documents, different results

AI output varies from run to run, even on identical documents. We completed 52 runs of these 11 trips on 5 November, and 19 failed the script’s criteria. All but one of those failures came before 07:21 UTC, while we were using these same documents to find and fix import bugs. That is what “tuned on them” means here, and it is why the table below looks worse than the results above. Eight more attempts stopped before any import began, because our test account had reached its limit of three agencies, so they aren’t counted.

| Trip | Completed runs | Failed | Bookings found | Extra bookings | Steps | Exact pin, lowest to highest share |
| --- | --- | --- | --- | --- | --- | --- |
| Pacific Northwest road trip | 8 | 5 | 11 to 13 of 13 | 0 to 2 | 70 to 71 of 71 | 38/49 to 40/51 |
| Barcelona congress | 7 | 5 | 10 to 12 of 12 | 0 to 3 | 47 to 48 of 48 | 10/17 to 13/18 |
| Mediterranean cruise | 5 | 3 | 6 to 9 of 9 | 1 to 3 | 65 to 68 of 68 | 25/41 to 38/44 |
| Lisbon and Algarve family | 5 | 2 | 9/9 | 0 to 1 | 66 to 72 of 72 | 36/40 to 38/40 |
| Kenya safari | 5 | 1 | 14/14 | 0 to 1 | 48 to 65 of 65 | 16/23 to 21/25 |
| Rome city break | 5 | 1 | 7/7 | 0 | 49/49 | 34/36 |
| Balkans group tour | 4 | 1 | 7/7 | 0 | n/a | 9/20 to 5/8 |
| Scotland by rail | 4 | 1 | 6/6 | 0 to 1 | 46 to 47 of 47 | 29/36 to 31/38 |
| Baltic ferry and rail | 3 | 0 | 8/8 | 0 to 1 | 52/52 | 20/34 to 21/34 |
| Maldives honeymoon | 3 | 0 | 9/9 | 0 | n/a | 3/5 to 4/5 |
| Paris and Loire family | 3 | 0 | 10/10 | 0 | 55/55 | 33/37 to 35/38 |
| **Total** | **52** | **19** | | | | |

The place totals in the pin column change because they count the places each run’s trip ended up with, and that set differs between runs: the Balkans tour held 20 in one run and 8 in another. We haven’t compared which places came or went.

The results above use the latest run per trip, picked by time and not by score. As far as our edit history shows, the import code didn’t change from 07:21 UTC on, and 14 runs used it: one each of 8 trips and two each of 3 trips. 13 of the 14 passed. Only those three repeats measure chance alone, and they are too few to give you a rate. The cruise passed with 67 of 68 steps at 07:21 and failed with 66 of 68 at 16:42. The road trip got 13 of 13 bookings both times, with pins of 40/51 and then 39/49. The Balkans tour got 7 of 7 both times, with pins of 6/11 and then 5/8.

Each run took between 11 and 45 minutes, but this was a queued test harness run on a local machine, so we make no claim about speed from it.

## What this means if you are choosing a tool

Expect to spend your time on review rather than retyping, and check three things on every import: extra bookings, times and references against the source, and places without a pin.

Viaondo’s import may not suit you if your bookings live in a GDS or a mid-office system, because it reads the documents you give it and doesn’t connect to those. It may also fall short if you import more than the 300 documents a month the plan includes. Viaondo costs US$50 a month or US$500 a year per agency, plus VAT where it applies, whole team included: see [pricing](https://viaondo.com/pricing), and [our price index](https://viaondo.com/blog/itinerary-software-price-index-2026) for how that compares with per-seat and per-passenger plans.

The fairest test is your own: run a real trip whose answer you already know, count the misses as we did here, and compare.

## Sources

- [mTrip: How travel agencies use AI to build itineraries, a practical guide](https://www.mtrip.com/en/resources/ai-itineraries-practical-guide/) (accessed 5 November 2026)
- [TripCreator: Import with AI](https://tripcreator.com/blog/import-with-ai) (accessed 5 November 2026)

## Frequently asked questions

### Can I turn booking confirmation PDFs into a client itinerary automatically?

You can get an automatic draft, and you should not send it unread. In Viaondo’s own test on 57 fictional documents, the import found all 104 expected bookings but also added 3 extra rows, showed one disembarkation an hour early, and 63 of 335 places had no exact map pin. Those documents were used to tune the import, so expect worse on new ones.

### How accurate is AI at reading booking confirmations?

On documents it was tuned on, Viaondo’s import matched 104 of 104 bookings and 525 of 527 plan steps, with 3 extra rows. That is a best case from Viaondo’s own measurement, with no outside party. A booking can count as found with a wrong time inside it. It says little about accuracy on documents the import has never seen, and nothing about other tools.

### Were real customer documents used?

No. Every document is fictional and built by Viaondo for testing. People, references, prices and most hotels and suppliers are invented. Real places and big carriers keep their real names.

### Does the AI send anything to my client?

No. In Viaondo the import writes to a draft. Nothing reaches the client until an agent reviews the trip and publishes it.
