Back office ops · Production

Dropbox smart move: ML-assisted file organization with human-in-the-loop review

The problem

Dropbox users, especially team managers and company administrators, had to move files one at a time on the web, making large-scale folder organization tedious and daunting. File organization is highly personal and varies widely across users, making automation difficult to scope.

First attempt

An initial prototype repurposed the existing 'suggested destinations' model, but internal testers found its results non-deterministic—changing based on recent user navigation rather than file/folder name relationships—and the model did not meet expectations for how suggestions should relate to filenames.

Workflow diagram · grounded in source
1
User triggers smart move
trigger
“you can drop a bunch of unorganized files into your home directory with existing folders, and smart move will try to place the files you added in the correct subfolders”
2
Encode filenames and folders
ai_action
“we tokenized the name of the file being moved (we also call this the context), candidate folder names, and names of potential sibling files and folders. The tokenized names were then passed to an in-house encoder we developed which used …”
3
Score and rank folder candidates
ai_action
“The model produced a score for each file to move/candidate folder pair. We ranked each candidate destination for a file based on this score, and the top ranked candidate was considered the suggested sub-folder recommendation for that file”
4
Confidence-tier filtering
validation
“only the top 20% of recommendations by score distribution were considered high-confidence and displayed to the user most prominently. Medium-confidence suggestions were less front-and-center, but still presented for users to select in th…”
5
Human reviews suggestions
human_review
“still presented for users to select in the human-in-the-loop review screen before any file moves are made”
6
Files moved in one click
output
“Smart move can move multiple files at a time, all with one click, reducing tedious work about work”
Reported outcome

Smart move launched in November 2021.
In online alpha testing, 61% of heuristic suggestions were accepted overall and 94% of high-confidence heuristic suggestions were accepted. The trained model reached 73% offline accuracy versus 64% for the heuristic baseline, and the model was reused for additional Dropbox feature prototypes.

Reported metrics
Trained model offline accuracy73%
Similarity heuristic offline accuracy64%
Evaluation dataset size57,921 files
Heuristic suggestion acceptance rate (online alpha, overall)61%
Show all 8 reported metrics
trained model offline accuracy73%
similarity heuristic offline accuracy64%
evaluation dataset size57,921 files
heuristic suggestion acceptance rate (online alpha, overall)61%
model suggestion acceptance rate (online alpha, overall)59%
heuristic high-confidence suggestion acceptance rate (online alpha)94%
model high-confidence suggestion acceptance rate (online alpha)90%
training examples from internal dataseveral million suitable training examples
Reported stack
GloVe
Source
https://dropbox.tech/machine-learning/smart-move-ml-ai-file-organization-automation
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Smart move launched in November 2021.

What tools did this team use?

GloVe.

What results were reported?

Trained model offline accuracy: 73%; Similarity heuristic offline accuracy: 64%; Evaluation dataset size: 57,921 files; Heuristic suggestion acceptance rate (online alpha, overall): 61% (source-reported, not independently verified).

What failed first in this deployment?

An initial prototype repurposed the existing 'suggested destinations' model, but internal testers found its results non-deterministic—changing based on recent user navigation rather than file/folder name relationships…

How is this back office ops AI workflow structured?

User triggers smart move → Encode filenames and folders → Score and rank folder candidates → Confidence-tier filtering → Human reviews suggestions → Files moved in one click.