OpenAI transcription - unexpected bill / token usage

[PreviouslyAsterisk]

Hey,
First, I'd like to say how impressed I am with the OpenAI integration - the transcript is great, but the summary is a feature I never knew I needed, and cannot now live without. It enables me to quickly see the nature of calls handled without needing to bother people. Idea: send a daily email with the summary and number of every call!

Now, the problem: The charges are as expected, however, I received a larger bill a few nights ago, in the early hours of the morning. This is for hundreds of tokens, in a very short space of time, rather than the dozens of tokens we use over a longer spread throughout the working day.

I can see a few events that look suspicious, so let me talk you through the events.

1. We receive an alert at 0200 on 4th August for an excessive and unusual token usage.





2. However, there are no calls noted at this time (Sunday morning). The last calls were made a few days ago, on the Friday:



3. Around the same time of the excessive token usage, I saw the transcription service flap:



4. Curiously, and I do not know if this is related, but there is a recording archive job at 0115 on Sundays to move all the call recordings to SharePoint.



Now, this is the first week we've been using OpenAI, and also the first week we've been archiving calls to SharePoint. But, the timeline does rather look like SharePoint has somehow caused a problem - and I have no idea how to investigate further. I wonder if something to do with the archiving could be running the call recordings through the transcription service over and over?

Anyway - I would be grateful if someone could help me investigate and suggest a resolution. Perhaps someone else has seen this?

Thanks in anticipation!

 

[TheodorosG_3CX]

Have you tried to Disable remote archiving and check if this is still the case?

 

[ivank]

What are your transcription settings?
If it includes Recordings, recordings that have no transcription will be transcribed during remote archiving. If you don't need this, set transcription options to Voicemail only.

 

[PreviouslyAsterisk]

Have you tried to Disable remote archiving and check if this is still the case?

I have turned it off - it was set for once a week, so we'll need to wait for next week to find out if this is the cause.

However, it would be nice to archive recordings - if this is the cause, could we submit this as a bug? I'll follow up next week.

 

[PreviouslyAsterisk]

What are your transcription settings?
If it includes Recordings, recordings that have no transcription will be transcribed during remote archiving. If you don't need this, set transcription options to Voicemail only.

Ah ha, this must be it!

The PBX has been around for a while, and has lots of calls recorded. However, transcription has only been in place for a week. I set up remote archiving last week, also.

So, from what you're saying, it sounds like every call stored in the system was being punted to OpenAI at once. If that's the explanation, then that's great, that's a one time activity - I just wanted to check it wasn't something more serious (and costly!) occuring.

 

[PreviouslyAsterisk]

Thanks both - I have just checked call recordings, and can see summaries for every call made.

At least we know the cause. Thanks!

 

[VoIPTools]

I encourage you to watch for the next version of our 3CX Recording Manager. This is a ground-up rewrite of one of our most popular tools. In the new version, we give you complete control over which extensions are processed by OpenAI, along with the ability to manually perform the action on individual recordings (if you only want to perform the analysis on selected recordings).

Further, we also make it possible to include your own OpenAI prompts to the analysis process. So, for example, if you want to perform agent scoring, now you can create your own prompts to score (as an example) whether the agent remembered to inform the caller that the call was being recorded. We have always had agent scoring built into our platform, but now you no longer need a human to perform the scoring. Now you can do it automatically with OpenAI!

I am really pushing the development team to get this new version out the door, but it looks like Friday next week at the earliest. The code is done and now we are just going through the Quality Assurance process, updating manuals, and getting help content added to the UI. Almost there!