Manage AI Use Credit Expectations

Asana Community Community Tips
, , ,
1

With AI Studio now well-established and Teammates proving their worth, it’s very easy to get caught up with what is possible and lose sight of the implications of jumping into the new tech with no oversight or budgeting.

It is very easy to blow through a month’s worth of credits in hours if you don’t test and prepare accordingly. So before you set your users’ expectations about the possibilities, I recommend to do the following:

  1. Test use cases and how many credits they consume.
    Create a private project and test common and known use cases. Se how many credits they consume and establish some benchmarks when these vary between similar activities (assigning tasks, setting custom field values etc)
  2. Determine frequency of consumption
    How many times a month would this automated workflow occur? Even a rough approximation gives you a reasonably accurate expectation of total credit usage when you combine it with the benchmark credit usage figures you identified in step 1.
  3. Compare to credit budget
    What proportion of your total credit budget would this total consume? How many credits does this leave you for all other automated processes? Is the time-saving worth the credit expenditure, and the potential need to then secure real-world budget for additional credits?
  4. Think about the ROI
    What benefit do your users and organisation get from this potential credit expenditure? If credits are being unused, then go for it, but if this would take you close to or over your monthly budget, you will need to secure additional credits, prioritise which automations should be in place, or risk having a potentially essential workflow break at some indeterminate point in the month.
5 Likes
2

I would also add that as others have shared in these forums, tweaking the prompts can impact the credit usage significantly. We’ve noticed a big improvement by changing the prompts themselves, including parameters for when to stop or ask for help. For example, we started with a prompt that gave suggested keywords and asked the AI to create a task based on those keywords. So if someone submitted a form that had the words “new image” it should create a task for our designer to create a new image (and include the details in the new task description).

But we found that the AI would often stretch its logic which caused more token burn. The AI was expanding our logic to assume that if a word was pluralized like “I need a new image of people holding balloons” it would assume that multiple images were needed because of ‘people’ and ‘balloons’. The extra tasks were annoying to consolidate, but it also was burning a lot of tokens trying to parse the plurals and how many tasks should be created for a nebulous word like “people” which could result in 2 tasks or 200 tasks.

We addressed this in two ways:

  1. Coaching our users to give a discrete number like “I need one image of people holding balloons”
  2. Updating the prompt to interpret the image count based on the request, not the content of the image

This reduced the cost of this specific automation by 75%!

1 Like