Forum Discussion
Helper IVHigh Copilot and AI cost
Hi there,
I'm the fabric adminstrator. I moved 2 tables of 500mb each into fabric datawarehouse F4 compute. I created data agent and used the copilot from fabric workspace to talk to data in 2 tables. I'm tracking the cost and noticed the cost seems to higher. the fabric run time is only around 4 hrs a day for POC. Only 1 person is querying. may be 20times a day. Is there any way to reduce the copilot AI cost? Im concerned if we were to open to 300+ users in the organization, dont want to go bankrupt. 😄
or am i using the AI feature incorrectly, may be there is any other cost efficient way? or is this the only cost effecient way?
If multiple people ask the same question like sales by region, does agent remember earlier answers or run back and forth costing up more even though data is static (any config to mention data is refreshed only once everyday)?
Hello AJAJ,
Thank you for posting your query in Microsoft Fabric Community Forum. Also, thanks to lbendlin , for those inputs on this thread.Copilot doesn’t charge based on how big your tables are. Even if the data is small, Copilot runs multiple background steps for every question a user asks. Because of that, the compute usage becomes a bit higher than what we normally expect with simple SQL or Power BI queries. So the cost you are seeing right now is normal behaviour.
When this goes to a bigger user base, the load comes mainly from the AI requests, not from the data size. That’s why higher capacities are usually suggested it’s just to handle the AI workload smoothly.
At the moment there isn’t any special setting to reduce Copilot-related cost. The best way is to limit who can use Copilot during the POC and keep checking the capacity metrics to understand where the usage is coming from.
Your current setup is fine, and you’re not using it incorrectly. This is simply how Copilot processes queries in the background.
Let me know if you want me to check anything specific.
Regards,
Community Support Team.
11 Replies
- lbendlin
Super User
F4 is way too small for that. for 300+ users you are looking at a minimum of F128
AJAJMy eyebrows jumped out of my face 😄
Can you plz help me what am i missing. why F128 is needed for 1GB of data in total? worstcase 5gb of data since dim / facts are designed compact. The queries are running in seconds whatever way data is sliced. 70m fact rows. Thats in F4. PBI would get copy of it and users would be hitting pre canned power bi report once a while. Not all users would be hitting the PBI at the same. I assume F64 itself would be waaaayyyyy too much power but kind a breaks even on PBI license. $15/month for 300 users= $4500 per month.
Just trying to understand- lbendlin
We recently did a test of Copilot queries an a similar sized semantic model on a F64. It took 15 concurrent users about 10 minutes to completely overwhelm the capacity.
Hello AJAJ,
Thank you for posting your query in Microsoft Fabric Community Forum. Also, thanks to lbendlin , for those inputs on this thread.Copilot doesn’t charge based on how big your tables are. Even if the data is small, Copilot runs multiple background steps for every question a user asks. Because of that, the compute usage becomes a bit higher than what we normally expect with simple SQL or Power BI queries. So the cost you are seeing right now is normal behaviour.
When this goes to a bigger user base, the load comes mainly from the AI requests, not from the data size. That’s why higher capacities are usually suggested it’s just to handle the AI workload smoothly.
At the moment there isn’t any special setting to reduce Copilot-related cost. The best way is to limit who can use Copilot during the POC and keep checking the capacity metrics to understand where the usage is coming from.
Your current setup is fine, and you’re not using it incorrectly. This is simply how Copilot processes queries in the background.
Let me know if you want me to check anything specific.
Regards,
Community Support Team.Hi AJAJ,
I hope the information provided above assists you in resolving the issue. If you have any additional questions or concerns, please do not hesitate to contact us. We are here to support you and will be happy to help with any further assistance you may need.
Regards,
Community Support Team.Hi AJAJ,
I wanted to follow up and see if you have had a chance to review the information that was shared. If you have any additional questions or need further clarification, please don’t hesitate to reach out. I am here to assist with any concerns you might have.
Regards,
Community Support Team.
the simplest way to reduce it use less tokens. i think the ai copilot that you are currently using the just dumping the data into the context window of the model which is causing it to bloat and increase cost very rapidly. it is costly to run for even 1 user and here we are talking about about 300+ users. the best way is to give your ai some memory, and store intermidiate outputs there, aggregate the data before giving it to ai copilot and let math be math, do not let ai do the maths for you.
also F4 should be enough to run it if properly used, F128 is just too much, i understand as developers think, more compute is better but more compute is more money, and here the money gets wasted.
we can get on a quick call if need be, here is my email, [email protected]- AJAJ
Though we got hijacked into few other conversation, will wait for somone to answer the original question. Thanks
- querixa
I totally get your concern about Copilot costs scaling with more users, it's a real issue many teams face when trying to bring AI insights to everyone. From my experience, one way to reduce AI-related costs is to explore custom solutions that run inside your existing dashboards without per-user licensing. For example, I built PowerMind as an embedded AI assistant that works directly in Power BI and doesn’t require Copilot licenses, which can help control costs when rolling out to hundreds of users. If you’re open to alternatives, you can check out a quick demo of how it works here:
