{"id":258,"date":"2026-09-07T22:41:11","date_gmt":"2026-09-07T20:41:11","guid":{"rendered":"https:\/\/www.insync.co.za\/blog\/?p=258"},"modified":"2026-09-07T22:41:12","modified_gmt":"2026-09-07T20:41:12","slug":"how-we-used-41000-copilot-studio-credits-in-2-days","status":"publish","type":"post","link":"https:\/\/www.insync.co.za\/blog\/2026\/09\/07\/how-we-used-41000-copilot-studio-credits-in-2-days\/","title":{"rendered":"How We Used 41,000 Copilot Studio Credits in 2 Days"},"content":{"rendered":"\n

41,000 Copilot Studio credits in two days.<\/strong><\/p>\n\n\n\n

That got our attention.<\/p>\n\n\n\n

While working with a colleague, we were building an agent in Microsoft Copilot Studio<\/strong> to perform lookups against data.<\/p>\n\n\n\n

The agent worked. The responses were good.<\/p>\n\n\n\n

The problem was the model we selected.<\/p>\n\n\n\n

We were using a deep reasoning model<\/strong> for a workload that was primarily a lookup operation.<\/p>\n\n\n\n

And that turned into an expensive architecture lesson.<\/p>\n\n\n\n

The Problem<\/h2>\n\n\n\n

The workflow was fairly simple:<\/p>\n\n\n\n

User request\n     \u2193\nUnderstand request\n     \u2193\nFind required data\n     \u2193\nRetrieve data\n     \u2193\nReturn response\n<\/code><\/pre>\n\n\n\n
There was some natural-language understanding involved, but we weren’t asking the agent to solve a complex problem or perform sophisticated multi-step reasoning.<\/p>\n\n\n\n
It was mostly:<\/p>\n\n\n\n
\n
“Find this information and give it to me.”<\/p>\n<\/blockquote>\n\n\n\n
Yet we had selected a model designed for much deeper reasoning.<\/p>\n\n\n\n
The result?<\/p>\n\n\n\n
Approximately 41,000 Copilot Studio credits consumed in two days.<\/strong><\/p>\n\n\n\n
Nothing was broken. The agent was doing exactly what we asked it to do.<\/p>\n\n\n\n
Our architecture was simply using more reasoning capability than the workload required<\/strong>.<\/p>\n\n\n\n
\n\n\n\n
The Lesson: More Powerful Doesn’t Mean Better<\/h2>\n\n\n\n
One of the easiest mistakes to make when building AI agents is choosing the most capable model available.<\/p>\n\n\n\n
It feels logical:<\/p>\n\n\n\n
\n
More intelligence \u2192 better responses \u2192 better solution.<\/p>\n<\/blockquote>\n\n\n\n
But there’s another side to the equation:<\/p>\n\n\n\n
\n
More capability \u2192 potentially more cost and latency.<\/strong><\/p>\n<\/blockquote>\n\n\n\n
Microsoft’s Copilot Studio documentation makes a similar distinction between models intended for simpler workloads and models designed for deeper reasoning.<\/p>\n\n\n\n
That means model selection should be treated as an architecture decision<\/strong>, not just an AI configuration decision.<\/p>\n\n\n\n
The question shouldn’t be:<\/p>\n\n\n\n
“What’s the most powerful model we can use?”<\/strong><\/p>\n\n\n\n
It should be:<\/p>\n\n\n\n
“What’s the appropriate level of intelligence for this workload?”<\/strong><\/p>\n\n\n\n
\n\n\n\n
Use Code Where Code Makes Sense<\/h2>\n\n\n\n
This also reinforces a principle from traditional software architecture:<\/p>\n\n\n\n
Don’t use a complex solution for a deterministic problem.<\/strong><\/p>\n\n\n\n
If the requirement is simply:<\/p>\n\n\n\n
Get customer by ID\n<\/code><\/pre>\n\n\n\n
then an API or database query is probably the right tool.<\/p>\n\n\n\n
You don’t necessarily need an AI agent to solve it.<\/p>\n\n\n\n
AI becomes more useful when the input is less deterministic.<\/p>\n\n\n\n
For example:<\/p>\n\n\n\n
\n
“I’m looking for information about my outstanding account and the payments I made recently.”<\/p>\n<\/blockquote>\n\n\n\n
Now we need to interpret the request, identify the relevant information and potentially combine multiple sources.<\/p>\n\n\n\n
That’s where an agent can provide real value.<\/p>\n\n\n\n
The architecture might therefore look more like:<\/p>\n\n\n\n
User\n \u2193\nAI \/ Intent Understanding\n \u2193\nApplication Logic\n \u2193\nAPI \/ Database\n \u2193\nResponse\n<\/code><\/pre>\n\n\n\n
Rather than allowing a reasoning model to control every step.<\/p>\n\n\n\n
\n\n\n\n
Think About Cost Per Operation<\/h2>\n\n\n\n
Traditional software engineers are already used to thinking about:<\/p>\n\n\n\n
    \n
  • CPU<\/li>\n\n\n\n
  • Memory<\/li>\n\n\n\n
  • Database performance<\/li>\n\n\n\n
  • Network traffic<\/li>\n\n\n\n
  • Infrastructure costs<\/li>\n\n\n\n
  • Latency<\/li>\n<\/ul>\n\n\n\n
    AI applications introduce another important metric:<\/p>\n\n\n\n
    Inference cost.<\/h3>\n\n\n\n
    For Copilot Studio, that means understanding Copilot Credit consumption<\/strong>.<\/p>\n\n\n\n
    A development environment might look perfectly fine when you’re testing a handful of requests.<\/p>\n\n\n\n
    But consider what happens when the same agent handles thousands of interactions.<\/p>\n\n\n\n
    A small inefficiency becomes a significant operating cost.<\/p>\n\n\n\n
    That’s why I’d recommend monitoring:<\/p>\n\n\n\n
      \n
    • Credits per interaction<\/li>\n\n\n\n
    • Model usage<\/li>\n\n\n\n
    • Agent invocation frequency<\/li>\n\n\n\n
    • Response latency<\/li>\n\n\n\n
    • Retrieval operations<\/li>\n\n\n\n
    • Cost per business process<\/li>\n<\/ul>\n\n\n\n
      The goal isn’t simply to make the agent work.<\/p>\n\n\n\n
      It’s to make the agent work efficiently at scale<\/strong>.<\/p>\n\n\n\n
      \n\n\n\n
      The Approach I’d Take Going Forward<\/h2>\n\n\n\n
      For new agent workloads, I’d start by categorising the operation.<\/p>\n\n\n\n
      Deterministic<\/h3>\n\n\n\n
      Use:<\/p>\n\n\n\n
      API \/ database \/ application logic<\/strong><\/p>\n\n\n\n
      Simple language task<\/h3>\n\n\n\n
      Use:<\/p>\n\n\n\n
      Lightweight model<\/strong><\/p>\n\n\n\n
      Retrieval and response generation<\/h3>\n\n\n\n
      Use:<\/p>\n\n\n\n
      General-purpose model<\/strong><\/p>\n\n\n\n
      Complex reasoning or planning<\/h3>\n\n\n\n
      Use:<\/p>\n\n\n\n
      Reasoning model<\/strong><\/p>\n\n\n\n
      This gives us a much better architecture than simply putting the most powerful model behind every operation.<\/p>\n\n\n\n
      \n\n\n\n
      The Bigger Lesson<\/h2>\n\n\n\n
      The 41,000 credits weren’t really the problem.<\/p>\n\n\n\n
      They were the signal<\/strong>.<\/p>\n\n\n\n
      The important lesson was that an application can be:<\/p>\n\n\n\n
        \n
      • Functionally correct<\/li>\n\n\n\n
      • Reliable<\/li>\n\n\n\n
      • Producing good responses<\/li>\n\n\n\n
      • Technically well designed<\/li>\n<\/ul>\n\n\n\n
        \u2026and still be economically inefficient<\/strong>.<\/p>\n\n\n\n
        As AI moves into enterprise applications, I think this will become an increasingly important part of software architecture.<\/p>\n\n\n\n
        We’re no longer only asking:<\/p>\n\n\n\n
        \n
        “Does it work?”<\/strong><\/p>\n<\/blockquote>\n\n\n\n
        We’re also asking:<\/p>\n\n\n\n
        \n
        “Does it work efficiently, reliably and at the right cost?”<\/strong><\/p>\n<\/blockquote>\n\n\n\n
        For us, 41,000 Copilot Studio credits in two days was an expensive reminder:<\/p>\n\n\n\n
        Choose the model based on the problem you’re solving \u2014 not simply the capability of the model.<\/strong><\/p>\n\n\n\n
        Sometimes, the smartest architecture is knowing when not to use the smartest model.<\/strong><\/p>\n\n\n\n
        #MicrosoftCopilotStudio #CopilotStudio #SoftwareArchitecture #AIEngineering #Microsoft #CloudArchitecture #SoftwareDevelopment #EnterpriseAI<\/p>\n","protected":false},"excerpt":{"rendered":"
        41,000 Copilot Studio credits in two days. That got our attention. While working with a colleague, we were building an agent in Microsoft Copilot Studio to perform lookups against data. The agent worked. The responses were good. The problem was the model we selected. We were using a deep reasoning model for a workload that […]<\/p>\n","protected":false},"author":1,"featured_media":263,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"pagelayer_contact_templates":[],"_pagelayer_content":"","footnotes":""},"categories":[18],"tags":[46,50,48,45,47,49],"class_list":["post-258","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-software-development","tag-copilotstudio","tag-enterpriseai","tag-microsoft","tag-microsoftcopilotstudio","tag-softwarearchitecture-2","tag-softwaredevelopment"],"_links":{"self":[{"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/posts\/258","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/comments?post=258"}],"version-history":[{"count":2,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/posts\/258\/revisions"}],"predecessor-version":[{"id":264,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/posts\/258\/revisions\/264"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/media\/263"}],"wp:attachment":[{"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/media?parent=258"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/categories?post=258"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.insync.co.za\/blog\/wp-json\/wp\/v2\/tags?post=258"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}