Microsoft didn’t have a Claude problem. Uber didn’t have an AI problem. They had a governance problem. And if your Salesforce org is rolling out Agentforce without a cost strategy in place, you may be walking into the same trap.
This is not a cautionary tale about AI being too expensive. It is a signal that enterprise AI has moved out of the experimental phase. The CFOs have started reading the token receipts.
What Actually Happened at Microsoft and Uber
In December 2025, Microsoft launched an internal Claude Code pilot across its Experiences and Devices division. Within months, the team had burned through its entire annual AI budget. The pilot was cancelled, effective June 30, 2026.
Uber’s story runs parallel. The company’s CTO sent an internal memo in April 2026 confirming that Uber had consumed its entire 2026 AI coding tools budget in just four months.
The culprit in both cases was the same: a shift from flat-seat licensing to token-based billing. Flat fees kept usage invisible. The moment billing tied to consumption, the true cost became impossible to ignore.
This is precisely the model Salesforce uses for Agentforce.
Why This Matters for Salesforce Orgs
Agentforce runs on a consumption-based pricing model. Each agent action, including updating a record, summarising a case, or executing a prompt, consumes Flex Credits. The current rate is 20 Flex Credits per action, at $0.005 per credit, which works out to $0.10 per action.
That sounds manageable in isolation. The problem scales.
Consider a mid-sized org handling 3,000 cases per day through an Agentforce service agent. Each case might involve 4 to 6 agent actions. At $0.10 per action, a single busy month can generate tens of thousands of dollars in credit consumption before anyone in finance notices.
Add voice interactions, where costs jump to 30 Flex Credits per action or 60 Flex Credits per minute depending on the billing model, and the numbers accelerate quickly.
The Microsoft and Uber situations are not edge cases. They are what happens when agentic AI meets unrestricted usage inside a large organisation.
The Agentforce Pricing Model: A Quick Orientation
Before you can govern costs, you need to understand the structure. Agentforce currently offers two primary pricing tracks.
Per-conversation pricing charges $2 per 24-hour conversation window. This model suits well-defined, contained use cases, like a customer service bot handling straightforward enquiries.
Flex Credits offer a pay-per-action model. Each standard action consumes 20 credits ($0.10). This model gives finer visibility into exactly what the agent is doing and why it costs what it costs.
Additionally, each standard action includes processing up to 10,000 tokens. If a single action exceeds that threshold, it counts as multiple actions. An agent loading extensive customer history during a voice call can easily trigger this, multiplying the credit cost unexpectedly.
The Digital Wallet inside Salesforce tracks usage in real time across orgs, agents, and teams. Most Salesforce teams are not using it proactively. They should be.
Four Governance Controls Salesforce Teams Should Implement Now
1. Set Hard Credit Caps per Agent
Agentforce’s Digital Wallet allows you to monitor consumption, but monitoring alone is reactive. Work with your Salesforce AE to establish hard spending limits at the agent level. Define what a reasonable monthly credit envelope looks like per use case before you go live, not after you receive the bill.
2. Audit Token Usage per Action
Each action processes up to 10,000 tokens before triggering additional credit charges. Review the context your agents are injecting into each action. Long knowledge article loads, extensive case histories, and verbose system prompts all push token counts higher. Trim context to what the agent genuinely needs.
3. Prioritise High-ROI Use Cases First
A common mistake is broad deployment before validation. Start with one or two use cases where deflection rates and resolution times are measurable. Case FAQ handling and basic field service scheduling are proven starting points. Expand only after you have real cost-per-resolution data to justify it.
This is the approach that separates pilot success from budget overrun. Microsoft’s mistake was enabling wide internal access before establishing consumption governance. Salesforce orgs can avoid that pattern.
4. Use Sandbox Usage for Development
Salesforce discounts Flex Credit usage in sandbox environments. Development and testing work should stay in sandbox wherever possible. Running complex multi-step agent tests in production environments burns production credits unnecessarily.
The Governance Checklist Before You Expand Agentforce
Before approving any new Agentforce agent or expanding an existing deployment, work through these questions.
- Have you defined the expected action volume per month for this agent?
- Have you reviewed the token consumption of each action type in testing?
- Is the Digital Wallet configured with alerts at 50%, 75%, and 100% of your credit budget?
- Have you confirmed whether per-conversation or Flex Credit pricing is more cost-effective for this use case?
- Does your finance team have visibility into Agentforce credit consumption?
- Have you scoped the Data Cloud dependency and associated costs?
That last point matters. Agentforce at scale requires a Data Cloud subscription, which adds significant infrastructure cost before a single agent action is executed.
The Broader Signal for Salesforce Practitioners
The Microsoft and Uber situations mark a shift in how enterprise leadership views AI tooling. The experimental budget era is closing. ROI expectations are arriving.
For Salesforce practitioners, this is actually a positive development. It pushes the conversation away from “can we deploy an agent?” toward “what business outcome does this agent drive, and at what cost?” That is the right framing.
Agentforce is a capable platform. The Flex Credit model, when governed properly, gives you more cost visibility than most enterprise AI tools on the market. The Digital Wallet, token-level billing, and sandbox discounting are all features that support responsible governance.
The organisations that succeed with Agentforce in 2026 will not be the ones that deployed fastest. They will be the ones that deployed with measurement built in from day one.
Key Takeaways
- Microsoft and Uber burned through annual AI budgets in months due to unmanaged token-based billing.
- Agentforce uses the same consumption model: Flex Credits charged per agent action.
- Each action processes up to 10,000 tokens before additional charges apply; prompt design directly affects cost.
- The Digital Wallet provides real-time visibility, but proactive credit caps are essential.
- Start with high-ROI, well-scoped use cases and expand based on cost-per-outcome data.
- Enterprise AI governance is no longer optional. It is a finance and architecture requirement.
Read More!
Your First Salesforce + Claude Project Without Writing a Single Line of Code
Build, deploy, and manage AI agents at scale.
Leave a Reply