Agentforce vs Copilot Studio vs Gemini vs the OpenAI Agents API vs Claude Managed Agents: five billing units, two worked bills, overage rules, audit, regulation and a scoring template.
The five platforms on every enterprise agent shortlist do not sell the same unit. Salesforce Agentforce bills the action (20 Flex Credits at $500 per 100,000) or the conversation ($2). Microsoft Copilot Studio bills the Copilot Credit (25,000-credit packs at $200 a month, with a published table from 1 credit for a classic answer to 10 for tenant graph grounding) and, since September 25, 2026, puts its long-running Autopilot agent on usage-based billing while chat stays on the per-user license. Google sells Gemini Enterprise by the seat and Gemini models by the token, with a Flash price doubling already scheduled for January 1, 2027. OpenAI bills tokens plus a hosted container session. Anthropic bills tokens plus $0.08 per session-hour, metered only while the session is running. A single dollar figure per platform on a slide is hiding the assumptions that produced it; this workbook makes you write them down. Two worked bills run the same workloads through all five. The first is Microsoft's own 900-conversation-a-day customer-service example extended to every platform, with the token shape stated and the arithmetic shown: from $729 a month on Gemini Flash at 2026 prices to $54,000 on the Agentforce conversation meter, with Copilot Studio at $1,800 in Microsoft's mix and $6,200 once tenant graph grounding is switched on. The second is a back-office order-processing agent that runs 15,000 times a month with no human in the loop, where the same four steps cost $200 as Copilot Studio agent flows and $20,800 as agent actions with a reasoning model, and where a hosted container minimum can be most of a small model's bill. Then the chapters buyers skip until the second invoice: what each platform does the day you run over (Microsoft disables custom agents at 125% of prepaid capacity; Salesforce bills overage at the contracted rate in arrears; Anthropic pauses the session at a budget you set), who lets you bring your own model and on which meter, and why cache-read price and US-only inference multipliers now decide agent cost more than the headline rate. The rest is the process. An identity, memory and audit chapter built on what the vendors actually publish: Autopilot's own tenant identity and Agent 365, Fin inside Salesforce AI Labs, Claude memory stores and the Compliance API, and the written questions to send where the pages say nothing. The regulatory overlay with exact dates and duties: EU Article 50 disclosure in force since August 2, 2026 with Annex III moved to December 2, 2027; Colorado SB 26-189 operative January 1, 2027; California SB 947 effective July 1, 2027. An evaluation-harness protocol that runs the same frozen 50 tasks on two finalists for one week with blind grading, a 12-item adversarial battery, and cost captured from the vendor's own meter. A 12-term negotiation checklist and a weighted scoring template with hard disqualifiers. Every price is quoted as the vendor page stated it, with the URL and read date in a sources chapter, and every figure that could not be read on a vendor, regulator or legislature page is labelled as such.
Every one of these is free and email-gated, same as this one.
Decode per-resolution pricing, audit your knowledge base, and run a real 200-question bake-off before you sign an AI customer support contract.
PDFA practical guide to moving enterprise AI from stalled pilots to production — governance, sequencing, and the patterns that actually ship.
PDFHow to build an FDE practice — the role that closes the gap between AI capability and real deployment, and why demand for it grew 800% since 2025.
PDF