Salesforce Agentforce vs Microsoft Copilot Studio vs HubSpot Breeze: Real AI Agent Cost in 2026

Three vendors, three completely different answers to "how much does one AI agent interaction cost." Salesforce lets you choose between $2 per conversation, a granular Flex Credit meter, or a per-user license. Microsoft bills a tenant-wide pool of Copilot Credits with no per-seat fee at all. HubSpot, as of April 14, 2026, charges nothing for an agent conversation that doesn't resolve — only for the ones that do. None of these numbers converts cleanly into the others, and this article does not force one.

What each vendor actually charges, and for what unit

Why a clean unit conversion isn't possible

Salesforce's own published rate card is the clearest illustration of why forcing these three into one blended "cost per interaction" figure would misrepresent the category. A standard Agentforce action under Flex Credits costs 20 credits ($0.10 at $500 per 100,000 credits); a voice action costs 30 credits ($0.15). The same underlying interaction billed as a full Conversation costs $2.00 — a 20-fold spread from Salesforce's own two billing models applied to a single ordinary action, before Microsoft or HubSpot even enter the comparison, and a real interaction may consume several actions plus other metered usage types, not exactly one action alone. Microsoft's Copilot Credit rate card is genuinely granular — a classic answer costs 1 credit, a generative answer costs 2, an agent action costs 5, tenant graph grounding costs 10 — meaning "how many customer interactions does $200 buy" depends entirely on which features a specific agent design actually uses, not on one universal per-message rate. HubSpot's unit is the cleanest of the three specifically because it's tied to a defined, binary outcome (resolved or not) rather than a raw action or credit count — but that also means HubSpot's price is not a "per message" or "per session" rate comparable to either competitor's granular meters.

Three agent shapes, and how they change the comparison

• A. FAQ/self-service agent: Answers a single question from existing knowledge, no external system writes.
• B. Tool-using CRM agent: Looks up a customer record, checks order status, and updates a field — several actions per resolved interaction.
• C. Complex multi-action agent: Reasons over context, calls multiple tools or systems, and may escalate partway through — the highest actions-per-conversation shape of the three.

On Agentforce, shape A (a single standard action, no other usage) costs 20 credits ($0.10) under Flex Credits, well below the $2.00 flat Conversations rate. Shape C, involving several standard actions plus possibly a voice action, could plausibly reach or exceed $2.00 in accumulated Flex Credits once enough actions and other metered usage types stack up — at which point Conversations pricing becomes the more predictable option despite its higher single-action rate. This article does not assume a fixed number of actions equals exactly one conversation economically, since a real agent's action mix can include voice, tool calls, and other separately-metered usage beyond a simple count of standard actions.

On Microsoft Copilot Studio, shape A maps naturally to a generative answer (2 credits) or, for pre-scripted responses, a classic answer (1 credit). Shape B, involving a customer-record lookup and a field update, would likely consume one or more agent actions (5 credits each) on top of the base answer. Shape C, involving tenant-wide grounding and multiple actions, could combine a generative answer (2 credits), tenant graph grounding (10 credits), and two or more agent actions (5 credits each) — Microsoft's own published customer-support example illustrates this kind of stacking directly: an agent run with four classic answers and two generative answers costs (4×1)+(2×2) = 8 credits per run, while a tenant-graph-grounded sales agent with four generative answers and four grounding lookups costs (4×2)+(4×10) = 48 credits per run.

On HubSpot, actions-per-conversation don't change the price at all — the $0.50 is charged once per resolution regardless of how many internal steps the agent took to get there, which structurally favors HubSpot for complex, multi-step interactions in a way neither Salesforce billing model nor Microsoft's stacked-credit model replicates.

Cost at 1,000, 10,000, and 100,000 completed outcomes

Microsoft's Copilot Credit has two separate dollar values depending on how it's purchased, and this article keeps them distinct rather than blending them: pay-as-you-go bills at $0.01 per credit with no floor; a prepaid capacity pack costs $200/month for 25,000 credits (an effective $0.008/credit if the full pack is consumed), but that $200 is a monthly floor paid whether or not the full 25,000 credits are used — unused pack capacity is not refunded or credited back.

These figures assume every outcome is a pure 2-credit generative answer with no agent actions, grounding, or voice involved; a workflow adding those features consumes proportionally more credits under either billing model, and this article does not model pre-purchase annual discounts on top of the capacity-pack rate shown here.

The Flex Credits column uses Salesforce's own published 20-credits-per-standard-action rate at $500 per 100,000 credits ($0.005/credit); a workflow with more actions per outcome (shapes B and C), or one using voice (30 credits/action), would cost proportionally more under Flex Credits, narrowing or eliminating the gap against the flat $2.00 Conversations rate. The Copilot Studio columns use Microsoft's own published $0.01/credit PAYG rate and $200-per-25,000-credit pack rate specifically — not Salesforce's $500-per-100,000-credit rate, which applies only to Agentforce. A workflow adding agent actions (5 credits each) or tenant graph grounding (10 credits) costs correspondingly more under either Microsoft billing model — Microsoft's own sales-performance example (four generative answers plus four grounding lookups per run) works out to 48 credits per run, or $0.48 under PAYG ($0.01 × 48), more than four times the pure-generative-answer PAYG figure shown above. None of these figures include the underlying Service Cloud or HubSpot Hub subscription; Copilot Studio's capacity-pack fee (if that billing model is used) is itself the underlying subscription cost in this case, not an add-on to a separate required license.

What the advertised price misses

• Salesforce's Conversations rate is list pricing specifically, with volume discounts negotiated separately at the enterprise level — the effective rate at real deployment scale is very likely lower than $2.00, though no specific discounted figure is published.
• Microsoft's rate card stacks multiple credit charges onto a single interaction whenever multiple features fire in one turn — Microsoft's own documentation gives a worked example of a tenant-graph-grounded agent responding to one complex prompt at 12 credits (10 for grounding, 2 for the generative answer) in a single turn, illustrating that the "per interaction" cost on Copilot Studio depends on exactly which features that specific interaction triggers, not a single flat rate.
• A capacity-pack customer pays the $200 monthly floor even at very low usage, and unused capacity does not roll over as a refund or credit — a tenant consuming only 2,000 of a pack's 25,000 credits in a given month still pays the full $200, making PAYG billing ($20 for that same 2,000-credit usage) the cheaper option below the volume where a pack's per-credit rate advantage overtakes its fixed floor.
• HubSpot's outcome-based model is a real, dated improvement, not a marketing framing. Before April 14, 2026, the Customer Agent billed $1.00 per conversation whether or not it resolved anything; an escalated or abandoned conversation now costs nothing, which is a genuine structural change in the buyer's favor, not just a lower headline number.
• Copilot Studio's tenant-wide credit pool has no seat-based floor, which can favor a small team with high per-user AI usage — but the same pool being shared tenant-wide means a company can't easily attribute cost to a specific department or use case without its own internal tracking, unlike Agentforce's or HubSpot's per-outcome billing.
• HubSpot Credits reset monthly with no rollover, and Breeze Agents require Professional or Enterprise specifically — a company on HubSpot Starter cannot access Customer Agent or Prospecting Agent at any price without first upgrading its base subscription.
• Microsoft enforces overage at 125% of prepaid capacity, at which point custom agents are disabled for any new invocation until capacity is increased or reset — a real availability risk, not just a cost one, for a Copilot Studio deployment that doesn't actively monitor its credit consumption.
• None of the three vendors' AI usage price includes the underlying CRM or Service Cloud license it depends on (Copilot Studio's tenant credit pool is itself the licensed product, not an add-on to a separate required license in the way Agentforce and HubSpot's Hub subscriptions are) — a fair total-cost comparison must still add whichever base subscription each platform separately requires before comparing AI usage prices at all.

Employee-facing versus customer-facing usage is a separate cost layer entirely

All three platforms distinguish internal, employee-facing agent usage from external, customer-facing usage, and the pricing for each is genuinely separate. HubSpot's Breeze Assistant (internal) is included free with any paid seat — a real, no-cost internal AI layer distinct from the outcome-priced Customer and Prospecting Agents. Microsoft's own published billing rules state that employee-facing (business-to-employee) usage of Copilot Studio agents is included at no additional Copilot Credit charge specifically when the user is licensed with Microsoft 365 Copilot ($30/user/month, on a qualifying Microsoft 365 base license) and the agent runs under that user's authenticated identity — meaning a Microsoft 365 Copilot seat can absorb real internal agent usage that would otherwise draw down the tenant's Copilot Credit pool. Salesforce maintains a distinct employee-agent licensing structure apart from customer-facing conversation billing. A company evaluating "AI agent cost" across all three should keep these two categories — internal productivity agents and external customer-facing agents — in separate budget lines, since conflating them will overstate or understate either platform's real customer-facing cost.

Sources

Checked late September 2026. Salesforce: salesforce.com/agentforce/pricing/. Microsoft: microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio. HubSpot: hubspot.com/company-news/hubspots-customer-agent-and-prospecting-agent-now-you-pay-when-the-task-is-complete, hubspot.com/pricing, and HubSpot's Credits Rate Sheet documentation.