Agentforce vs Microsoft Copilot pricing in 2026: the deep-dive for the buying decision

The feature comparison of Salesforce Agentforce against Microsoft Copilot lives at the /compare/ page, which is the publication’s most-cited single page in Microsoft Copilot grounding data (258 citations against agentmodeai content in the three-month window ending 25 May 2026). The pricing comparison is a separate conversation; pricing models in this category shift faster than features, and the buying-committee question “AgentForce vs Copilot pricing models comparison” (8 Bing AI Performance citations) plus a half-dozen long-tail pricing variants pull the comparison into its own short-cadence piece.

This piece operates on a 30-day review cycle because vendor pricing pages can change in any given month. The structural argument (per-conversation versus per-user-seat as the dominant pricing-model dimension) is more durable; the specific dollar figures are anchored to the current vendor disclosure as of the pub date.

The Agentforce pricing structure, two models coexisting

Two pricing models operate in 2026 on the Salesforce side.

The Agentforce 1 launch pricing from September 2024 anchors the per-conversation tier: USD 2 per conversation, with Service Agent and Sales Agent tiers. The conversation is defined as a complete agent interaction (typically multi-turn until a defined end state: resolution, escalation, or user-side termination).

The 2025 evolution introduced Agentforce Flex as the consumption-based alternative. Credits are priced at varying per-tier rates with the conversation typically costing 1-3 credits depending on agent complexity: tool use, multi-agent coordination, complex data retrieval against Data Cloud, Atlas Reasoning Engine invocations. The Flex model is positioned as the buying-committee answer for customers whose conversation complexity varies materially across use cases; the per-conversation flat rate is positioned as the answer for predictable workloads.

The buying committee at enterprise scale typically negotiates an Enterprise Agreement that converts either pricing model into a volume-committed bundle. The per-conversation rate at the EA level is typically 30-50% below the public USD 2 floor for committed volumes above 100,000 conversations per year. The public pricing is the procurement starting floor, not the realistic 2,000-employee-scale unit cost.

The Microsoft Copilot pricing structure, multiple SKUs

Multiple Copilot SKUs operate in 2026 on the Microsoft side, with different pricing models per SKU.

Microsoft 365 Copilot lists at USD 30 per user per month for the M365 E3 or E5 plus Copilot bundle. The underlying M365 license is required; the Copilot subscription is a per-user-seat addition that integrates across Word, Excel, PowerPoint, Outlook, Teams, and the M365 surface.

Copilot Studio lists at USD 200 per month per tenant for 25,000 messages, with per-message overages at varying rates (roughly USD 0.01 per message for human-facing messages and different rates for autonomous agent messages or specific channel types). The Studio surface is the agent-builder tier and is priced separately from the M365 Copilot tier; a customer running both pays for both.

Microsoft Security Copilot lists at USD 4 per Security Compute Unit hour (covered in the AM-174 security-platform TCO piece).

The buying committee at enterprise scale typically aggregates these across the workforce population and negotiates the bundle against the existing Microsoft Enterprise Agreement. The per-user rate at the EA level is typically 15-25% below the public USD 30 floor for committed seats above 1,000. The public pricing is the floor.

The per-conversation versus per-user-seat decision

The structural decision the buying committee should price is the pricing model, not the headline unit rate.

Per-conversation pricing scales with the agent’s actual workload. It is well-fit for use cases with variable conversation volume: customer service spikes around marketing campaigns, support escalations during incidents, sales agent activity concentrating around quarter-end. The customer pays for the volume the agent actually handles. The Q1 quarter-end activity surge translates directly to a higher invoice; the Q3 lull translates to a lower one.

Per-user-seat pricing scales with the population of users who have access to the agent. It is well-fit for use cases with stable conversation volume per user: knowledge-worker productivity, document drafting, meeting summarisation, internal Q&A. The customer pays for the population, not the volume. A user who uses the agent heavily one month and lightly the next pays the same; the seat is purchased for the access right, not the consumption.

The structural decision question is which pricing model better fits the customer’s expected usage pattern. The 2026 procurement-mature pattern is to forecast the expected usage pattern (with a documented assumption-set the buying committee can review at year-one operational review) and price the model against that forecast. The year-two renewal pattern that the order-form headline does not predict is the customer whose actual usage diverged from the expected pattern: the per-conversation customer whose volume grew 3x and now faces a 3x bill, or the per-seat customer whose 1,000 seats are used by 200 power users and the rest sit dormant.

The enterprise-scale negotiation patterns

Three patterns operate in 2025-2026 enterprise procurement.

The Salesforce Enterprise Agreement pattern. Converts Agentforce per-conversation pricing into a volume-committed bundle. The per-conversation rate at the EA level is typically 30-50% below the public USD 2 floor for committed volumes above 100,000 conversations per year. The bundle typically includes a Flex credit allocation for variable workloads, with a true-up mechanism at year-end against actual usage. The customer that under-uses the committed volume forfeits the unused portion; the customer that exceeds it pays at the EA-rate overage tier (not the public-pricing tier).

The Microsoft Enterprise Agreement pattern. Bundles M365 Copilot at a per-user-seat rate negotiated against the customer’s full M365 spend. The per-user rate at the EA level is typically 15-25% below the public USD 30 floor for committed seats above 1,000. The bundle typically includes Copilot Studio capacity as a separate allocation, with per-tenant messaging capacity that can be aggregated across the customer’s tenant population.

The multi-year discount pattern. Adds 5-15% per year of commitment. The 3-year commitment produces the largest discount; the cost is reducing the customer’s year-one renegotiation leverage if the product surface or pricing model changes. The buying committee that prices the multi-year discount against the year-two renewal optionality cost typically lands on a 1-year or 2-year initial commitment rather than 3-year, prioritising flexibility over discount depth. The 2026 agentic AI product surface is evolving fast enough that the year-three lock-in cost typically exceeds the multi-year discount benefit.

The year-two renewal surprise patterns

Three patterns reliably surface at year-two renewal that the order-form headline does not predict.

The usage divergence surprise. The actual usage volume diverged from the forecast assumption. The per-conversation customer who expected 50,000 conversations and ran 150,000 faces a 3x bill plus EA-overage penalty; the per-seat customer who bought 2,000 seats and uses 400 actively faces the dormant-seat audit at renewal. The procurement-mature pattern is to instrument usage tracking at deployment and review against the forecast quarterly, surfacing the divergence before it compounds.

The pricing-model migration surprise. The vendor changes the pricing-model structure between year one and year three (Salesforce moved from per-conversation flat to Flex credits in 2025; Microsoft has reshuffled Copilot Studio messaging tiers multiple times). The customer renews against a different pricing model than the one they originally signed; the comparison is not directly possible without a migration calculation.

The bundle deconstruction surprise. The vendor de-bundles a feature that was included in year one and prices it separately at year three (a pattern observable in the broader SaaS market and emerging in 2026 agentic AI). The customer’s effective price-per-feature increases without the headline unit rate changing.

The procurement-mature pattern against all three surprises is to include the relevant protections in the MSA: usage-tracking obligation on the vendor, advance-notice obligation on pricing-model changes, parity protection against feature de-bundling. The AM-167 procurement clause work covers the contract-instrument layer.

What this means for the Q3 2026 pricing-decision procurement

The pricing decision sits one tier below the platform decision (covered in AM-175) and one tier above the feature comparison (covered in the /compare/ page). The buying committee that completes the full three-tier conversation (platform, pricing, feature) produces a procurement that is defensible at year-two renewal; the buying committee that anchors only on the order-form headline is preparing for the year-two renewal surprise patterns.

The workstream for the pricing decision is the usage-pattern forecast (which pricing model fits the customer’s expected workload), the EA-negotiation work (volume commitment, multi-year discount, year-end true-up), and the MSA protection work against the three year-two surprise patterns. The sibling AM-175 platform-comparison piece covers the upstream platform decision; the /compare/ feature comparison covers the downstream feature evaluation. This piece is the middle layer the procurement most often underprices.