CRM AI Pricing Shake-Out 2026: Seats, Tokens, Outcomes

The CRM line item used to be simple: seats times months. In 2026 the same renewal arrives with three meters running at once — seats, tokens, and outcomes — and the vendor will let you pick whichever one is most expensive for you. Pick wrong and you’re funding their R&D.

The shake-out is not theoretical. Two recent customer audits we ran landed bills 40% over the procurement model on identical workloads. Same vendor, same SKU, same usage. The difference was which meter the contract foregrounded.

How the shift happened in 18 months

Look back at vendor pricing pages from early 2024. Seats and seats only. By Dreamforce 2024, Salesforce announced Flex Credits. By mid-2025, every vendor had a consumption tier. By Q4 2025, three vendors had outcome SKUs in market.

Customers drove some of this — early Agentforce deals at flat seat AI pricing got “broken” by power users running 200 actions a day, and the vendor losses spooked finance. Vendors drove the rest — once outcome pricing was on the table at Sierra and Intercom, every legacy vendor had to answer.

The lesson for buyers: the pricing model in your renewal contract has a half-life of about 12 months. Build that into negotiation. Don’t lock into year 3 of a model that won’t exist by year 3.

The three pricing models on the table

Seat-based. Per-user-per-month. Predictable. Aligns to headcount, not value. Salesforce Sales Cloud, Dynamics 365 Sales, HubSpot Sales Hub all still anchor here.

Token-based / consumption. Per million input/output tokens, per agent action, per “Flex Credit”, per “Power Platform Request”. Agentforce Flex credits, Copilot Studio message packs, Now Assist consumption units, Zoho’s Zia AI credits, Freshworks Freddy Copilot credits all live here. Predictably unpredictable.

Outcome-based. Per resolved case, per qualified lead, per booked meeting, per closed-won influenced deal. Sierra charges per resolved conversation. Decagon and Ada too. Salesforce’s “Agentforce 360 Outcomes” SKU prices per autonomous resolution in tier-1 service. Intercom’s Fin charges per resolution. The outcome unit varies — what counts as a “resolution” is contractual, not natural.

Three meters. Three risk profiles. Three different things the vendor sales engineer will model for you, optimistically. The model in their deck is not the model in your invoice — the gap is your operating problem.

Why the shake-out is happening now

Two forces. First, AI flipped the seat-per-user assumption — one rep with five agents does the work of four reps, so the vendor wants the value back. Second, customers got burned by uncapped consumption in 2024–2025 and are demanding ceilings or refunds. Vendors are responding with hybrid SKUs that bundle a token floor with outcome upside.

The honest read: vendors are running an experiment on you. The pricing page will change three times before your renewal.

What each model actually costs

A 500-seat mid-market sales org, mature pipeline, moderate AI use:

Model	Year 1 list	Year 1 likely	Year 3 trajectory
Seats only (no AI)	$900k	$720k	flat
Seats + token add-on	$1.4M	$1.6M (overage)	+25% / yr
Seats + outcome SKU	$1.1M	$1.2M	depends on volume
Pure outcome (rare)	$0 + per-unit	$1.8M at scale	tied to revenue

The token row is the trap. Vendors quote average consumption from pilot customers — usually engineering-heavy users on capped workloads. Production sales orgs with agentic email drafting, meeting prep, and account research routinely exceed quotes 2–3x in month two.

Why seat pricing is dying

Seat pricing assumes a 1:1 mapping between licensed user and value delivered. AI breaks that assumption. One AE with three agents working renewals does the work that took four AEs in 2024. The vendor sees the same seat count and the same MRR while customer value has 3x’d. From the vendor’s CFO chair, that’s a structural revenue leak.

So seats are getting unbundled. Year 1 you keep the seat-priced tier. Year 2 the AI features move into a separate SKU. Year 3 the “Sales Cloud” SKU is renamed and the AI add-on is required for parity with last year’s experience. By year 4 the seat fee is half what it was and the AI fee is 2x.

This is not theoretical. It’s the pricing trajectory of every major CRM vendor’s last three release cycles.

The token math nobody runs

Forge a back-of-envelope before signing anything:

# token_budget.py — sanity check before signing
SEATS = 500
ACTIONS_PER_SEAT_PER_DAY = 40        # vendor says 15. trust nobody.
WORKING_DAYS = 220
AVG_INPUT_TOKENS = 8_000             # context-heavy CRM grounding
AVG_OUTPUT_TOKENS = 600
INPUT_PRICE = 3.00 / 1_000_000       # $/token (vendor passthrough +30%)
OUTPUT_PRICE = 15.00 / 1_000_000

annual_actions = SEATS * ACTIONS_PER_SEAT_PER_DAY * WORKING_DAYS
input_cost = annual_actions * AVG_INPUT_TOKENS * INPUT_PRICE
output_cost = annual_actions * AVG_OUTPUT_TOKENS * OUTPUT_PRICE
print(f"Annual AI burn: ${input_cost + output_cost:,.0f}")
# Annual AI burn: $1,455,600

Run that with your numbers before the AE runs theirs. If the vendor quote is below this, ask which assumption they shaved.

What CFOs ask now (and what to answer)

The CFO has stopped asking “what’s per-seat.” They now ask:

• What’s our cost per resolved case? Track it monthly. It should trend down with model improvements.
• What’s our cost per qualified lead? Outcome SKUs let you compare to traditional SDR cost.
• What’s the unit economics of AI vs human? Honest accounting, including model error tax.
• What happens at the renewal cliff? Year-3 pricing is the bear, not year-1.
• What’s the contractual exit? Termination-for-convenience, data portability, vector export rights.

If you can’t answer all five before signing, don’t sign.

Outcome pricing: the fine print

Outcome pricing sounds aligned. It often isn’t. Read for:

• Outcome definition. “Resolved” — by whom? Vendor model, customer CSAT, or 7-day no-reopen? Sierra defines resolution as “customer didn’t return within 14 days for the same issue.” Decagon uses session-level deflection. Different math, different bill.
• Floor commit. “Up to 10,000 resolutions / month” almost always means a minimum, not a maximum. Read the contract.
• Attribution windows. Outcome-influenced lead pricing requires an attribution rule. If the vendor sets it, the vendor wins.
• Reset clauses. If resolution rate drops below X%, do you still pay per attempt? Some vendors quietly switch you to per-call billing during quality dips.
• Audit rights. Can you replay the conversation logs and dispute? If not, walk.

Negotiating leverage by company size

• SMB (<200 seats). Almost no leverage. List price minus discount. Best play: shorter terms, walk-away willingness.
• Mid-market (200–2k seats). Some leverage on token ceilings and outcome definitions; little on unit price.
• Enterprise (2k–20k seats). Real leverage on structure, BYOM tier, multi-year ramps, and overage caps. Use it.
• Strategic (>20k seats or marquee logo). Custom contracts. Outcome SKUs become negotiable. BYOM passthrough below 20% markup is achievable.

If you’re in the lower bands, accept the structure and focus on cap clauses. If you’re in the upper bands, the structure itself is the negotiation.

Hybrid models — the actual 2026 norm

What most large customers are landing on:

• Seat base for license, identity, audit, and admin UX (the boring stuff that doesn’t scale with AI).
• Capped consumption pool for general agentic use (with hard kill switches, not soft alerts).
• Outcome-priced SKU for tier-1 deflection where the value is measurable.
• Bring-your-own-model discount where the vendor passes through cost rather than re-marking it.

This is the structure Agentforce 360, Copilot Studio Enterprise, and HubSpot’s Breeze Intelligence Suite are all converging toward. None of them will offer it on the first call. Ask explicitly.

The hybrid is harder to model upfront but easier to live with. It gives you escape velocity on the consumption tier (you cap it) while letting outcome value accrue (you measure it). The seat layer keeps your AP team’s spreadsheet looking familiar to the CFO.

The pricing dimension nobody negotiates: tier-mix

Vendors quote a single price per credit / token / outcome. Production usage is bimodal — most actions are cheap, a few are expensive. The vendor’s billing aggregates across the mix, but the unit economics differ wildly:

Action class	Tokens / action	Cost share
Field summarization	500–2,000	5%
Email draft	3,000–6,000	15%
Account research	15,000–40,000	35%
Multi-step agentic resolution	50,000–150,000	45%

If you only model “average action cost”, you’ll be off by 3x on the high-cost tier. Negotiate tier-based pricing if your usage skews high — many vendors will offer it under pressure, no vendor will offer it unprompted.

BYOM and the markup question

Every major CRM vendor now supports bring-your-own-LLM-key in some form — Salesforce via Einstein Trust Layer model gateway, Microsoft via Azure OpenAI passthrough, HubSpot via Breeze model selector, ServiceNow via Now Assist BYOM tier. The markup on vendor-managed inference ranges 30%–120% over raw provider rates. BYOM is worth pursuing if your spend exceeds ~$200k/yr on inference.

Tradeoff: BYOM means you own the Anthropic or OpenAI commitment, the rate limit headroom, the quota negotiation, and the failover. For enterprise tier customers already on direct contracts this is a no-brainer. For everyone else, vendor passthrough is fine until it isn’t.

The internal-clearinghouse model

Some larger orgs (banks, telcos, large retailers) have moved to an internal clearinghouse for AI consumption. Central FinOps team negotiates with vendors, owns the BYOM relationships, charges back to business units at a flat marked-up rate. Units don’t see direct vendor consumption pricing.

Pros: predictable BU budgets, centralized governance, leverage in negotiation. Cons: internal markups can grow opaque, BUs lose direct connection to unit economics, slower iteration.

Worth considering at >$5M annual AI spend. Smaller orgs should keep it simple.

Procurement playbook

What to negotiate, in priority order:

1. Token ceiling with auto-shutoff — soft alerts at 80%, hard cap at 100%, no overage billing without written approval.
2. Outcome definition baked into the order form — not the MSA, not the data sheet. Order form.
3. Quarterly true-up, not annual — gives you escape velocity if usage explodes.
4. Model-version lock — vendor can’t unilaterally swap you to a more expensive model mid-term.
5. Logged-replay audit rights for outcome-priced SKUs.
6. Multi-year ramp — year 1 at 40% commit, year 2 at 70%, year 3 at 100%. Lets you measure before you’re cornered.

Red flags

• “Average customer uses X tokens/month.” Average is the wrong stat. Ask for p95.
• “Outcome pricing scales with your success.” Sure — and the vendor’s success scales with your scale. Find the cap.
• “We’ll work with you on overage.” Get it in writing or assume zero leniency.
• “Token costs will come down.” They will. The markup won’t.

What each vendor’s pricing actually looks like in May 2026

A snapshot — pricing changes monthly, but the structural shape is sticky.

• Salesforce Agentforce. Flex Credits + Outcomes SKU. ~$2 per credit list, credits consumed per agent action with weight by complexity. Outcomes SKU charges per autonomous tier-1 resolution; floor commits typical.
• Microsoft Copilot Studio. Message packs (25k messages per pack) + per-action billing for autonomous agents. M365 Copilot seat at $30/user/mo separate.
• HubSpot Breeze. Tiered Breeze Intelligence credits, included floor per Hub seat, overage at credit rates. Breeze Agents priced per published agent + usage.
• ServiceNow Now Assist. Per-agent consumption units, bundled with Pro+ / Enterprise+ SKUs. BYOM tier available for large customers.
• Zoho Zia. Predominantly bundled into seat price; AI credits for higher-volume use cases.
• Freshworks Freddy. Per-seat AI add-on with bundled credits; overage at fixed rate.
• Intercom Fin. Pure outcome — per resolution, ~$0.99 default list.
• Sierra. Outcome-only — per resolved conversation, custom-negotiated.

The fault line: legacy vendors still anchor on seats with AI add-ons; AI-native vendors anchor on outcomes. The middle is hybrid.

Modeling outcome SKUs honestly

For a tier-1 service workload considering Sierra-style outcome pricing:

inputs:
  monthly_contacts: 80,000
  current_deflection_rate: 0.42       # human agent baseline
  expected_ai_deflection: 0.65         # vendor claim
  expected_ai_deflection_skeptical: 0.55  # discounted
  price_per_resolution: 1.00
  human_cost_per_resolution: 6.40
  
optimistic_monthly_ai_cost: 80000 * 0.65 * 1.00 = $52,000
skeptical_monthly_ai_cost: 80000 * 0.55 * 1.00 = $44,000
human_cost_avoided_optimistic: 80000 * 0.65 * 6.40 - 52000 = $281,200
human_cost_avoided_skeptical:  80000 * 0.55 * 6.40 - 44000 = $237,600

Even with the skeptical number, the unit economics work. But run the same model when AI deflection comes in at 0.40 — slightly below human baseline — and it’s still break-even because price-per-resolution is low. Outcome pricing can survive disappointing results in a way token pricing can’t. That’s the structural advantage.

The “AI rebate” trick

Watch for a clause that lets the vendor reclaim AI savings. Some agreements include a “value share” provision — if your tier-1 deflection rate exceeds X%, the vendor takes back a percentage of seat cost savings. It’s framed as “shared success.” It’s a clawback.

Strike it or cap it. Your headcount savings are yours.

Budgeting under uncertainty

You cannot accurately forecast year-1 AI consumption. Don’t try. Instead:

• Allocate a “discovery budget” of 20–30% of expected AI spend for unknown unknowns.
• Set quarterly true-up reviews on the calendar before the contract is signed.
• Maintain a kill switch (finops controls) you can flip without renegotiation.
• Track unit economics from day one: cost per resolved case, cost per qualified lead, cost per booked meeting.

After 90 days you’ll know your real consumption shape. Renegotiate then if the contract permits — and make sure it permits.

A note on multi-vendor stacks

The harder modeling problem in 2026 is cross-vendor unit economics. Many large customers run Agentforce on Salesforce, Copilot Studio on Dynamics, Breeze on HubSpot, plus a layer of in-house agents on top. Each has a different meter. Each invoice arrives separately. Nobody has a single dashboard for total AI cost per resolved case.

Build that dashboard. It’s the only number that lets you compare. The vendor will not build it for you — they’re not motivated to make their stack look more expensive than the competing one.

Bottom line

• Run your own token math before the vendor runs theirs — assume 2.5x their quoted average.
• Outcome pricing is only aligned if you control the outcome definition.
• The 2026 winning structure is hybrid: seats for the boring layer, capped consumption pool, outcome SKU only where deflection is measurable.
• BYOM pays off above ~$200k/yr inference; below that, vendor passthrough is fine.
• Negotiate token ceilings with auto-shutoff before you negotiate price.