One agentic run can bill as 150 assists instead of 1, so the unit definition moves more money than the rate ever will

ServiceNow's published guidance charges 25 assists for an agent execution using 0 to 4 tools and 150 assists for one using 9 to 20, and Salesforce bills a single $0.10 action three times once it crosses 20,001 tokens. Every one of those multipliers sits in the denominator, not the rate card, which means a 10% discount you win on price can be erased by a tool-count band you never read. Redline the unit first, then argue rate.

Prepared by Redress Compliance · September 4, 2026 · GenAI and agent licensing advisory. Salesforce, ServiceNow, Microsoft, Zendesk and Intercom engagements, 2025 to 2026.

Put the four live agent meters next to each other and the pattern is obvious: the published price per unit is the least interesting number on the page.

Salesforce sells an action at 20 Flex Credits, or $0.10, but wraps a 10,000-token ceiling around it, so an action that consumes 15,000 tokens bills as two and one that crosses 20,001 tokens bills as three. Nothing in that escalation touches the rate card.

ServiceNow goes further and multiplies openly: its June 2026 guidance bands an agentic execution at 25 assists for 0 to 4 tools, 50 assists for 5 to 8, and 150 assists for 9 to 20.

And it publishes no universal dollar price for an assist at all, which means the only number you can negotiate is a pool size you cannot yet forecast.

Microsoft's rename from messages to Copilot Credits in September 2025 changed the label and not the meter, at $0.01 pay as you go or roughly $0.008 prepaid in $200 packs of 25,000, while a single tenant-grounded reasoning response has been measured at 112 credits.

Zendesk and Intercom sell outcomes rather than actions, which sounds safer until you read who decides an outcome occurred.

Treat these as four different denominators, not four different prices, and use the agent contract redline checklist to attack the definition before anyone opens a discount conversation.

The same logic runs the other way on voice. If both Voice Minutes and Voice Actions are enabled, Voice Minutes takes precedence and Voice Actions are disabled, so the meter you thought you negotiated silently stops applying.

Write the elected meter into the order form by name and add that any change of meter requires your written consent at a price no higher than the elected meter would have produced on the same volume.

Agent deals break the muscle memory that procurement built over twenty years of seat licensing. In a seat deal the vendor controls the price and you control the count: you know how many people you employ, you can freeze headcount, and a 15% discount on a known quantity is a real 15%.

In an agent deal the vendor controls the price and the counting rule. That second lever is worth more than the first, and every vendor sales leader knows it, which is exactly why they concede on rate so cheerfully.

Watch the choreography. You open on price. The rep resists for two calls, escalates for approval theater, and lands you at 10 to 20% off the credit or resolution rate.

In my experience across these deals, consumption growth recovers that concession inside two quarters, because agent volume in year one is a fraction of year two and the vendor's own forecast already assumes it.

The discount is real, the recovery is faster, and nothing in the contract stops the recovery from accelerating.

Now count the levers the vendor keeps. Tool-count bands mean the same business outcome bills 25 assists or 150 depending on an integration decision your architect makes six months after signature. Token ceilings mean a wordier prompt template triples the cost of an action with no notice.

Retry and failure billing means you pay for the model's mistakes unless the contract says otherwise. Non-production metering means your test cycles burn the pool you sized for production.

Resolution windows and premium message classes mean a unit can be recategorized upward without anyone amending the agreement. Every one of those is a price increase that requires no signature, no renewal, and no conversation.

That is the structural asymmetry: your discount needed a negotiation, their increase does not.

The denominator is negotiable, and Zendesk proved it. Its May 2026 restructure split one billing bucket into three, made Assisted Escalation and Contained Resolution free, and billed only Verified Resolution.

Under the model it replaced, a conversation the customer simply abandoned could be billed as a resolution. That change was not generosity, it was the aggregate weight of buyers refusing to pay for outcomes they could not verify.

Compare Intercom, where assumed resolutions still count when a customer exits without replying, and where Intercom cites roughly 71% resolution against independent reports in the 42 to 50 percent range.

That 20-plus point gap is not a performance debate, it is a definitional one, and it lands entirely on your invoice.

The asymmetry that decides all of this is telemetry. The vendor measures your consumption in real time, at unit granularity, across production and sandbox.

You get a monthly report in the vendor's own format, built from the vendor's own definitions, with no independent way to reconstruct why one execution billed 25 assists and the next billed 150.

ServiceNow publishes no universal assist price and no universal overage rate, which means even your unit economics are inferred rather than known. When a definition is ambiguous, it is resolved by whoever holds the meter, and that is never you.

So invert the sequence. Fix the unit, the ceiling, the retry treatment, the non-production exclusion and the reclassification bar first, then argue rate, and expect the rate concession to shrink once the vendor can no longer earn it back through the counting rule.

A tighter denominator with 8% off beats a loose denominator with 22% off in every twelve-month model I have run.

Pair the definitional work with a price hold clause that survives a repricing year, because a fixed unit definition and a fixed rate together are the only combination that produces a forecastable bill.

Eight definitions decide whether your committed pool lasts twelve months or seven. Write them as defined terms in the order sheet, not as aspirations in an email thread.

First, the billable unit: one user-initiated request equals one unit regardless of internal tool calls, retrieval hops, reasoning steps, or model invocations.

That single sentence is the difference between ServiceNow's published 25 assists for a 0 to 4 tool execution and 150 assists for a 9 to 20 tool execution, a 6x swing you never priced.

Second, measurement method and source of truth: the vendor's telemetry is the meter, but you get raw exportable logs at unit granularity, and disputed units default to your reading until reconciled.

Third, included usage: state the pool numerically, state that it resets annually on the contract anniversary (ServiceNow's does), and negotiate rollover of unused units rather than accepting expiry.

Fourth, overage rules and a hard cap: Microsoft's published overage runs $0.01 per standard message and $0.02 per generative, a 2x reclassification risk sitting inside one meter; cap total overage at a stated dollar figure per quarter and require written notice before any burn above 80% of pool.

Fifth, minimum commitment, ideally a floor you can hit in month nine, not month one. Sixth, retries and failures: a failed or timed-out execution is not billable, and a retry of a failed execution is not separately billable.

Silence here defaults against you because token and credit plans bill resources consumed regardless of outcome. Seventh, reporting cadence: monthly, machine-readable, unit-level.

Eighth, service levels tied to the meter: if the agent fails its accuracy or latency SLA, those units are credited, not just apologized for. Token ceilings belong in the same list. Salesforce's $0.10 action assumes 10,000 tokens; 15,000 tokens bills as two, 20,001 bills as three.

Fix that ceiling numerically and bar downward revision mid-term.

Outcome models need a ninth definition: a successful outcome requires an affirmative user signal and expressly excludes abandonment, because Intercom counts assumed resolutions when a customer exits without replying.

And the gap between vendor-reported 71% resolution and independent 42 to 50% is definitional, not technical.

Expect three counters. The vendor will offer a rate discount instead of a unit definition, will claim the meter is "product behavior" and not contractual, and will propose a side letter that expires at renewal. Reject all three.

A strong outcome looks like this: unit defined as one user request, retries and failures excluded, token ceiling frozen at signature, pool rollover of at least 25%, overage capped, and every one of those terms carried into the renewal by reference in the master agreement rather than the order form.

Pair this exhibit with the price hold clause language that survives a repricing year so the rate you fixed and the denominator you fixed expire on the same date.

Volume growth is forecastable. Reclassification is not, and it is where the term-length money actually goes. The vendor does not need to raise your rate to double your bill; it needs only to move traffic you are already sending into a class that costs more.

Microsoft's premium message category is the clearest exhibit: practitioners report it surprises most enterprise deployments precisely because the categorization is not surfaced in the public pricing summary or in the Power Platform admin center. You cannot audit what the console will not show you.

Layer on standard overage at $0.01 versus generative at $0.02 and the reclassification of a single high-volume workflow is a 100% increase on that stream with no rate change and no amendment.

The September 1, 2025 rename from messages to Copilot Credits is the pattern in miniature: pack size unchanged, meter rate unchanged, label changed. Renames are harmless. Reclassifications are not, and both arrive in the same release note.

The redline is three sentences and it is worth more than most rate concessions. First: the vendor may not unilaterally reclassify any unit consumed by Customer into a premium, generative, or higher-consumption class during the term, and any such reclassification requires a signed amendment.

Second: tool-count bands, token ceilings, and channel scope are frozen at their signature-date values and listed in an exhibit with the numeric thresholds spelled out (0 to 4, 5 to 8, 9 to 20 tools, and the corresponding multipliers).

Third: any new consumption category, unit type, or premium tier introduced mid-term is billed at the Customer's then-existing rate for the nearest equivalent unit until the next renewal. That third sentence is the anti-drift clause, and it is the one that survives product roadmaps you have not seen.

Channel scope is the quiet version of the same problem.

Confirm in writing whether the per-resolution or per-action rate covers email automation as well as web chat, because scoping to chat alone leaves email to be repriced later as a "new category." Salesforce adds a mutual exclusivity trap on top: enabling Voice Minutes takes precedence and disables Voice Actions.

So you are forced to pick a meter, and voice actions bill 30 credits ($0.15) against 20 credits ($0.10) for standard.

Price both paths before signature, not after your contact center picks one. Vendors will resist the freeze by arguing the product is evolving too fast to fix bands. Answer that you will accept new categories at renewal, priced then, with your existing traffic protected in the interim.

In practice that trade closes, and it pairs directly with capping the AI uplift at renewal so drift and increase are both bounded.

Two clauses get skipped in nearly every agent deal I see, and both of them are worth more than the rate concession the buyer spent three calls chasing. The first is scope of environment.

ServiceNow's own guidance is explicit that the contracted assist pool is monitored across production **and** non-production instances, which means your regression suite, your load tests.

Your scheduled validation jobs and your sandbox refresh cycles all burn against the same commercial pool as live customer work.

That is not an accident of implementation, it is a deliberate denominator choice, and it converts your own QA discipline into vendor revenue.

Demand a written exclusion for all non-production instances, UAT, developer sandboxes, automated test harnesses and scheduled validation runs, with the vendor obliged to tag those instances at provisioning rather than leaving you to reconcile after the fact.

If the vendor refuses a full exclusion, take a defined non-production allowance of 15 to 20% of the production pool at zero cost, which is roughly what a mature test estate consumes in my experience across ServiceNow and Microsoft agent deployments.

The second clause is visibility plus a hard ceiling. There is no single native dashboard showing all Copilot Credit consumption across a tenant, a single tenant-grounded reasoning response can burn 112 credits, and an autonomous agent can trigger itself thousands of times overnight.

Zendesk's January 2026 change auto-bills overage with no notification, no cap and no grace period. Put together, you have a meter you cannot see, a workload that self-multiplies, and a billing engine that charges without asking.

Ask for four things: tenant-level daily burn reporting delivered by API or scheduled export, a hard spend ceiling with a **vendor-side throttle** at 100% of the committed pool, 72-hour anomaly notice when consumption exceeds 150% of trailing 30-day average.

And an express prohibition on auto-billing above the committed pool without written approval from a named signatory.

The vendor will counter that throttling degrades service and offer alerting instead. Alerting is not a control, it is a receipt.

Hold the throttle and trade it against a modest overage rate rather than against the cap itself, and pair it with the wider set of GenAI and agent contract redlines so the ceiling is not the only clause carrying the weight.

Five patterns repeat across every rate card and every engagement I have worked in the last eighteen months.

Units get renamed but not repriced: Microsoft moved from "messages" to "Copilot Credits" on 1 September 2025 with pack size and meter rate unchanged, which is a naming exercise that resets buyer familiarity without resetting cost.

Tool-count and token bands live in guidance documents rather than in the contract, so ServiceNow's 25/50/150 tiers and Salesforce's 10,000-token action ceiling can be revised without a signature.

Pools expire annually on the contract anniversary with no rollover, converting overbuy into pure vendor margin.

Overage rates for ServiceNow assists and Zendesk resolutions are unpublished, which means they must be extracted in writing before signature or they will be set unilaterally when you are already dependent.

And vendor-reported outcome rates run about 20 points above independent measurement, so any outcome-priced commitment sized on vendor data is oversized from day one.

The direction of travel favors buyers who push.

Zendesk narrowed its own denominator in May 2026 under buyer pressure, splitting the single bucket so Assisted Escalation and Contained Resolution are free and only Verified Resolution bills, while dropping the Advanced AI Agents add-on and capping included resolutions at 5 to 15 per agent month.

Salesforce's roughly $3.6bn Intercom Fin acquisition in June 2026 now leaves two competing outcome definitions ($0.10 actions with token multipliers, $0.99 billable outcomes including assumed resolutions) inside one vendor, and that inconsistency is leverage you should name in the room.

Sequence the unit fight ahead of the rate fight, then lock the result with a price hold clause that survives a repricing year.

Related advisory: the full topic guide.

What is the difference between an assist, an action and a credit?

They are three vendor-specific denominators for the same idea, priced very differently. A Salesforce action is 20 Flex Credits, or $0.10, with a 10,000-token ceiling above which one response bills as two or three actions.

A ServiceNow assist is banded by tool count, with 25 assists for a 0-to-4-tool agentic execution and 150 for a 9-to-20-tool one, and no published dollar rate. A Microsoft Copilot Credit is $0.01 pay-as-you-go or roughly $0.008 prepaid in $200 packs of 25,000.

Do vendors bill for failed agent tasks and retries?

Usually yes, unless you write otherwise. Token, API, compute and credit-based plans generally charge for resources consumed even when the task fails, while outcome-based contracts charge only after a defined successful result.

Silence defaults against the buyer, so the agreement must state explicitly that failed or timed-out executions are not billable and that a retry of a failed execution is not separately metered.

Is non-production and sandbox usage metered?

On ServiceNow it is. Published guidance confirms the contracted assist pool is monitored across production and non-production instances, and that consumption grows as testing and scheduled activity grow.

Excluding development, test, regression and load-testing traffic from the metered pool is one of the highest-value redlines available and typically protects 10% to 20% of first-year consumption.

How much does tool count actually change the bill?

By up to 6x on identical business output. ServiceNow's June 2026 guidance prices one agentic execution at 25 assists for 0 to 4 tools, 50 for 5 to 8, and 150 for 9 to 20.

Because agents accumulate tools over a contract term, an unbanded contract gives the vendor a built-in consumption escalator with no price increase and no notice requirement.

What is wrong with outcome-based pricing like per-resolution billing?

The definition of the outcome is written by the vendor. Intercom Fin bills $0.99 per billable outcome and counts an assumed resolution when a customer exits without replying, which is why vendor-cited resolution rates near 71% appear at 42% to 50% in independent measurement.

Zendesk improved on this in May 2026 by making Assisted Escalation and Contained Resolution free and billing only Verified Resolution, which proves the definition is negotiable.

Can a vendor reclassify my usage into a more expensive category mid-term?

Yes, unless you block it. Microsoft's premium message category is not surfaced in the public pricing summary or the Power Platform admin center, and standard overage runs $0.01 against $0.02 for generative messages.

Insert language prohibiting unilateral reclassification of any unit into a higher-consumption class during the term, and require any new category introduced mid-term to bill at the existing rate until renewal.

What overage protection should I demand before signing?

A hard spend ceiling, a written overage rate, and no automatic billing above the commitment.

Zendesk's January 2026 change auto-charges overage at the per-resolution rate with no prior notification, no cap and no grace period, and third parties place the rate near $1.50 on committed volume and $2.00 pay-as-you-go.

Cap overage at the committed rate plus 10% and require written approval before any spend above the pool.