Nobody Counts Conversations

I have priced enterprise software and I have bought it, and the argument I had with myself across the table took about a decade to resolve.

At the Irvine Company I was buying real estate software while the industry moved off user-based pricing.

At MRI Software I was on the other side, setting the value metrics.

Commercial products priced by lease. Residential by unit. Some by square foot, some by contract. Whatever the countable thing was that drove how the product delivered value to the customer, that was the metric, and we used it wherever the product allowed.

That transition is mature now. Value-based pricing is the market expectation in real estate software, and the vocabulary is settled enough that a buyer can read a quote and know what it means.

Every vendor has a pricing metric, because every vendor has to bill on something. Only some of them have a value metric, which is a pricing metric that also tracks what the customer is getting. An entire vertical worked out the difference, agreed on it, and named the results. Buyers compare quotes. Vendors explain their pricing in one sentence.

The AI pricing market has not reached step one.

The vendors cannot pick one either

Salesforce runs three models at once and Microsoft runs four

Start with what is actually being quoted, because it is worse than five vendors disagreeing.

Salesforce's Agentforce pricing page, checked in September 2026, offers two consumption meters and four per-user price points. Flex Credits at five hundred dollars per hundred thousand, where a standard action draws twenty credits and a voice action thirty. Conversations at two dollars each. Then a user license at five dollars a month that still requires Flex Credits. Add-ons at a hundred and twenty-five, industry add-ons at a hundred and fifty, and full editions from five hundred and fifty. Three buying models sit on top of that.

Flex Credits and Conversations cannot run in the same org, so a buyer picks a meter and lives inside it. Unused credits do not roll over.

One product. One page. Six prices and three ways to pay for them.

Microsoft lists four ways to pay on a single page. The Microsoft 365 Copilot enterprise add-on at thirty dollars per user per month, paid yearly. Copilot Studio credit packs at two hundred dollars a month for twenty-five thousand credits, sold tenant-wide rather than per seat. A pay-as-you-go meter with no upfront commitment. And a pre-purchase plan of prepaid Commit Units at up to twenty percent off. Copilot Credits replaced per-message billing in September 2025, so the metric changed underneath customers already using it.

Microsoft describes the charge in its own words: whenever an action or response is completed by an agent, a varying number of Copilot Credits will be billed depending on the specific usage.

That is the vendor telling you, on the pricing page, that the bill is not predictable.

Salesforce goes further and uses the language of value directly. Its page says Flex Credits "scale up and down with usage to align cost to value."

Read that against what a credit measures. A standard action is twenty credits. A voice action is thirty. Salesforce defines what counts as an action, its platform executes them, and the customer pays per action. The word value is doing work the metric cannot.

Oracle went the other direction. It announced twenty-two Fusion Agentic Applications at the Oracle AI World Tour in London on 24 March 2026, and tied them to Fusion Cloud release 26B in May. Oracle states that its AI Agent Studio is available at no additional cost, and its service descriptions list a separate Agentic Applications subscription, part number B112535, that carries the right to develop, test and use them.

That subscription carries a pooled allocation of two and a half million AI Units a month, on top of the twenty thousand a month a Fusion production environment already gets, and every use of an agent draws them down.

Sierra prices per outcome, resolution among them, which is the only genuine attempt at a value metric in this group. It publishes no pricing page, no tiers and no rates. Reported figures run between one and two and a half dollars per resolution and the company has never confirmed them. A 2024 post on its own blog floats a blended model where routing and greetings may bill per conversation and other interactions bill per outcome, with the definition of a successful outcome set in each contract.

Forethought sells into the same category on the same metric. Its pricing page carries three tiers and a Get a Quote button where each number should be.

Salesforce has since added one of its own. Its rate card of June 17, 2026 prices a Help Agent Resolution at 400 credits, billed only when the agent resolves the case. The same card cut prompt multipliers sharply from the October 2025 version, which is what a metric repricing underneath live customers looks like.

Notice which vendors those are. Per resolution comes closest to passing all three tests of any metric on offer. Salesforce now prices one at a published rate. The two vendors that sell on it exclusively publish nothing at all.

Which is the second half of the bargain real estate made. Named metrics were necessary and not sufficient. What made the trade work was that the numbers were published, so a buyer walking into a renewal knew the market rate. A metric nobody quotes delivers value to the vendor and nobody else.

Comparing two approaches

Those are not four bids for the same work, and it is worth saying so before anyone objects. Copilot is a horizontal assistant across Microsoft 365. Oracle's agents are function-specific applications inside Fusion. Agentforce and Sierra both aim at customer service. Nobody runs all of them through one evaluation.

So narrow it to the two that genuinely compete, and watch what a buyer has to do.

Before comparing anything, they have to pick which Salesforce meter to model. Conversations or Flex Credits, and the choice binds the whole org. Getting it right means forecasting how many actions an average conversation will consume, on their own data, before the agent is live. At those published rates, twenty actions is the crossover point where the two meters cost the same.

Then they need a number from Sierra, which publishes none. That number arrives at the end of a sales cycle, attached to a contractual definition of a resolution written for that deployment.

Only then can they compare a per-action forecast they built themselves against a per-outcome rate they negotiated, using a resolution rate neither vendor will commit to.

The honest bake-off, between two direct competitors selling the same thing, has no arithmetic in it. There is a forecast, a negotiation, and a hope. And hope is not a strategy.

The cross-category version is the one a CIO actually faces. A thousand employees on the Microsoft 365 Copilot enterprise add-on runs three hundred sixty thousand dollars a year, independent of usage, before any Copilot Studio credits for agents that reach outside the tenant. Twenty-two agents inside Fusion cost nothing beyond licenses already paid, until the AI Units run out. Agentforce costs whatever the meter you picked turns out to produce. Those land in the same budget, in the same fiscal year, and there is no way to reason about which one earns the next dollar.

A procurement team can compare prices. It cannot compare these.

What separates a pricing metric from a value metric

Three tests: counted, controlled, and tracking the customer's value

Tokens are countable. So are conversations. Countable is the low bar, and clearing it is what makes something billable rather than useful.

The real bar is higher. A customer should not have to invent a metric in order to buy a product. A property manager already knows its units, leases and square feet. An insurer already knows its claims, policies and premiums. A lender already knows its loans, balances and accounts. The vendor is not introducing a new unit of economic reality. It is attaching a price to one that already exists.

A pricing metric becomes a value metric when it passes three tests at once. Call it the value metric test.

• The customer already counts it. A commercial operator knows its lease count. A residential operator knows its unit count. That number exists on their books for reasons that predate the software, and it was there before the vendor showed up.

• The customer controls it. Units arrive through acquisition or development. Leases get signed. Those are deliberate decisions made by people in the business who intended to make them. When the software bill moves, it moves because the business chose to grow.

• It tracks the customer's value. More units means more of the work the software does. The bill and the benefit move together, which is the entire argument for value-based pricing and the reason buyers eventually accepted it.

Run them through those tests. Per token. Per conversation. Per action. Per user. Per resolution. Bundled with the license.

None of them passes all three.

Plenty of businesses count conversations. Contact centers have counted contacts, tickets and interactions for decades. What none of them do is treat a conversation as the underlying unit of economic value the way a lease, a unit, a claim or a loan is. It is an activity count, not an economic one, and nobody sets a budget or values a portfolio on it. The customer does not control token consumption the way it controls its leases, units or claims. The model's behavior and the task determine token consumption, and a more verbose model version costs you more for identical work. Worse, you can improve a process and pay more for it, because efficiency measured in your terms and efficiency measured in tokens are different things. Per user is the pricing metric the industry spent years moving away from, reappearing at the exact moment the users stop being the ones doing the work.

Per resolution is the only one of those that seriously attempts the value side, and it deserves the credit. It also puts the vendor in the position of grading its own homework on what counts as resolved.

Telecom similarly evolved for thirty years

Minutes and texts metered the carrier's costs, and then they stopped

Long distance was billed by the minute, because circuits were scarce and minutes were what a carrier's costs were made of. Text messages were billed one at a time, because they rode a signalling channel the carrier had already built. Roaming carried its own charge, because settlement between carriers was real money.

Every one of those charges metered something expensive for the seller. Not one of them metered anything a customer wanted more or less of.

So customers counted. They watched the clock on long distance calls. They rationed texts. Families argued about who had run up the overage.

Then it stopped. Unlimited did not win on price, and the carriers did not stop metering because bandwidth got free. It changed the relationship. The carrier absorbed the variance instead of billing the customer for it, and the market followed.

The result is on your phone bill right now, and the line item that used to bill for minutes does not exist.

Tokens are minutes. They meter the vendor's cost of goods and nothing else. Conversations are text messages: a pricing metric that exists because it is easy to bill, and nothing more than that.

The trade buyers are being asked to make

What a buyer gives up, and what real estate got in return

Here is the part I got wrong for a while, sitting in the buying seat.

User-based pricing has one real virtue: headcount is a number the buyer sets. You know it a year out. You can defend it in a budget, and you can reap the benefits of efficiency, because automating a process cuts seats and the savings land on your side of the table. Moving to value-based pricing means giving that up, and it took the real estate market years to accept the trade.

What made it acceptable was the other side of the bargain. Real value metrics, named. Published benchmarks. Two vendors quoting per unit per year compare line by line, and a buyer could walk into a renewal knowing what the market rate was.

The AI market is asking for the same concession and offering none of that in return. No agreed metric. No benchmark. No way to know whether two dollars a conversation is expensive, because there is nothing to hold it against.

That is not a pricing problem. It is a market maturity problem. If it follows the earlier pattern, it resolves slowly, through buyers refusing to sign things they cannot compare.

Buyers have already said this out loud

IDC ran its Future Enterprise Resiliency and Spending Survey in March 2026 with 903 respondents, and published what enterprise buyers most want from AI vendors.

Simple, flexible and transparent pricing for AI capabilities ranked third, behind an AI roadmap with agentic extensibility and information on security and data privacy, and ahead of the value modeling and ROI justification tools vendors built specifically to close these deals.

The same survey found that when evaluating agentic AI specifically, more than sixty percent of buyers rank transparency around business, IT and security risk as their number one requirement, above productivity. IDC states plainly that this challenges the assumption that automation and productivity should always lead these conversations.

Two questions for the vendor

What the metric costs when the agent is wrong, and what else you could pay on

Every vendor will quote you a metric. Run the test on it, then ask two questions.

Ask what the metric costs you when the agent is wrong.

Per conversation, you pay for the failure and then you pay a person to fix it. Per token, a chattier model version raises your bill for identical output and you had no say in the version. Per user, you pay for people whether or not they are the ones doing the work. Per resolution, the vendor bears more of the cost of failure, which helps explain why that model can price higher and why the definition of resolved repays a close reading. Bundled free with the license, you are paying in the currency of not being able to leave.

Then ask the second question, which is the one that actually moves a negotiation.

Can you price this on something I already count?

Your business still counts what it counted before. Leases. Units. Square feet. Contracts. Claims. Tickets. Loans. Those numbers did not stop existing because the work behind them got automated, and they remain an honest measure of the economic activity the software supports.

Watch which vendors can answer that and which ones explain why their metric is better. The answer tells you whether you are buying a product with a business model or a metering scheme with a product attached.

Where this ends

Two exits, and why domain expertise decides which vendors take the second

The special counting stops. Businesses will always count things, and the ones worth counting were already being counted before any vendor arrived. What ends is the separate tally kept only because a bill demands it. That part is not in doubt, because metering the seller's cost structure has not held anywhere else for long.

There are two ways out and the market has already shown you both.

One is unlimited, or close enough. Bundle the agents into the license and stop asking the customer to count. Oracle looked like it took that direction in March, bundling twenty-two agentic applications into Fusion rather than metering them. It read as generosity. Then read the contract. Oracle's Fusion Cloud service descriptions, version 091026, effective 10 September 2026, price a separate Agentic Applications subscription, metered in AI Units, and say the units "are counted for all use of AI agents and Agentic Apps, whether during development, testing or in production use, and likewise are counted for both Your Production and any associated non-production environments." Oracle did not end the counting. It moved it to a unit the customer has never counted before, and it counts in non-production environments too.

The other is repricing onto something the customer already counts. Per lease. Per unit. Per claim. Per loan. That is where real estate software landed, and it is the road the AI market has not seriously walked down yet.

That road requires more domain expertise than most of these vendors currently have. Pricing by lease means knowing what a lease is worth to the customer and how many of them there are. Pricing by claim means knowing an insurer's book. A value metric is domain-specific by definition, and a vendor selling the same product into twelve industries has twelve different metrics to learn.

Per token requires none of that. The same metric works everywhere, which is exactly its appeal and exactly its defect.

Both endings share a mechanism. The special counting becomes the vendor's problem instead of the customer's. What separates them is who carries the variance and what the buyer can compare at renewal.

IDC has put a date on part of it. In its FutureScape predictions it says that by 2028 pure seat-based pricing will be obsolete as AI agents rapidly replace manual repetitive tasks with digital labor, forcing seventy percent of vendors to refactor their value proposition into new models.

Read that alongside what Salesforce and Microsoft are doing now. Both have added per-user options for agents in the last year. The metric IDC expects to disappear is the same one vendors are reaching for when consumption pricing proves too hard to forecast.

That is a market trying two directions at once, which is what the end of a pricing era looks like from the inside.

I would not bet on tokens and conversations surviving this either. They are minutes and texts, and we know how that went.

In the meantime I would be cautious about signing multi-year commitments denominated in a pricing metric that may not survive the next stage of the market, and start asking every vendor in the process whether they can price on something your business already counts.

Vendors who can answer are further along than they are letting on. The rest are telling you where their costs sit.

There are four seats in the name. The fifth one is yours.