Agentforce vs Fin vs Breeze: Three Meters That Do Not Measure the Same Thing

Three platform vendors sell customer support agents on three different meters. Fin bills $0.99 per outcome, HubSpot Breeze bills $0.50 per resolved conversation, and Salesforce Agentforce bills $0.10 per action. Anyone comparing Agentforce vs Fin vs Breeze on those three numbers is comparing three different units, and the 10x spread mostly disappears once you fix that. What survives the fix is the interesting part.

The three support agent prices are not three prices

A buyer reading three vendor pricing pages in one afternoon sees $0.99, $0.50 and $0.10. That looks like a ten-to-one range for the same job. It is not, because none of the three vendors is selling the same billable event.

This is the most common mistake in AI support procurement right now, and it is not the buyer's fault. Every vendor picked the unit that makes its own product look disciplined, and none of them publishes a conversion into anyone else's unit. Comparison is left as an exercise for the person signing the contract.

Fin bills for outcomes, including the ones it did not finish

Fin's published pricing lists four billable outcomes at $0.99 each, plus lead qualification at $9.99. A resolution is defined as no further help being requested after Fin's last answer. A procedure handoff, where Fin completes a workflow you configured to end at a human, is also billable. So is a disqualification, where Fin decides a prospect does not match your criteria.

Read the resolution definition again. It bills on silence, not on confirmation. A customer who gives up and closes the tab produces the same invoice line as a customer whose problem was fixed. The meter cannot tell you which of the two you paid for.

There is a fair defence of this. Confirmation-based billing requires the customer to answer a survey, most customers do not, and a meter that only fires on replies would undercount real work badly. Silence is a proxy that errs in the vendor's favour, but every available proxy errs somewhere.

Agentforce bills for work, not results

Salesforce sells Agentforce capacity as Flex Credits at $500 per 100,000 credits, with a standard action costing 20 credits. That is $0.10 an action, and $0.15 for a voice action. The meter runs on work performed, with no reference to whether the customer went away satisfied.

The older per-conversation model at $2.00 is still available. Salesforce did not remove it, which matters for the arithmetic below. Two published prices from the same vendor for the same job let you infer what that vendor privately thinks a conversation costs.

What Agentforce, Fin and Breeze each count as a resolution

The billing unit is the visible half of the decision. The definition sitting behind the unit is the half that moves your invoice, and it is written in documentation nobody reads before signing.

The 72-hour clock does not run backwards

HubSpot is the most precise of the three, and its precision is worth reading closely. A Breeze conversation counts as resolved when the agent shared a content source or performed an action. There must also be no qualifying handoff to a human within 72 hours of the last visitor response.

Then comes the sentence that should change how you model this. HubSpot states that subsequent actions, including messages from a live agent, a transfer request or negative feedback, will not change the resolution status.

That is a defensible engineering choice. Billing needs a settled state, and an open-ended clock never settles. It also means a conversation your customer rated badly on day four is still on your bill as a success, priced at 50 credits.

HubSpot moved Breeze Customer Agent to this model on 14 April 2026, cutting the rate from $1.00 per conversation to $0.50 per resolved conversation. On its own numbers that is a price cut for most buyers, because the old meter charged whether the agent resolved anything or not. The mechanics of the credit conversion are covered in the breakdown of what HubSpot's credit repricing actually changed.

Rates are the vendors' published list prices as of 24 August 2026. Enterprise contracts are negotiated, and none of these three publishes its volume discount schedule. Treat the rate column as a ceiling, not a quote.

Normalising Agentforce vs Fin vs Breeze to one unit

To compare the three you need a common denominator. The only one that works is cost per inbound conversation the agent handles, because that is the volume figure you already have in your helpdesk.

Salesforce's own two prices set the crossover at 20 actions

Salesforce publishes $0.10 per action and $2.00 per conversation for the same product. Divide the second by the first and you get 20. That is Salesforce's own implied estimate of how many actions a support conversation takes, and it is a far better anchor than any consultancy's guess.

The inference is worth stating plainly, because it does the work of a benchmark nobody has published. A vendor offering two meters prices them near indifference, otherwise buyers would arbitrage the gap immediately. Below 20 actions, credits are cheaper. Above it, the conversation price is.

Now take 1,000 inbound conversations. Assume 65% are handled by the agent without a human, which is the rate HubSpot reports for Breeze Customer Agent across more than 8,000 activated customers. Apply it to all three for comparability, which is generous to nobody in particular.

Breeze bills 650 resolutions at $0.50, so $325. Fin bills 650 resolutions plus, say, 100 configured procedure handoffs at $0.99, so $743. Agentforce bills every action on all 1,000 conversations, including the 350 it lost, which is $800 at eight actions each and $2,000 at twenty.

The inversion holds for a structural reason rather than a pricing accident. Fin and Breeze charge for a subset of conversations, Agentforce charges for a multiple of them. A per-action meter is the only one of the three where a failed conversation still generates revenue for the vendor, and failed conversations are roughly a third of the volume.

That does not make Agentforce a bad purchase. It makes the $0.10 figure the least informative number on any of the three pricing pages. The Agentforce conversation-cost breakdown works through the full arithmetic for Salesforce, including the licence floor sitting underneath the credits. The wider pattern is covered in the piece on bundling against metering.

One practical warning about the arithmetic. Every one of these meters is sensitive to a variable you control, and it is a different variable in each case. Breeze rewards you for escalating early. Fin rewards you for keeping procedures out of the configured-handoff category. Agentforce rewards you for scoping the agent so tightly that it takes fewer steps. Optimise for the wrong one and you will pay for the privilege.

Deflection quality is the number nobody publishes independently

Price per unit only matters if the units are comparable in quality. Here the evidence base gets thin, and the honest answer is that no independent, cross-vendor benchmark of support agent quality exists in public.

Every published resolution rate is vendor-run

Fin's headline claim is a 76% average resolution rate across its customer base, which Salesforce repeated in its acquisition announcement. Its Apex 1.0 model scored 73.1% on Intercom's own benchmark, against 71.1% for GPT-5.4 and 69.6% for Claude Sonnet 4.6. HubSpot reports 65% resolution and a 39% cut in resolution time across more than 8,000 customers.

Three vendors, three numbers, three definitions of the thing being counted, and three vendor-run measurements. That is not a comparison. It is three press releases printed next to each other.

VentureBeat's coverage of the Apex benchmark made the limitation explicit. Intercom declined to disclose the base model and did not publish a sample size. The result may well be genuine. It is also unfalsifiable as published, which is the wrong side of a trade for a claim this load-bearing.

There is a second problem underneath the first. Fin's 76% and HubSpot's 65% are averages across thousands of deployments with wildly different knowledge bases. An average across that much variance tells you almost nothing about your own likely result, because the driver is not the model. It is how much of your product knowledge is written down, how deep the integrations go, and how narrowly the agent is scoped.

My position is that you should ignore all three headline rates in a bake-off. Run each agent against 200 of your own historical tickets, grade the transcripts yourself, and treat the vendor number as marketing until your own sample contradicts you. Two weeks of that work is worth more than any published benchmark, and it is the only measurement anyone will ever let you audit.

Escalation is where the three support agents diverge most

Every deployment escalates. The design question is what happens in the seconds around the handoff, and the commercial question is who pays for it. The three vendors answer both differently.

Fin is the only one of the three whose published price list contains a line for a handoff. A procedure handoff bills at $0.99, the same as a resolution. The logic is coherent, because completing a configured triage workflow is real work that saved a human real time. It is still the case that Fin can invoice you for a conversation it could not finish.

Breeze is the opposite. A qualifying handoff inside 72 hours removes the charge entirely, which aligns HubSpot's revenue with your outcome more tightly than either competitor manages. The catch is the direction of the clock. Escalate late and you pay anyway, and nothing you do afterwards reverses it.

Agentforce is indifferent. The actions consumed before the handoff are billed exactly as they would be in a resolved conversation, so the meter carries no signal at all about whether the agent helped. For a buyer that is the cleanest possible separation of price from performance, and it is worth understanding before you sign, alongside the contract clauses worth arguing over on any agent purchase.

None of the three publishes what it transfers to the human alongside the ticket. Full conversation history, extracted intent, attempted actions and the results of those actions are four separate things. A handoff carrying only the first makes the customer repeat themselves, which is the moment that turns a deflection into a complaint.

Ask for a live demonstration of the human agent's screen immediately after an escalation, not a slide about it. If the answer is a chat transcript pasted into a ticket note, the agent is not integrated with your service desk, whatever the integration page says.

Lock-in: what each AI support agent drags in behind it

The per-unit rate is the part of the bill you negotiate. The platform underneath is the part you cannot leave, and it is usually the larger of the two by a wide margin.

Fin is the only one that runs on someone else's helpdesk

Fin publishes native connections to Zendesk, Salesforce Service Cloud, Freshdesk and HubSpot. It also sells a deployment mode that sits on top of an existing helpdesk rather than replacing it. That is a genuine architectural difference, not a marketing one, and it has been the strongest argument for Fin against both platform incumbents.

Agentforce and Breeze are the reverse. Each is a feature of a platform you are already committed to, and neither is sold to anyone who is not. That is not a criticism. It is the whole reason both can price aggressively, because the agent does not have to carry the acquisition cost of the account.

This table describes architecture, not cost. Published implementation and data-platform figures for Agentforce circulate widely but trace to consultancies rather than to Salesforce's own price book, so none is quoted here.

The pattern is the one platform absorption of point tools predicts. A capability that arrives as a line item on an existing contract does not have to win a bake-off. It has to be adequate, and adequate is a much lower bar than better. Every specialist in this category is fighting that asymmetry rather than fighting a product.

Salesforce buying Fin collapses the comparison to two roofs

On 15 June 2026 Salesforce signed a definitive agreement to acquire Fin, formerly Intercom, for approximately $3.6 billion. The deal is expected to close in the fourth quarter of Salesforce's fiscal 2027, and brings more than 30,000 business customers with it.

That changes the shape of this comparison more than any pricing move could. A Salesforce customer running Fin was, until June, running a hedge against its own platform vendor. After close, the hedge and the thing it hedged against report to the same executive.

Salesforce did not need to buy resolution volume. Agentforce ARR reached $800 million in its fiscal fourth quarter, up 169% year on year, on 29,000 closed deals. It bought a product that resolves support conversations better than its own by the only numbers either company publishes, and a purpose-built model to go with it. That is a candid assessment of Agentforce's own support agent, delivered in the form of a cheque.

The practical question for a buyer is whether Fin's per-outcome pricing survives integration. Nobody has said it will not, and nobody has committed that it will. If you are signing a Fin contract this year, the renewal term and any price-protection clause are worth more attention than the rate. That judgement sits inside the wider argument about whether Salesforce is the platform or the roadkill of the agent era.

Where this AI support comparison is weakest

Three things in this post would not survive a hostile review, and it is better to say so than to have a vendor say it for you.

The normalisation rests on assumptions, not measurements

The 65% handling rate applied to all three vendors is HubSpot's published figure for its own product. Using it for Fin and Agentforce is a modelling convenience, not a finding. Your rate will be set by your knowledge base quality, your integration depth and how narrowly you scope the agent, none of which appear anywhere in this arithmetic.

The eight-action figure is likewise illustrative. Salesforce's own $2.00 conversation price anchors the ceiling at 20, and I have used a lower number deliberately to avoid the least flattering reading of its meter. A buyer with real transcripts can replace both numbers in an hour, and should, before quoting any figure from the chart above.

The strongest case against this argument

A Salesforce or HubSpot customer can reasonably argue that per-unit price is close to irrelevant. If the agent is a line item on a renewal you were signing anyway, the incremental spend is small, the procurement cycle is zero, and the integration risk is already absorbed. Against that, Fin has to be better by enough to justify a separate contract, a separate security review and a separate vendor relationship.

That argument usually wins, and it should. The counter-case is narrow but real. If support is the majority of your customer contact, a two-point difference in resolution quality is worth more than the entire pricing delta, and the specialist earns its separate contract several times over. Both positions are defensible. Which one applies to you is an empirical question about your ticket mix, not a matter of taste.

The last weakness is the one I could not fix. No independent audit of resolution quality across these three products exists, and I do not believe anyone has attempted one. Every figure in the quality section traces back to the vendor selling the product. That gap will stay open until either a neutral party runs the benchmark or one of the three publishes its methodology and its sample.

Frequently asked questions

Where to start this week

Pull the last 200 closed tickets from your helpdesk. Label three columns: what the customer wanted, how many distinct system actions the resolution required, and whether a human was needed. That gives you the action count and the handling rate the vendors will not give you, and both are inputs to every calculation above. It takes an afternoon and it survives every pricing change any of these three vendors makes.

Then send one email to each vendor on your shortlist with a single question. Ask them to state, in writing, what happens to the bill when a customer stops replying without confirming their problem was solved. The three answers will differ. The differences are worth more than any demo, and the vendor who answers fastest and most precisely is telling you something useful about the next two years of the relationship.

Related reading