AI Sales Agent: What It Costs in 2026 and Where It Breaks

I spent 12 September 2026 reading pricing pages for AI sales agents. Ten vendors. Four of them publish no price at all.

11x is the cleanest example. Its Alice agent is backed by a16z and Benchmark, and 11x.ai/pricing returns a 404. There is no page on which to hide a number. Artisan, Qualified and Outreach each have a pricing page and none of them carries an entry rate. Outreach is the odd one, because it publishes its meter, seats multiplied by AI credits, and then declines to say what a seat or a credit costs.

The six that do publish something publish it in four different units, and a seventh company, HubSpot, adds a fifth by billing per lead. Per seat. Per credit or action. Per outcome. Per lead. Per agent. The $49 a month on one row and the $500 a month on another measure different things, and nothing on the open web converts between them.

So this page maps the category, the ranked picks sit in my comparison of the best AI agents for sales teams, which orders them by the sales job you want done, and what follows here is the three things that come before that decision: what the category charges, what its own vendors disagree about, and the rules that cap what any of these agents is allowed to do.

Everything below was read from a vendor pricing page, a vendor doc, a platform policy or a named publication on 12 September 2026. I have not run, piloted or deployed any of these products, and I have not spoken to their customers. Where a figure was published by a company selling into this category, I say so in the same sentence as the figure.

What an AI sales agent costs, and who will not say

Four of the ten vendors in the table below publish no entry price. That is the first fact about this category and almost nobody writing about it leads with it, because a comparison table with four empty cells looks broken. Those empty cells are the finding. The column that matters most here is the second one. Not the price. Whether there is a price.

Read from each vendor's own pricing page on 12 September 2026. Monthly rates unless the row says otherwise. Apollo's Basic figure is the annual-billing rate its page shows by default. Gravity's row is my own product and is marked as such.

Artisan shows what the missing cell does to the web around it: third-party write-ups I found quote Artisan anywhere from a few hundred to several thousand dollars a month, a spread wide enough to be useless, which is itself the point. I will not pick a figure out of that range and neither should anyone else, because a spread that wide means the number comes out of sales calls sized to the buyer.

Apollo hides in a different way, because its pricing page renders entirely in JavaScript, so the numbers are visible to a human with a browser and invisible to anything reading the page programmatically, which in a year when buyers increasingly ask an assistant to compare tools for them is a self-inflicted wound.

Qualified's silence has a reason attached. Salesforce announced a definitive agreement to acquire the company on 17 December 2025, describing it as "a leading provider of agentic AI marketing solutions", with the deal expected to close in the first quarter of Salesforce's fiscal 2027, a window that ended on 30 April 2026. The purchase price was not disclosed. I did not find a Salesforce completion announcement when I looked on 12 September 2026, so I cannot tell you from a primary source whether the deal has closed. What I can tell you is that Piper's pricing is now being set inside, or on the way inside, a far larger vendor, and that no entry rate is published either way.

Then there is the guidance problem. Salesforce's own AI sales agent guide tells readers that cost "varies based on business needs, the number of conversations, and level of integration". Salesforce publishes a complete Agentforce price list one click away on the same domain. Both statements are technically true. Together they teach a buyer that the number is unknowable, which it is not.

Fin.ai's roundup goes further. Its title claims the tools were "Tested & Reviewed", and the page publishes no test protocol, no sample size and no dates. Its price column lists numbers in four different units side by side with no normalisation, and at least one of them has been out of date since 14 April 2026: it records HubSpot Breeze as "Included in HubSpot plans", when HubSpot's own newsroom announced outcome pricing for the Breeze agents on that date, for Pro and Enterprise tiers. The page also carries no publication date and no author byline. A page you cannot date cannot be trusted on prices that move quarterly.

Five billing units, and why none of them convert

The reason AI sales agent prices cannot be compared is that the vendors are not selling the same countable thing. One sells you a person's access. One sells you a machine's effort. One sells you a finished job. One sells you a name. One sells you a headcount that happens to be software. Below is what each unit really bills for, with the arithmetic done.

Per seat is still here, and the seat is often just the door

Apollo.io charges $49 per user a month on its Basic annual plan and then meters credits on top, which is the familiar SaaS shape: the seat buys access, the credits buy work. Nothing surprising, and the bill is at least predictable within a range.

Salesforce does something more interesting with the seat. The Agentforce User License is listed at "$5 USD/User/Month", with the parenthetical "(Requires Flex Credits)" attached. Read that carefully. The $5 is not the price of the agent. It is the price of being allowed to buy the agent's fuel. Every unit of actual work is billed separately through the credit meter, which means a $5 line item can sit above a four-figure consumption bill and both are correct.

One "action" costs a cent at one vendor and ten at another

Regie.ai's Pro plan is $49 a month for 5,000 credits, which works out at roughly one cent per credit. Clay's self-serve floor is $167 a month on Launch, and Clay runs two meters at once, actions and data credits, with unlimited seats on top. Two meters means two ways to run out, and the free tier tells you they are genuinely separate: 500 actions and 100 data credits a month, counted apart. I go through what Clay's meters reward and punish in more detail in the piece on running Clay as an agent rather than a spreadsheet.

Salesforce prices Flex Credits at "$500 USD/Per 100k Credits", which is half a cent per credit. Its docs then state that "Agentforce actions are 20 Flex Credits, while Agentforce Voice actions are 30 Flex Credits". Multiply it out and an Agentforce action costs ten cents, with a voice action at fifteen. So one action is a cent at Regie, ten cents at Salesforce, and at Clay it is a unit that has a sibling unit. A tenfold gap inside a single word, on a single SERP, with no vendor obliged to define it. This is the single most useful thing to know before you sit in a demo, because the demo will quote you a per-action figure as though the word were standardised.

Per outcome sounds like one unit and bills as three

Intercom Fin publishes the cleanest outcome definition in the category, and it defines three of them. A "Resolution: No further help is requested after Fin's last answer." A "Procedure handoff: Fin completes a Procedure you've configured to hand off to a human." And a "Disqualification: Fin determines that a prospect doesn't match your qualification criteria."

Two details in the fine print change the maths. Intercom states that "You're only charged once per conversation, even if multiple questions are answered", which caps the downside on a chatty thread. It also states that a "Minimum monthly commitment applies (e.g. 50 outcomes)", which puts a floor under the bill whether or not the agent earns it. Then the rate itself: Fin's standard outcome is $0.99, and a qualification outcome is $9.99. Ten times the headline, for the specific job a sales buyer is shopping for. I have not found another comparison page in this category that reports the $9.99 rate, which means most buyers are modelling their spend on a number that does not apply to them.

HubSpot moved the same direction on 14 April 2026, announcing $0.50 per resolved conversation for the Customer Agent, down from a previous "from $1.00 per conversation", with a 28-day free trial and availability limited to Pro and Enterprise tiers. Note what changed alongside the number, though: the word "resolved" means you now pay only for the conversations that reached an end state.

Per lead and per agent arrived last

Per lead and per agent are the two newest meters, and the per-lead example comes from HubSpot, which sits outside the ten vendors in the table above. HubSpot's Prospecting Agent bills "$1.00 per lead recommended for outreach", a unit with an unusual property: it charges for a recommendation, before anyone has replied and before anyone has opened anything. The agent is paid for its opinion. Lavender prices its Ora agent at $500 per agent a month with 1,000 emails included, so the included allowance works out at fifty cents an email at the cap. That is the most expensive per-message rate on this page by a wide margin, and also the most honest one, because it is the only unit a sales manager can compare directly to a human's cost per touch. You build for free and pay at launch.

Agentforce runs every unit at once

Salesforce is worth a section of its own because its price list stacks five models on top of each other. Salesforce Foundations sits at $0. Flex Credits are $500 per 100,000 credits. Conversations are "$2 USD/Per conversation". The Agentforce User License is $5 per user a month and requires credits. Add-ons run $125 per user a month for Sales, Service and Field Service, and $150 per user a month for Industries. Agentforce 1 Editions start "from $550" per user a month and include 2.5 million Flex Credits per org per year.

Do the derivation on that last one and it gets concrete. At 20 credits per action, 2.5 million credits is 125,000 Agentforce actions per org per year, shared across every user on the edition. For a fifty-seat sales team that is 2,500 actions per person per year, or about ten working-day actions each. Whether that is generous depends entirely on how chatty your agents are, which is the point: the edition price tells you almost nothing until you model the action count.

The history matters too. Agentforce launched at a flat $2 per conversation, and Salesforce introduced Flex Credits on 15 May 2025 to sit alongside it. Both meters still run. One more detail worth pricing in before you sign: the unmetered $125 per user add-on covers employee-facing agents only, so the customer-facing sales motion stays on the meter.

The vendors cannot agree what the word means

Three well-ranked definitions exist and they do not agree on the boundaries of the category. That is worth knowing before you take any of them as the standard.

Creatio's glossary, updated 23 July 2026, says: "AI sales agents are AI-powered tools that can autonomously perform sales tasks and support sales-related activities traditionally handled by salespeople."

Salesforce's guide, updated 13 February 2026 and bylined Erin Hueffner, says: "AI sales agents are autonomous applications that analyze and learn from your sales and customer data to perform tasks with little or no human input."

Fin.ai, which is Intercom's domain and therefore a vendor in the table above, says: "An AI sales agent is software that runs sales development work on its own: engaging buyers, answering product and pricing questions, qualifying leads against your ICP, booking meetings, and updating the CRM." Read that definition knowing who wrote it: scoping the category to sales development is the scoping that most flatters Intercom's own product.

Notice the drift. Creatio includes support for human activity, so a copilot counts, while Salesforce requires autonomy, so a copilot mostly does not. The Fin.ai boundary excludes the account research, CRM enrichment and post-call work that the other two include. Three definitions, three different category sizes. If you want the ground-level version of what the word "agent" means outside sales, I keep that in the explainer on what an AI agent actually is.

Analysts divide the category by autonomy

Creatio uses a two-bucket split of supportive versus autonomous agents. Salesforce's guide uses a two-bucket split of autonomous versus assistive. Same shape, different labels, and both put the dividing line at whether a human is in the loop on every action. It is a clean taxonomy. It is also the wrong axis for a buyer, because autonomy is a configuration setting on most of these products.

Fin.ai splits it by funnel stage instead

Fin.ai, the Intercom property from the table above, uses three buckets instead: inbound qualification, outbound prospecting, and full customer journey. This is more useful than the autonomy split for shortlisting, because a tool built for inbound chat qualification genuinely cannot do outbound sourcing and vice versa, though it is still too coarse to match against a purchase order, as the product pages show.

The product pages ship seven named jobs

Microsoft's Dynamics 365 Sales documentation, last updated 18 July 2026, lists seven separately named agents: the Sales Qualification Agent, the Sales Opportunity Agent, the Sales Close Agent, the Sales Research Agent, the Recommended Actions Agent, the Sales agent in Microsoft 365 Copilot, and AI-powered Data Enrichment. Microsoft then splits the Qualification Agent again, into a "Research-only mode" and a "Research and engage mode".

That is eight distinct behaviours where the analyst taxonomies see two. And it produces a contradiction inside a single company: Salesforce's guide tells readers the category divides into autonomous and assistive, while Salesforce's own product line sells named, job-specific agents with separate SKUs and separate meters. The guide describes one category; the price list describes a catalogue. My reading is that the vendors' product pages are the more reliable taxonomy, because they are the ones attached to invoices. Buy by named job. The named jobs are what you will be billed for.

Where "CRM hygiene agent" and "forecasting agent" came from

"CRM hygiene agent" and "forecasting agent" turn up repeatedly in write-ups of this category. Neither appeared as a named SKU on any of the ten vendor pricing pages I read on 12 September 2026, and neither is in Microsoft's list of seven. The closest a vendor gets is Microsoft's "AI-powered Data Enrichment", which is narrower than what the phrase "CRM hygiene agent" implies. The work itself is real whatever it is called, and people absolutely do hand deduplication and field-cleaning to an agent, which is the job I cover in the walkthrough of an AI agent for Salesforce data hygiene, with the same pattern applying to cleaning an exported Apollo lead list before it ever reaches a sequence. Just do not walk into a procurement conversation asking for a category by a name the vendor's price list does not use.

The delivery rules that cap what any sales agent can do

No vendor in this category mentions the three rules below on its pricing page, and all three are enforced, dated and outside any vendor's control. They set a ceiling on the outbound motion that no amount of model quality can raise.

Gmail's spam-rate ceiling is the binding constraint

Google's sender guidelines state that "Starting February 1, 2024, email senders who send more than 5,000 messages per day to Gmail accounts must meet the requirements in this section", one of which is to "Keep spam rates reported in Postmaster Tools below 0.3%", with below 0.10% recommended. The consequence is stated plainly: "If you don't meet the requirements described in this article, your email might not be delivered as expected, or might be marked as spam."

Now do the arithmetic that nobody does. A spam rate of 0.3% is three complaints per thousand messages. A sender at the 5,000-a-day threshold therefore has a budget of roughly fifteen spam complaints in a day before it trips the limit, and roughly five if it wants to sit at Google's recommended 0.10%. Fifteen annoyed strangers. That is the entire margin for error on an automated outbound programme, and an agent that personalises badly at scale will spend it in an afternoon.

This is the number I would put on the wall before buying any AI SDR. Volume is not the constraint. Complaint rate is, and complaint rate gets worse as volume goes up.

Outlook rejects the message at the door

Microsoft began enforcing equivalent authentication requirements on 5 May 2025 for senders pushing 5,000 or more messages a day to outlook.com, hotmail.com and live.com. The difference from Google's approach matters operationally: Microsoft returns the error "550 5.7.15 Access denied" at the door, where Google files the message in junk. A rejection is louder than a junk-folder placement and easier to detect, which is the small mercy here. Your sending tool sees a hard bounce and you find out the same day.

A note on this source, because it affects how much weight to put on it. Microsoft's own announcement page returned title-only content to my fetcher on 12 September 2026, so I could not read the primary. The detail above comes from dmarcian's 4 April 2025 write-up reproducing Microsoft's announcement. Treat it as reliable secondary reporting, not as text I read on microsoft.com.

LinkedIn's agreement prohibits the motion half this category advertises

LinkedIn's User Agreement, effective 3 November 2025, prohibits members from doing the following at section 8.2.13: "Use bots or other unauthorized automated methods to access the Services, add or download contacts, send or redirect messages, create, comment on, like, share, or re-share posts, or otherwise drive inauthentic engagement".

Read that against the marketing copy for almost any AI SDR product. "Multichannel outreach across email and LinkedIn" is one of the most common feature bullets in the category. Automated LinkedIn messaging is the exact behaviour section 8.2.13 names. The vendor is selling a capability, and the platform's contract with your rep forbids using it.

I am not predicting enforcement and I have no evidence of a campaign against these tools. The contractual position is simply unambiguous, and the account that gets restricted belongs to your rep. If LinkedIn is genuinely central to your motion, the safer shape is an agent that researches and drafts while a human sends, which is the pattern I describe in the piece on AI agents and LinkedIn Recruiter outreach.

What has already gone wrong, with dates attached

The category's best-documented failure is 11x. On 24 March 2025, TechCrunch reporters Dominic-Madori Davis and Marina Temkin published claims from named companies listed as 11x customers. ZoomInfo told them: "We did not give them permission to use our logo in any manner, and we are not a customer", and said that during a one-month trial the product "performed significantly worse than our SDR employees". Airtable said the product "was never used in production and never rolled out to our sales team". A former employee told the reporters: "We were losing 70-80% of customers that came through the door."

11x disputed the picture. The company said it reports contracted ARR, that its investors had reviewed the underlying contracts, and that its "retention rate is currently 79%". It denied misleading sales practices. Both sides of that are in the same article and both belong in any honest summary of it, including this one.

The wider pattern has numbers too. Gartner predicted on 25 June 2025 that "Over 40% of agentic AI projects will be canceled by the end of 2027", a forecast published alongside a January 2025 poll of 3,412 webinar attendees. Gartner also estimated that of the thousands of vendors claiming agentic AI, only around 130 are real, a phenomenon it labelled agent washing. Gartner's site blocks automated fetching, so these figures come to you through consistent reproductions across multiple independent outlets.

S&P Global Market Intelligence found something similar across AI generally. In research reported by CIO Dive on 14 March 2025, covering over 1,000 respondents in North America and Europe, "The share of companies abandoning most of their AI initiatives jumped to 42%, up from 17% last year", and the average organisation scrapped 46% of its AI proofs of concept before they reached production.

One thing I deliberately will not give you is a reply-rate benchmark. Every cold-email number circulating in this category comes from cold-email vendors reporting on their own customers' send data, the published figures vary widely depending on which vendor published them, and none of it has been independently audited. Any single figure you see quoted as the benchmark was picked out of that range by someone with a product to sell.

Two numbers in this category measure the wrong thing

Both of the following are real, published and widely repeated. Both mislead in a way that is easy to miss if you only read the headline.

A high accuracy score can hide a model that always says no

A study published in Frontiers in Artificial Intelligence on 7 March 2025 tested fifteen classifiers against 16,600 cleaned CRM records. The best model reached 98.39% accuracy with an AUC of 0.9891, which is the kind of number that ends up in a vendor deck.

The paper is honest about what sits underneath it. The target variable is severely imbalanced, and the authors kept it imbalanced on purpose because the rebalanced versions scored worse. On a severely imbalanced target, a model that predicts the majority class every single time already scores close to the headline accuracy. The 98% is measuring how rare qualified leads are at least as much as it is measuring the model.

This is why I treat any accuracy figure in lead scoring as uninformative on its own. Ask for precision and recall on the positive class, or ask how many genuinely qualified leads the model missed. Those are the numbers that decide whether a scoring agent earns its place, and they are the ones I work through in the practical guide to lead scoring with a HubSpot agent.

Salesforce disagrees with Salesforce about how much reps sell

Salesforce's AI sales agent guide states that reps spend "28% of their time actually selling". Salesforce's State of Sales, sixth edition, published 3 February 2026, states that "The average seller spends 40% of their time selling". Two Salesforce properties. Two numbers. The same claim about the same population.

A twelve-point gap on the most quoted statistic in sales productivity, inside one company's own publications, is a good reason to stop treating it as a fact. When a vendor uses the low number to size the pain, check whether the vendor's own research says something else.

The State of Sales research itself is worth reading with its provenance in view. It surveyed 4,050 sales professionals across 22 countries, fielded in August and September 2025, and reports that 87% of sales organisations use some form of AI and that 54% of sellers say they have used agents. Those are Salesforce's own numbers, published by the company that sells Agentforce. They may well be right. The point is that the adoption figures repeated across this category mostly originate from a party with a commercial interest in adoption looking high.

The line where the agent stops and a person starts

Buyers still want the rep, and they say so in the research. Reporting from Demand Gen Report on 26 May 2026, covering Gartner buyer research released on 20 May 2026, found that 69% of B2B buyers turn to sales reps to validate AI-generated insights, that 67% would prefer a rep-free experience, and that 51% had encountered misleading information generated by AI. Those first two findings look contradictory until you notice they describe different moments: people want to self-serve the research and want a human to confirm it before they commit.

One caveat on that source, stated because it changes how much weight it carries. Neither Gartner's release nor Demand Gen Report's write-up published a sample size, fielding dates or geography for those two surveys. I am reporting them as directional signals.

The 6sense 2025 B2B Buyer Experience Report, published 12 November 2025 and based on more than 4,000 buyers, is more specific about timing. It found that 94% of buying groups had already ranked a preferred vendor before first contact with any seller, and bought that favourite 77% of the time. First contact with a vendor now happens around 61% of the way through the buying journey. Put those together and the job description for an AI sales agent narrows usefully. By the time your agent gets a reply, the buyer has usually decided who they prefer. The agent's realistic contribution is being present, accurate and fast in the 61% of the journey that happens before any conversation, and being clean enough with the data that the human who does eventually show up knows what has already been said.

Here is the boundary I would write into a rollout, and I would call it the commitment line. Everything that produces information stays with the agent: account research, enrichment, drafting, scheduling, logging, list cleanup, follow-up that restates what was agreed. Everything that creates an obligation goes to a person: price, discount, contract terms, delivery dates, anything a buyer would reasonably treat as a promise from your company. The test is whether your company would have to honour the sentence.

Intercom already treats the handover this way, which is the strongest evidence available that the boundary is a design choice with money behind it. A "Procedure handoff" to a human is one of Fin's three billable outcomes. The vendor charges for the handoff, the same as it charges for a resolution. In the product's own economics, passing the conversation to a person counts as a job completed. Cold pipeline is a good place to see the line in practice, since the agent can do all the research and drafting while a human owns the actual offer, which is how I structure it in the cold lead follow-up walkthrough.

Matching the billing unit to the job

Start from the job and the shortlist writes itself. Start from the category and you will spend a fortnight comparing units that do not convert.

If you need a genuine outbound SDR motion at volume, buy a specialist. Regie.ai is the one I would look at first among the dedicated tools, purely because it publishes a real entry price at $49 a month with a stated credit allowance, so you can model the bill before a sales call. Build the target list somewhere that is good at lists; Clay is the obvious candidate and its dual meter is at least documented.

If your company already runs on Salesforce or Dynamics, the native agents deserve the first look regardless of what a comparison page says, because the integration work you avoid is usually worth more than the per-action difference, provided you budget for consumption and model the action count before you sign, since the seat price on its own will tell you almost nothing.

Inbound website qualification is its own product shape. Fin and Qualified both live there, and Fin is the one publishing rates, with the caveat that the qualification outcome bills at $9.99 rather than $0.99. Conversation intelligence is a different purchase again, sitting on recordings instead of outbound. If that is the gap, what a Gong AI agent does with call recordings covers the shape of it.

Now Gravity, last and honestly. We are not an AI SDR and we do not send cold outbound at volume. What we do is take a specific, bounded piece of sales work and hand it back finished: research this list of accounts, draft these follow-ups from the call notes, clean this CRM export and flag the duplicates. You describe the task in plain words, an expert-built agent runs it, and the finished output comes back to you to check and use.

The pricing is deliberately boring, because the point of this whole page is that the meters are the problem. Your first agent is free, with no card. After that Autopilot is $20 a month worldwide, ₹1,999 a month in India, with a lot of usage included, and Minipilot is ₹399 a month in India with a smaller allowance. If you run out, you buy more usage. There is no action to count and no outcome to define.

If your actual requirement is a thousand cold emails a day into new accounts, we are the wrong tool and Regie.ai or a comparable specialist is the right one. I would rather say that here than have you find out in week three.

FAQ

Sources