What an AI SDR Costs If It Fails: 1.2x to 2.9x Sticker

Here's the trap. You get quoted $45,000 a year for an AI SDR, you hold it against the $134,000 to $154,000 a human SDR costs you fully loaded, and it looks like the easy call. Then the thing underperforms.

You work that out around month four, and you keep paying until month twelve, because the contract you signed contains no clause that lets you stop.

The sticker is a price per month only where every month is one the tool earned. In the other case you divide the same annual figure by four instead of twelve, then add what you paid to get it running and what you lose on the way out. Signed deals in the AI SDR category run $18,000 to $65,640 a year, and neither of those two additions is in the quote.

It works like a twelve-month season ticket bought for a job you might leave in April. The price assumes you travel all year, the railway keeps the balance when you stop, and every journey you took ends up costing three times what the brochure said. The sticker price is a price for the good outcome only. To get the real one you divide by the odds the deployment survives, and the published figures a buyer could use for that run from 5% to 89%. That is the difference between paying 1.2 times the sticker and paying 2.9 times it.

So what does an AI SDR actually cost if it fails?

On a $45,000 annual contract with no exit, where you conclude at month four that it has not worked, the cost per month the tool actually earned its place lands between $4,575 and $10,875 depending on which published survival figure you use.

The sticker implies $3,750. Every figure in that band is computed from the published inputs below, not measured.

Two things there matter more than any single row. The multiple is set by the survival probability and the month you find out, never by contract size: an $18,000 deal and a $65,640 deal carry the identical multiplier, which is $1,500 and $5,470 a month of premium.

And the seven rows disagree by a factor of seventeen while every one is the best available figure from a real, dated, published source.

This page will not pick one.

Picking one would be the most useful thing to do and the least honest, because none of those seven numbers measures what a buyer of an AI SDR needs measured. What follows is why each falls short, then how to make the answer stop mattering.

What a stranded month is, and why it is the only number that matters

A stranded month is a month you pay for after concluding the tool is not working. It is not a month of poor performance. The decision has already been made and the invoice arrives anyway, because the contract runs longer than the evidence took to arrive.

Stranded months are the entire cost of being wrong, and the only part of it you can price before signing.

You cannot know whether the tool will work. You can read today, from published pages, exactly how many months you would keep paying if it does not.

The three other words this page turns on

Survival probability is the chance the deployment is still running at the end of the term, written as p. The notice window is how far ahead of the term's end you have to say you are leaving, and missing it renews you. A break clause is a right to end the contract early without proving the vendor did anything wrong, which is a different thing from a termination clause and much rarer.

The fourth term does the most damage. An Order Form is the one-page document you sign alongside the public terms, and it is where the term length, the renewal mechanism and the notice window actually live. The public page tells you the rules. The Order Form tells you the numbers, and you do not see it until you are already in a sales process.

Why the contract finds out later than you do

Assume you conclude at month four: ramp plus one full quarter of real pipeline, about the earliest a reasonable buyer stops blaming onboarding. Nothing in the published record fixes that month, and moving it moves the answer more than the churn rate does.

Now read the exit. 11x's published Terms and Conditions, last updated June 2026 and read on 2026-08-25, has a section numbered 5.1 and titled "Term and Renewal". In full: "The term of this Agreement begins on the Effective Date and will remain in effect for the term described in the Order Form." That is the whole clause.

A section titled Term and Renewal containing no renewal mechanism, no term length and no notice window, because all three sit in an Order Form the prospect has not seen.

Section 5.2 supplies the only two early exits: a material breach uncured after thirty days' written notice, or the other party's insolvency.

There is no termination-for-convenience clause anywhere in the published terms. That is a vendor's own document, not a characterisation of it.

The one performance-linked exit 11x publishes, and its 60-day clock

Section 7.3 warrants that the Service "will function as described in the Documentation and the DPA". Section 7.4 makes repair the exclusive remedy, then says that if 11x "cannot repair such deficient Service as warranted within sixty (60) days after receipt of written notice", the customer may terminate that Order Form and recover a pro-rata share of prepaid fees.

Count the calendar. You conclude at month four, write the notice, the sixty-day clock starts, and the earliest you are out is month six. What you must demonstrate is not that the tool failed to book meetings. It is that the Service did not function as described in a Documentation the vendor wrote.

Treating that clause as an exit is a reading, not a promise.

The model, written out

Five inputs, all of which you can read or set before signing. S is the annual subscription. I is implementation, everything you pay to get it live. X is the switching cost, what it takes to rebuild the motion somewhere else. M is the month you conclude it has failed. N is the month your payments actually stop, which the contract sets, not you.

In the good case you spend S plus I for twelve useful months, so the cost is (S + I) / 12. In the bad case you spend S multiplied by N over 12, plus I, plus X, for M useful months, so it is (S x N / 12 + I + X) / M. Weight the two by p.

Run it bare first, with I and X at zero and N at 12, because that is the version where every input is published and nothing is estimated.

On a $45,000 contract failing at month four, p = 0.50 gives $7,500 a useful month against a $3,750 sticker. Exactly 2.00 times. At p = 0.79, $5,325 and 1.42 times. At p = 0.21, $9,675 and 2.58 times.

Where the survival probability would have to come from

No AI SDR vendor is public, so none of them files. The nearest evidence is what comparable public software companies disclose about their own installed base, and those disclosures are useful precisely because the companies are legally on the hook for them. Five of them, read off EDGAR on 2026-08-25:

Two of those rows are more useful than the numbers. Semrush's 10-K states that its "customer cohorts typically experience their lowest dollar-based net revenue retention rate during their second full year after becoming a customer". A company disclosing that its own retention bottoms out in year two is telling you the first renewal is the cliff, not the first quarter.

The Braze rows say the other half. Retention among customers at $500,000 or more of ARR runs a point above the whole book, and ZoomInfo's 10-Q states that it expects retention to improve as it moves upmarket because "larger customers have historically exhibited higher net revenue retention". Both are filed disclosures, and both put the small buyer on the wrong side of the vendor's own average.

Why net revenue retention is not your survival probability

Every figure in that table is dollar-based. Each company takes the revenue from a cohort of customers twelve months ago, measures what the same cohort spends now, and divides. Klaviyo's 10-Q spells the method out and so do the other four. Nothing in the calculation counts companies.

Expansion hides departure. A book where four customers in ten leave and the remaining six double their spend reports 120% net revenue retention and a 40% logo loss in the same breath, and only the first gets disclosed. It is not a survival probability and no filing presents it as one.

The inference that survives runs one way. Net revenue retention below 100% means the installed base is shrinking net of every expansion the vendor could book, a hard floor on how much churn is happening. ZoomInfo at 89% is disclosing that.

Above 100% tells you almost nothing about logo survival.

One disclosure travels with the ZoomInfo figure. A securities class action was filed 2026-06-25 in the Western District of Washington, covering purchasers between 2025-11-03 and 2026-05-11, alleging the defendants "overstated their confidence in statements related to the Company's projected revenue outlook, the growth of its AI-driven products and its sustained improvement in net revenue retention". That is an allegation in a pending case, not a finding, and the 89% is quoted as filed rather than proven.

The one company here whose churn actually got reported

On 2025-03-24 TechCrunch published an investigation by Marina Temkin into 11x, the only substantial press reporting on churn inside this category.

It needs care, because it carries two figures that are near mirror images of each other. One employee told the publication: "We were losing 70-80% of customers that came through the door." The same article carries 11x's on-the-record position that its "retention rate is currently 79%", and its statement that its "highest churn" was in "initial cohorts in late 2023" with retention improving since. Both are reported claims. Neither is a finding of fact, and this page does not treat either as one.

Where the 21% row in the table above comes from

The article reports one employee saying the company might state $14 million in annual recurring revenue when "the number of contracts that passed the three-month trial period totaled only about $3 million". Three over fourteen is 21.4%. It deserves every hedge available: one anonymous employee's figure, one outlet, about revenue rather than customers, from March 2025, and contradicted by 11x's own published position.

The reporting also supplies the contract shape. TechCrunch reports 11x was "adamant" that prospects wanting a pilot sign a one-year contract, quoting a prospective customer: "They were resistant to signing any sort of trial or letting us experiment." Instead the company offered a break clause "typically at three months", and 11x's spokesperson told the outlet that some enterprise customers "require a 12-month contract with an opt-out after 3 months."

Read that last phrase as a buyer rather than as a reader. "An opt-out after 3 months" does not say whether it is a right you hold for the remaining nine months or a window that closes at month three. On a $45,000 deal the gap between those two readings is $30,000, and it is one sentence in an Order Form.

Ask which it is, in writing, before the sentence becomes contested.

What eight vendors actually publish about getting out

Read on 2026-08-25 from the vendors' own pricing and terms pages. Cells reading Not stated publicly are not gaps in the research but the finding itself: a term you cannot read before the sales call is a term you will negotiate without a comparison.

Four of the nine rows are almost entirely "Not stated publicly", and those four sell the largest contracts. That is not a conspiracy. A vendor whose published terms contain a notice window has given every prospect a number to negotiate against; a vendor whose terms point at an Order Form has not.

"Cancel anytime" and what cancelling actually costs

AiSDR is the most transparent vendor here on price, and the clearest example of why price transparency and contract transparency are separate things. Its pricing page, read 2026-08-25, carries the headline "Plans from $250/mo - cancel anytime" and the line "Most AI SDR vendors push annual contracts. We want you to see the results first." Both are true as far as they go.

Its Terms of Service, read the same day, define what cancelling means. Under Early Cancellation: "You may choose to cancel your subscription early at your convenience provided that, we will not provide any refunds of prepaid fees or unused Subscription, and you will promptly pay all unpaid fees due through the end of the Subscription term." The next section adds: "This Agreement may not otherwise be terminated prior to the end of the Subscription term."

So "cancel anytime" means you may stop using it anytime. It does not mean you stop paying. On AiSDR's quarterly terms that is bounded exposure and a better deal than an annual one; on the annual plan the same sentence carries very different weight. Vendor's published terms against vendor's published marketing is the only kind of contradiction worth printing.

The notice window catches more people than the cancellation clause does. AiSDR's non-renewal deadline is "10 days prior to the expiration of your then-current Term"; Salesforge's is seven days, requestable "only via our live chat support". A ten-day window on a quarterly term means four dates a year on which the decision has to already be made. Put them in a calendar the day you sign.

The implementation fee nobody in this category prices

I in the model is what you pay to get the thing live, and no AI SDR vendor publishes it as a line item. Artisan's pricing page states that "Every plan comes with a white-glove rollout: CSM support from day one, guided onboarding, and sending infrastructure sized to your motion" and separately promises "No hidden line items." Together those vendor-authored sentences mean implementation exists, costs something, and is not separable from the subscription.

One published number does put a dollar on the human labour that makes an AI SDR work, and it is AiSDR's own. Its Scale tier offers a per-campaign managed service at +$149/campaign; its Enterprise tier offers a fully managed service at +$2,500/mo, on top of a $2,500/mo software tier. Read that ratio slowly.

The service that operates the software costs exactly what the software costs.

Three months of that rate is $7,500, the implementation figure this page loads into the model. It is a proxy and labelled as one: a published rate for a related service at a different price point, not an implementation quote from anybody. Ask your own vendor to price theirs and substitute the real number.

I matters more than its size suggests because it is entirely sunk. At p = 0.50 with a $7,500 implementation and no switching cost, the cost per useful month goes from $7,500 to $8,750, or 2.33 times sticker rather than 2.00. Add $10,000 of switching cost and it reaches $10,000, or 2.67 times.

What you cannot take with you

X is the switching cost, and the terms that set it are published, short and almost never read. Salesforge is the plainest. Its terms state that "as of the effective termination date of your account, You will cease to have access to your account and to the Data stored therein". No export window, no format, no grace period. If you intend to leave, you export before the termination date, not after, and that is a scheduling constraint rather than a feature request.

11x handles it under confidentiality rather than portability. Customer Data is defined as the customer's Confidential Information, and Section 5.3 requires each party, on the other's request, to "within seven (7) days, return or destroy (and certify destruction of) all copies". Return or destroy is the vendor's choice, and no export format appears in the document.

The training clause, where the two vendors take opposite defaults

11x's Section 3.2.2 states that "11x shall not use any Customer Data to train any artificial intelligence or machine learning models", an unusually clean commitment. AiSDR's terms take the opposite default: "We may use Customer Data for machine learning... You instruct us to use Customer Data for such purposes", with an opt-out by email. Both are the vendors' own words, and the second is a reason to send that email on day one rather than on the way out.

The month your payments stop is worth more than the churn rate

Hold p at 0.50 and the failure month at four, and vary only N. This is the most controllable input in the model and the one buyers spend the least negotiating capital on.

The top row is the trap inside the solution. A break clause fixed at month three, exercised on a deadline rather than held as a right, forces the decision one month before the evidence lands. It costs less per useful month than the good outcome, which is arithmetic telling you the premise is wrong: you are cancelling deployments you have not finished evaluating.

The bottom row is where most of this category sits. Annual, no break clause, no termination for convenience, $3,750 a month of pure premium on a $45,000 deal at even odds. Nothing about the product differs between the two rows.

One sentence in the Order Form does.

Move M instead and the effect is larger. At p = 0.50 with N held at 12, concluding at month two costs $13,125 per useful month, month three $9,375, month four $7,500, month six $5,625 and month nine $4,375. A vendor who gets you to real pipeline quickly is worth more than one who is cheaper, because early failure is the expensive kind and a slow ramp is the expensive kind twice over.

The contract-shape audit: 40 minutes, published pages only, no trial

Run this before your next vendor call, from your own laptop, with no demo and no trial. It is the audit this page ran on nine vendors, at roughly eight minutes each.

• Find the terms URL and time-stamp it (3 min). Try /terms, /legal/terms, /terms-of-service. Record the "Last Updated" date; 11x's reads June 2026. Fail if: no public terms page exists, because then the whole contract lives where only a sales process reaches it.
• Find the term and renewal section (5 min). Search "renew". You want a term length, an auto-renewal statement and a notice window in days. Fail if: it names an Order Form instead of a number.
• Search "convenience" (2 min). A termination-for-convenience clause is the only exit that does not require you to prove fault. Fail if: the word does not appear. In this category it usually does not.
• Search "refund" and "non-refundable" (3 min). Establish whether early cancellation stops the fees or only the access. AiSDR's and Salesforge's published answers both say the fees continue. Fail if: cancelling does not stop the invoices on a term longer than a quarter.
• Search "export", "delete", "return", "Customer Data" (5 min). You want an export window in days and a named format. Fail if: the only obligation is return-or-destroy at the vendor's option, or access ends on the termination date.
• Search "train" and "machine learning" (3 min). Record whether the default is opt-in or opt-out and whether the opt-out is self-service or an email. Fail if: the clause is silent, which is worse than either default because it is unbounded.
• Read the uptime SLA remedy (4 min). Check whether the credit is capped at a share of one month's fee. Capped that way it is a rounding error, not a remedy, and never a substitute for an exit.
• Price both versions (15 min). Put S, I, X, M and N into the formula at p = 0.50 and again at p = 0.79. If the two answers straddle a number your finance team would refuse, the survival probability is a live question in the negotiation.

What this audit cannot detect. Published terms are the opening position, not the deal. It cannot tell you what a redline gets you, what the Order Form says, or whether a vendor who publishes hostile terms enforces them. Nor can it read a vendor with no published terms, which here is four of nine. For those the output is "we could not check", which still belongs in the procurement file.

Is a break clause worth more than a 20% discount?

This is the part that changes what you do on Monday, because buyers spend their negotiating capital on price and hand back the exit terms for free.

Twenty percent off a $45,000 contract saves $9,000, guaranteed, in every outcome. A break clause you can exercise at month four avoids eight stranded months, or $30,000, but only in the case where you cancel. Its expected value is $30,000 multiplied by the probability of failure.

The two are equal at exactly 30.0%.

Below that, take the discount. Above it, take the clause, and the gap widens fast: at a 50% failure rate the clause is worth $15,000 against the discount's $9,000, and at 70% it is worth $21,000. Against a 10% discount the breakeven falls to 15%, which means a small discount almost never beats an exit.

One reading breaks this and it belongs here rather than in a footnote. If you accept 11x's published figure of 79% retention, the implied failure rate is 21%, below the 30% breakeven, and the discount becomes the better trade. That number is vendor-reported, treat it as a ceiling not a floor: it measures account retention rather than project survival, and it sits in the same article as an employee saying close to the opposite. It is still the only company-published figure on the record here, and if you believe it, the arithmetic says negotiate on price.

So the advice is conditional and the condition is checkable. Ask the vendor for their twelve-month logo retention rate in writing. If they give you a number, price the discount against the clause using it. If they decline, price the clause at 50%.

What the practitioners say about month three

There is almost no practitioner corpus on AI SDR deployments specifically, which is documented below. The adjacent corpus, on generative AI pilots inside companies, is large and on point. On Hacker News item 44941118, from 2025-08-18, 230 points and 167 comments, discussing the widely reported MIT finding that 95% of generative AI pilots at companies were failing, one commenter wrote: "LLMs get you 80% of the way to a solution almost immediately but that last 20% when it comes to missing knowledge, data, or accuracy is a complete tar pit and will wreck adoption."

That is one person's characterisation, not a finding of fact. It names the timing problem exactly: an evaluation that looks strong in week two and collapses in month three is the worst possible shape for a twelve-month contract.

The two comments that argue against this whole page

In the same thread, commenter trenchpilgrim asked: "What's the failure rates if technology pilots in general for comparison? For example, I heard that SAP has an 80-90% deployment failure rate back in the day, but don't have a citable source for it." Nobody produced one. If most software pilots fail at a comparable rate, AI SDR churn is a fact about buying software rather than about AI.

A reply made the definitional point that undermines every headline failure rate, including the one in that thread's own title. Commenter RaftPeople wrote: "ERP 'failure' is frequently defined as went over budget or over time, even if it ultimately completed and provided the desired functionality. It's a much smaller percentage of projects that are either cancelled or went live and significantly did not function as the business needed." This model needs cancellation. Most published failure rates measure disappointment.

On item 44421803, from 2025-06-30, discussing Gartner's cancellation prediction, the top comment is the right response to any analyst figure: "I can easily believe that number. However it's unclear how they arrived to it. Can anyone with a Gartner subscription take a look to see if this passes a smell test?" Nobody in the thread did. This page uses that prediction for one table row, as a prediction, from the headline of Gartner's press release dated 2025-06-25, and does not quote its reasoning because that release body was not read first-hand.

What the vendors' own job boards say about whether deployments survive

A vendor's open roles are a checkable signal about where its effort goes. Two boards in this category were reachable on 2026-08-25. Artisan's Ashby board carried three open roles across three teams named BDR, Forward Deployment and Marketing: Business Development Representative, Forward Deployed GTM Architect and Growth. Two of the three teams put humans next to the deployment, one doing the outbound and one making the rollout work.

That is not hypocrisy. A Forward Deployed GTM Architect is what a category looks like when the product does not configure itself, and every enterprise software company that ever worked hired that role. What it means for the model is that implementation is a staffed cost centre at the vendor, which makes it improbable that I is zero for the buyer.

Regie.ai's Greenhouse board carried exactly one open role, updated 2025-03-13: a Customer Success Manager in Bellevue, required in office. The duties include "onboarding and training, renewals and minimizing churn". Every SaaS CSM req says something similar, and it confirms renewal here is contested rather than automatic.

What job boards cannot tell you. Board tokens change, roles close without being removed, and a company hiring through referrals posts nothing. 11x, Qualified and Salesforge returned nothing on Greenhouse or Ashby on 2026-08-25, and the correct inference is that we could not find their board. Report it as a floor, never as a share.

Why there is no litigation to read, and what that tells you

When a deployment fails badly enough that money is disputed, it usually surfaces in a docket. A CourtListener RECAP search on 2026-08-25 for "11x.ai", "Artisan AI", "Regie.ai" and "AI SDR" returned no contract disputes involving any AI SDR vendor. The only 11x-adjacent hit was an unrelated immigration case.

Three innocent explanations come before any interesting one. The contracts are small enough that litigation costs more than the disputed amount. The category is three years old and disputes take longer than that to be docketed. And most commercial software agreements carry arbitration clauses, which keep disputes off the public record.

The consequence is that no case law tells you how a court would read an ambiguous break clause here, and there is no settled meaning for "an opt-out after 3 months". Where a market has no precedent, the drafting is the whole protection.

Get the ambiguity out of the Order Form, because there is nothing behind it to fall back on.

Who should stop reading here

If you are buying at the self-serve end, most of this page does not apply to you. Clay from $185 a month, Regie.ai's Pro tier with cancellation at the end of the billing period, and AiSDR's Solo plan month-to-month all cap stranded exposure at one billing period, so the churn premium is close to zero by construction.

This page is for buyers asked to sign twelve months at $18,000 or more, which is where 11x, Artisan, Piper by Qualified and the unpriced half of this category live. If that is not you, take one line away: check your notice window and put it in a calendar.

It is also not for you if you have already run the motion manually and know it works. The premium priced here is a premium on uncertainty, and a buyer who has proved the sequence, the list and the offer by hand is buying execution rather than a hypothesis. They just cannot prove that to the vendor, and the vendor will not price it either way.

What to ask for before you sign

Five things, all cheaper to get in writing now than to argue about in month five, in the order they are worth asking for.

• A break clause at month four, held as a right rather than a deadline, in the contract rather than an email. Four, not three, because three forces the decision before the evidence. Ask what it costs to add: paying 10% more for an exit beats 20% off without one at any failure rate above 15%.
• The vendor's twelve-month logo retention rate, in writing. The share of customers who signed twelve months ago and are still paying. Not net revenue retention, which as five filings above show can sit above 100% while customers leave. A refusal is an answer.
• A written definition of month-four success, in meetings held, not booked and not replies. At $7,500 per useful month against this site's $960 to $1,100 per held meeting, break-even against an in-house SDR sits near eight held meetings a month, sustained.
• An export window in days, with a named format. Ask specifically: on the termination date, do I still have access, and for how long? Salesforge's published answer is no. 11x's is return-or-destroy on request within seven days. Neither is discoverable afterwards.
• The implementation figure, itemised. Even where it is bundled, ask what it would cost standalone. That is what you forfeit if the deployment dies, and a vendor who will not name it has made the sunk portion of your spend unmeasurable.

Two of those five change no commercial term and cost the vendor nothing. If a rep resists the export window or the itemised implementation figure, that resistance is information about how the rest of the negotiation will go.

What this page does not know

The missing measurement is logo survival at twelve months: of the companies that signed a twelve-month AI SDR contract, what share were still paying a year later, and what share had stopped using the product while still paying. Nobody publishes either. No AI SDR vendor is a public filer, and the five retention disclosures above are dollar-based and cannot separate a customer leaving from a customer shrinking.

This page has not run that experiment either. We hold no contract data, surveyed no buyers, and observed no deployment succeed or fail. That is why the model runs as a sensitivity analysis across seven survival probabilities rather than one answer: asserting a single rate would have been the only invented sentence here. The full account, including what the practitioner search returned, is below the FAQ.

So what does an AI SDR cost if it fails?

Between 1.22 and 2.90 times the sticker on the bare model, and between 1.66 and roughly 4 times once implementation and switching are loaded in. On a $45,000 annual contract that is $4,575 to $10,875 for every month the tool earned its place, against a sticker implying $3,750. The width of that band is not sloppiness. It is what a category whose survival rate nobody has measured in public actually looks like.

The number to aim for is the sticker itself, and the distance between the two is one clause rather than a better vendor. A break clause exercisable the month you decide takes N to 4, which takes the multiple to exactly 1.00 at any survival probability and any contract size. The tool is no more likely to work.

You are simply no longer funding the months after you found out.

Do one thing this week. Open the terms page of whichever vendor is furthest along in your process and search it for "convenience". If it is not there, you know the exact question your next call has to answer and what it is worth: $30,000 on a $45,000 deal, or 2.00 times sticker at even odds, or eight held meetings a month you would otherwise have to find somewhere else.

The observed contract values for the three vendors most buyers choose between sit side by side in 11x vs Artisan vs AiSDR. The tools that skip the annual commitment are counted in AI SDR tools with no annual contract, and what happens when exit terms were not agreed up front is in the ZoomInfo contract exit.