Unit Economics Teardown: What a Sales-Development Agent Actually Costs at Scale

Table of Contents

• Why the SDR Agent Is the Cleanest Test Case in GaaS
• The Outcome You're Actually Selling
• Cost Stack: Walking the P&L Line by Line
• Inference: Smaller Than You Think
• Data and Enrichment: The Quiet COGS
• Deliverability Infrastructure
• The Human-in-the-Loop Tax
• A Worked Example: 10,000 Contacts a Month
• Where the Margin Actually Leaks
• Pricing Models and Which One Survives Contact
• Insights Most People Overlook
• References

Why the SDR Agent Is the Cleanest Test Case in GaaS

Of all the vertical agents being sold as a service right now, the sales-development rep is the one most people can price in their head. The outcome is legible, a booked meeting, ideally a qualified one, and the comparison is obvious: a human SDR costs somewhere between $5,000 and $9,000 a month fully loaded, books maybe eight to fifteen qualified meetings in a good month, and quits in eighteen. That makes the unit math tempting. If an agent can book the same meeting for $30 instead of $400, the pitch writes itself.

Except the pitch is where the trouble starts. The reason the SDR agent is such a clean test case for Agentic AI-as-a-Service economics is the same reason it's brutal: the outcome is measurable, so customers will actually hold you to it. A coding agent can hide behind "developer productivity." A support agent can hide behind deflection rates that nobody audits. An SDR agent either put a meeting on the calendar or it didn't, and the buyer's CRM knows the answer. That measurability is exactly what forces honest unit economics, and exposes the vendors who've been quietly subsidizing their own cost-per-meeting to win logos.

The Outcome You're Actually Selling

Before any cost analysis, you have to decide what counts. This is the trap. "Meeting booked" and "qualified meeting that showed up and was a real opportunity" are separated by a chasm of cost.

Walk the funnel an SDR agent actually runs:

1. Sourced contact, a name and company pulled from a list or ICP definition.
2. Enriched contact, verified email, role confirmation, recent trigger event.
3. Sent sequence, a multi-touch email/LinkedIn cadence, personalized.
4. Positive reply, someone said "tell me more" or "sure."
5. Meeting booked, a calendar event exists.
6. Meeting held, they showed up.
7. Qualified opportunity, sales said it was worth the AE's time.

Each step has a survival rate, and the survival rates compound viciously. If 35% of sequences get a reply, 8% of those are positive, 70% of positives book, 65% of bookings show, and 60% of shows qualify, then a sourced contact converts to a qualified opportunity at roughly 0.6%. That's not pessimism; that's a decent outbound funnel. It means roughly 165 sourced contacts per qualified opportunity. Whatever you charge per qualified meeting has to absorb the cost of those other 164 contacts that went nowhere, and most of them still incurred enrichment and inference spend on the way down.

This is the single most important framing in the whole teardown, and it connects directly to the broader GaaS debate about defining cost-per-completed-task as the category's core unit. For an SDR agent, "task" is ambiguous until you pin it to a funnel stage, and the stage you pick changes your COGS by an order of magnitude.

Cost Stack: Walking the P&L Line by Line

Let's build the cost of serving a single sourced contact through a multi-touch sequence, then roll it up.

Inference: Smaller Than You Think

Here's the counterintuitive part. The LLM bill is usually not the dominant cost, and vendors who obsess over it are optimizing the wrong line.

A personalized cold sequence is maybe 4-6 emails plus a couple of LinkedIn touches. Generating each touch, pulling context, drafting, self-critiquing, regenerating once, runs somewhere in the 8,000-20,000 token range per touch when you include the research context window. Call it 80,000 tokens of input-heavy work across a full sequence per contact, plus reply handling. On a mid-tier frontier model at blended rates in the low single-digit dollars per million tokens, you're looking at roughly $0.10-$0.40 of raw inference per fully-sequenced contact. Reply handling and meeting negotiation add a bit, but those only fire on the ~3% who engage.

The catch is retries and reasoning. If your agent uses extended "thinking" tokens to personalize, or it loops because a tool call failed, that $0.20 can quietly triple. This is the dynamic explored in the hidden cost of retries, when one task becomes fifty model calls, and it's real here, particularly in reply handling where the agent has to interpret ambiguous human responses and may thrash. Even so, inference for an SDR agent typically lands at 15-30% of true COGS. Anyone telling you tokens are the whole story is either subsidizing the rest or not measuring it.

Data and Enrichment: The Quiet COGS

This is the line that eats SDR agents alive, and it's the one demo decks never show.

Every sourced contact needs a verified email, ideally a mobile number, role confirmation, and increasingly a "trigger" signal (job change, funding round, tech-stack signal) to make personalization land. Enrichment and verification providers charge per record, credits that, depending on data depth and provider, run anywhere from a few cents to $0.30+ per fully enriched contact. Email verification alone, to keep bounce rates under the thresholds that protect deliverability, is its own per-record cost. The economics of buying B2B contact data have been well documented in analyses of the data enrichment and sales intelligence market, and the per-record costs are structurally sticky because the data decays, roughly 30% of B2B contact data goes stale annually as people change jobs.

Here's why it compounds: you enrich the whole list, not just the contacts who convert. Going back to the funnel, if you enrich 10,000 contacts at $0.15 each to eventually produce ~60 qualified meetings, that's $1,500 in data, or $25 of enrichment per qualified meeting before you've written a single word. Data is frequently the largest single line in an SDR agent's COGS, often rivaling or exceeding inference. It's also where margin discipline lives: caching enrichment, deduping across customers, and refreshing only on signal, the same quiet levers discussed in caching, memory, and agent gross margin.

Deliverability Infrastructure

You cannot send 10,000 cold emails a month from one domain and expect them to land. Doing outbound at scale means a fleet of secondary domains, dozens of warmed inboxes, rotation logic, and ongoing reputation monitoring. Inbox providers and warmup tools charge per mailbox per month; at the volumes a real SDR agent runs, this is a recurring four-figure infrastructure cost spread across customers, and it scales with volume, not with success.

This line is insidious because it's mostly fixed-and-rising relative to outcomes. The 2024 bulk-sender requirements from Google and Yahoo on email authentication and spam thresholds raised the floor: SPF, DKIM, DMARC, one-click unsubscribe, and a hard spam-complaint ceiling of 0.3%. Cross that line and your domains burn, your send volume collapses, and your cost-per-meeting spikes because the denominator just cratered. Deliverability isn't a setup cost; it's an ongoing tax on every send, and it's the part of the SDR-agent business that looks least like software and most like running a logistics operation.

The Human-in-the-Loop Tax

Almost every credible SDR agent at scale has humans in the loop, and almost no pitch deck admits how many. Someone reviews edge-case replies. Someone handles the "actually, I meant next quarter" responses that the agent misreads. Someone monitors deliverability and pauses domains. Someone QAs personalization before the truly cold, high-value accounts get touched.

If one reviewer can oversee the output for, say, 3,000-5,000 contacts a month, and that reviewer is loaded at $4,000-$6,000/month, you're adding $1-$2 of human cost per contact, which, divided across the ~0.6% that become qualified meetings, can be $150-$300 of human oversight per qualified meeting in a poorly-tuned operation. The whole GaaS thesis rests on driving this toward zero, but the human-intervention rate as the new churn signal cuts both ways: it's also your largest controllable cost. Vendors who've truly automated reply handling have a structural margin advantage that's invisible in a demo and decisive in a P&L.

A Worked Example: 10,000 Contacts a Month

Let's make it concrete. One customer, 10,000 sourced contacts/month, a funnel that yields ~60 qualified meetings. Rough monthly COGS:

Line item	Per contact	Monthly	Per qualified meeting
Inference (full sequence + replies)	$0.22	$2,200	$37
Data / enrichment / verification	$0.15	$1,500	$25
Deliverability infrastructure	$0.08	$800	$13
Human-in-the-loop (well-tuned)	$0.10	$1,000	$17
Platform / orchestration overhead	$0.05	$500	$8
Total COGS	$0.60	$6,000	$100

So the fully-loaded cost to produce one qualified meeting is around $100 in a reasonably efficient operation, and notice that inference and data together are more than half of it, with everything else still material.

Now price it. If you charge $300 per qualified meeting, you're running a ~67% gross margin, healthy, and roughly what a sustainable GaaS business should target, in line with what a healthy GaaS gross margin looks like in 2026. If you charge a flat $2,000/month subscription and the customer pushes 10,000 contacts, your margin collapses to negative on a heavy month. And if you guaranteed meetings and the funnel underperforms, say 35 qualified meetings instead of 60, your cost-per-meeting jumps to $171 against a $300 price, and you've quietly halved your margin without changing a line of code. That sensitivity is the whole game.

Where the Margin Actually Leaks

Three leaks turn a 67% gross-margin model into a 20% one, and they rarely show up until you're at scale:

The no-show. You can book a meeting and still lose money if the buyer pays per held or qualified meeting and the prospect ghosts. No-show rates of 25-40% are normal in cold outbound. If your contract pays on "held," every no-show is fully-loaded cost with zero revenue. Vendors who price on "booked" and customers who pay on "held" are not in the same business, and the contract language is where margins are won or lost.

Funnel decay over a contract. Month one, you're hitting fresh ICP accounts. Month six, you've burned through the best-fit list and you're enriching and sequencing lower-quality contacts to hit the same meeting quota. Cost-per-meeting rises monotonically across a contract unless the customer keeps feeding new total addressable market. This is the SDR-specific version of net revenue retention questions for agents, usage may hold steady while profitable usage quietly collapses.

Token volatility. Your inference cost is denominated in a unit you don't control. A model price change, a switch to a more expensive reasoning mode, or a customer demanding deeper personalization can move your inference line 2-3x. Budgeting against this is a known hard problem, covered in the token-volatility problem of budgeting when inference costs swing weekly.

Pricing Models and Which One Survives Contact

Three structures dominate, and they distribute risk very differently:

• Per-qualified-meeting (per-outcome). Cleanest story, hardest to deliver profitably. You eat all funnel risk. Works only if you can define "qualified" tightly enough that you're not paying for junk, and price with enough buffer to absorb bad months. The measurement problem here is genuine, see the broader debate on whether per-outcome pricing can actually measure the outcome.
• Per-sequenced-contact (per-task). You charge for work done, not results. Margins are predictable; the customer eats funnel risk. Easier to run, harder to sell, because the customer wants outcomes, not activity.
• Subscription + volume cap. A flat fee with a contact ceiling. Protects you from runaway usage but caps your upside and invites the customer to max out the cap every month.

The honest answer, drawn from watching these models meet reality: most durable SDR-agent businesses land on a hybrid, a platform fee that covers fixed deliverability and oversight cost, plus a per-qualified-meeting component priced with a real margin buffer. Pure per-outcome pricing is a great logo-acquisition tool and a slow way to go broke if your funnel assumptions are even slightly optimistic. The analysts at a16z have written extensively about how AI application businesses are rethinking pricing around outcomes rather than seats, and the SDR agent is the sharpest version of that tension because the outcome is so easy to count and so expensive to guarantee.

Insights Most People Overlook

1. Your most expensive contacts are the ones who almost convert. A contact who replies, asks two questions, schedules, and then no-shows has consumed full enrichment, a full sequence, and reply-handling inference (the most token-expensive part), and produced nothing. These near-misses, not the silent majority who ignore you, are where inference COGS concentrates. Cheap contacts are the ones who never reply.

2. Deliverability is a denominator risk, not a cost line. Everyone budgets deliverability as infrastructure spend. The real danger is that a single spam-complaint spike doesn't just cost money, it torches your send capacity, which collapses the meeting denominator and makes every other cost per-meeting spike at once. It's the one line item that can take your unit economics from healthy to underwater in a week, and it's correlated with nothing else in the model.

3. "Qualified" is a margin lever the customer controls. Whoever defines "qualified meeting" controls your COGS. If the customer's AE team tightens the qualification bar mid-contract, and they will, once the AEs complain about junk, your cost-per-qualified-meeting jumps overnight with zero change to your agent. Smart vendors negotiate the qualification definition as carefully as the price, because it is the price.

4. The agent's quality improvements can raise your costs. Counterintuitively, a "better" agent that personalizes more deeply uses more research context and more thinking tokens per touch. If that lifts reply rates by 10% but inference cost by 40%, you may have improved the funnel and degraded the margin. Quality and unit cost are not aligned by default, you have to measure both, which ties directly into building a GaaS metrics dashboard every operator should track.

5. Per-outcome pricing leaks information to your competitors. When you price per qualified meeting, you're publishing an implied cost structure. Sophisticated buyers (and competitors who buy a contract to reverse-engineer you) can back out your funnel assumptions and margin from your price. Per-task pricing hides this; per-outcome pricing broadcasts it.

References

• Gartner, Sales Technology and Sales Intelligence
• Google Workspace Admin Help, Email Sender Guidelines (Bulk Sender Requirements)
• Andreessen Horowitz (a16z), The Economics of Generative AI