AI Agent › Use Cases › Sales • aiagent.app
AI Lead Generation: What Works, What Breaks, and What It Costs
Most guides to AI lead generation are written by companies that sell one piece of the pipeline. That shapes what they tell you. A data vendor explains that better data is the bottleneck. A sequencer explains that sending is the bottleneck. A CRM explains that your CRM is the bottleneck. Each is describing a real problem, and each stops describing it precisely where their product stops.
This page tries to be the other thing: an account of what AI lead generation actually does, what it costs, where it breaks, and what the law requires — written for a founder or a two-person growth team who has to make the whole loop work, not just one node of it.
What AI lead generation actually means
AI lead generation is the use of machine learning and language models to do four jobs that used to require a person: finding accounts that resemble your best customers, enriching them with usable contact and context data, ranking them by likelihood to convert, and making the first contact.
The phrase gets used loosely. It is worth separating three things that are routinely conflated:
- Predictive lead scoring — a statistical model that ranks existing leads by conversion probability. This is a decade-old technique and the most mature of the three.
- AI-assisted prospecting and enrichment — using models to identify accounts, find contacts, and infer firmographic, technographic, or intent signals. This is where most of the recent tooling money has gone.
- Agentic outbound — an autonomous system that researches a prospect, drafts a message, sends it, reads the reply, and decides what to do next. This is the newest and by a wide margin the least reliable.
When a vendor says "AI lead generation," they usually mean the second. When a founder imagines it, they usually mean the third. The gap between those two is where most disappointment lives.
Adoption is real but earlier-stage than the marketing suggests. Salesforce's seventh State of Sales report, which surveyed 4,050 sales professionals across 22 countries between August and September 2025, found that 54% of sales teams use AI agents today and that 34% of those teams use them for prospecting specifically (). Salesforce sells the agents it is surveying, so read the enthusiasm accordingly — but the methodology is disclosed, which is more than most numbers in this category can claim.
The four jobs, and which ones AI is actually good at
Sourcing. Given a description of your ideal customer, find companies matching it. Language models are genuinely good at this now, because the task is mostly retrieval and classification over text. The failure mode is not accuracy but sameness: everyone querying similar models against similar databases converges on the same account list, which is precisely why response rates on "obvious" targets keep falling.
Enrichment. Attach emails, job titles, tech stack, headcount, funding, and recent triggers to those accounts. This is a data-coverage problem more than an intelligence problem. The AI part is mostly waterfall logic — try provider A, fall back to B, then C — and the quality ceiling is set by the underlying providers, not the model orchestrating them.
Scoring. Rank the resulting list. This is the job with the strongest evidence behind it and the most misleading marketing, which is worth its own section below.
Contact. Write and send the first message. Models are fluent here and that fluency is the trap: fluent, generic, high-volume outreach is exactly the pattern spam filters are trained to catch, and exactly the pattern recipients have learned to ignore.
A useful rule: AI is reliable in this pipeline in inverse proportion to how much of the outcome depends on a human choosing to reply.
How accurate is AI lead scoring, really?
The honest answer is: measurably good on the right data, and close to meaningless on the wrong data.
The best public evidence is a 2025 peer-reviewed case study in Frontiers in Artificial Intelligence, which built a B2B lead-scoring model on a real software company's CRM records spanning January 2020 to April 2024. Testing fifteen algorithms, a Gradient Boosting Classifier reached 98.39% accuracy with a 0.9891 ROC AUC ().
That number will get quoted at you as "AI lead scoring is 98% accurate." It does not mean that, and says so itself. Lead conversion data is severely imbalanced — the overwhelming majority of leads never convert — and on imbalanced data a model that simply predicts "will not convert" for everything scores extremely well on accuracy while identifying none of the leads you actually care about. The authors are explicit that headline accuracy is misleading here, and had to weigh rebalancing techniques against AUC and F1 tradeoffs rather than optimise the number that reads best.
Two practical conclusions follow. First, when a vendor quotes an accuracy figure for lead scoring, ask for precision and recall on a holdout set instead; if they cannot produce those, the accuracy figure is decoration. Second, that study is one company's historical data, not an industry benchmark. It shows what is achievable with four years of clean CRM history. It says nothing about what you will get.
What breaks at small data volumes
This is the part almost nobody writes about, because almost nobody selling lead-scoring software benefits from you knowing it.
Supervised lead scoring learns from your conversion history. The signal it needs is examples of leads that converted. A company with four years of CRM data and thousands of closed deals has that signal. A pre-product-market-fit startup with three hundred leads and eleven customers does not — and eleven positive examples spread across a dozen features is not a training set, it is noise with a confidence interval wrapped around it.
The class-imbalance problem described above gets worse, not better, as volume drops. Below roughly a few hundred conversion events, a model will happily fit the idiosyncrasies of the deals you already closed, then rank new leads by how much they resemble your first eleven customers. That is not prediction, it is anchoring, and it will systematically steer you away from any segment you have not already sold to.
For most early-stage companies the correct answer is a rules-based score you wrote yourself — company size band, a technology signal that implies need, a trigger event, a title match — reviewed monthly. It is transparent, it is debuggable, and at low volume it will usually outperform a model. Graduate to a learned model when you have enough conversions that you can hold out a test set and still have something left to train on.
What AI lead generation actually costs
Public information here is poor, and the reason is structural: cost depends on how many enrichment "credits" or "actions" each lead consumes, which depends on your waterfall depth and list quality, which nobody can know in advance. So the marketing settles on per-lead figures that cannot be checked.
What can be checked is list price. As of August 2026, taken from the vendors' own pricing pages rather than third-party roundups: Clay's entry paid tier is $167 per month for 15,000 actions, with a free tier at 500 actions (). Instantly's Starter plan is $94 per month for 5,000 emails and 1,000 uploaded contacts, with Scale at $194 per month for 100,000 emails ().
Note how quickly that composes. A conventional stack — a data source, an enrichment layer, a sequencer, and a CRM — lands most teams somewhere in the mid-hundreds per month before anyone has sent a single message, and each tool meters a different unit. You are billed per action in one, per contact in another, per seat in a third. There is no single number that tells you your cost per qualified lead, which is why every article confidently quoting one is quoting something it made up.
When we checked third-party "cost per lead" articles against live vendor pricing while researching this page, they disagreed with the vendors and with each other, in the same month, about the same plans. Treat any specific cost-per-lead claim — including a hypothetical one from us — as uncheckable unless it comes with the assumptions attached.
Deliverability is the real constraint
You can solve sourcing, enrichment, and scoring perfectly and still fail, because the binding constraint on AI-driven outbound is not finding people. It is landing in an inbox.
Google's sender requirements apply to everyone, not just bulk senders. Every sender must authenticate with SPF or DKIM and keep the spam rate reported in Postmaster Tools below 0.30%. Anyone sending more than 5,000 messages a day to personal Gmail accounts must additionally have SPF and DKIM and DMARC, and provide one-click unsubscribe. Google's guidance is to keep spam rates below 0.10% and never let them reach 0.30% ().
Those thresholds are less forgiving than they look. The 0.30% ceiling in is three complaints per thousand delivered messages. At any meaningful volume, a single badly-targeted campaign crosses it.
There is also a mechanism worth understanding, rather than the folklore version. Google's spam classifier was rebuilt around a text vectorizer called RETVec specifically to be robust to character-level and adversarial text patterns; swapping it into the Gmail spam filter improved spam detection by 38% and cut false positives by 19.4% (, paper at ). That work predates the current wave of AI outbound and was not aimed at it. But it establishes the relevant fact: modern filters classify on statistical properties of text, not on whether a human typed it. Templated generation at volume produces exactly the regularity these models are built to detect.
For calibration on what good actually looks like, Instantly — which sells cold-outreach software and is therefore an interested party, aggregating anonymised data from its own platform — reports an overall average reply rate of 3.43%, with the top decile of senders above 10.7%, and 58% of all replies arriving from the first email rather than the follow-ups (). Even taken at face value from a vendor, that is the shape of the game: a low single-digit reply rate is normal, and the first message carries most of the outcome.
The legal floor: CAN-SPAM and GDPR
Two misconceptions are worth killing directly.
"CAN-SPAM doesn't apply to B2B." It does. The FTC's compliance guide states plainly that the Act "makes no exception for business-to-business email" and that it covers all commercial messages, not just bulk mail. Each individual non-compliant email is subject to penalties of up to $53,088, opt-out requests must be honoured within 10 business days, and you cannot charge a fee or require anything beyond an email address to unsubscribe ().
"AI scoring triggers GDPR's automated-decision rules." Usually not — but the rule people reach for is the wrong one. GDPR Article 22 restricts decisions based solely on automated processing that produce legal or similarly significant effects. Most lead scoring does not meet that bar, because a human sales rep reads the score and decides what to do. The obligation that does bite is lighter and more often ignored: individuals have the right to be told how to object to profiling, including profiling for marketing purposes ().
The practical version: your scoring model probably is not an Article 22 problem, and your privacy notice probably does need to explain the profiling and how to object to it.
Why AI SDR tools have a reputation problem
It is not only marketing backlash. The first generation genuinely underperformed, and some of the people who built it have said so on the record.
Artisan's founder told TechCrunch that first-generation AI SDRs "get a pretty low response rate" and have "relatively high churn" among customers, that his own product had "extremely bad hallucinations when we first launched," and that a realistic target for AI SDR outreach is around a 1% response rate. The company also declines to sell into segments where agentic outbound does not work ().
That is a vendor describing the ceiling of their own category, which makes it considerably more credible than a statistics roundup. If you are modelling AI outbound, model it against a low single-digit reply rate and a real risk of domain damage — not against the numbers on a pricing page.
Tools versus agents: the stitched-stack tax
Here is the thing the incumbent guides structurally cannot tell you, because each of them is one node in the stack.
The default recommendation you will find — including from Google's own AI Overview for this query — is a three-or-four tool pipeline: an enrichment tool, a data provider, a sequencer, and your CRM. That works. It also means you own the integration. Every handoff between those tools is a place where a field mapping drifts, a sync silently fails, a lead gets enriched twice, or a contact who replied last week gets a cold open this week because the sequencer never learned about the reply.
Nobody quantifies that overhead, because no vendor in the chain is incentivised to. But it is the actual work: the tools are individually good and the seams between them are where the time goes.
The alternative framing — one agent that holds the whole loop and calls each system as a tool — is not automatically better. It concentrates failure rather than distributing it, and an agent that misunderstands your ICP will do so consistently across every step. What it does change is where the integration lives. Instead of you maintaining four sets of field mappings, the agent reads and writes each system directly, and the state of a lead lives in one place.
Which is right depends on volume and on how much of your week you can afford to spend on plumbing. Below a certain scale, the stitched stack costs more in attention than it saves in capability.
Running lead generation on AI Agent
To be concrete about what this looks like here, and to stay inside what is actually true of the product: AI Agent exposes 40 connected applications, and the ones relevant to this loop are Attio, HubSpot, Pipedrive, Salesforce, Gmail, Mailchimp, LinkedIn, Google Sheets, and Notion. An agent reads and writes those directly rather than through a chain of syncs.
A realistic setup is one agent with a written ICP definition, scheduled to run on a cadence: pull new signups or list additions, enrich against your CRM to check they are not already known, score them against explicit criteria you wrote, and draft outreach for the ones that clear the bar. The draft goes to you for approval before anything sends.
That approval step is the part worth defending. Given everything above — the deliverability thresholds, the CAN-SPAM exposure, and the roughly 1% reply-rate ceiling that an AI SDR vendor's own founder describes as realistic for fully-autonomous outbound () — the sensible division of labour is that the agent does the research and the assembly, and a human decides what actually leaves the building. Fully autonomous send is the feature that sounds most impressive and carries nearly all of the downside.
What to measure
Most AI lead-gen dashboards report the metrics that make the tool look busy: leads sourced, emails sent, contacts enriched. None of those are outcomes, and optimising them actively makes things worse — sourcing more and sending more is precisely how deliverability degrades.
Four numbers are worth tracking, in this order.
Spam complaint rate. Not a growth metric, a survival metric. It is the one number that can quietly end your ability to send at all, and Google publishes the threshold you are being judged against. Watch it weekly, before volume, before reply rate.
Reply rate, segmented by list source. A blended reply rate hides the thing you need to know. The same headline figure can be a healthy rate on inbound-adjacent contacts averaged together with a near-zero rate on a scraped list — and the scraped list is simultaneously destroying the domain reputation that the good list depends on. Segment, or you will draw the wrong conclusion and scale the wrong source.
Qualified-conversation rate. The share of replies that are a real conversation rather than an unsubscribe, an auto-responder, or a request to stop. This is the metric AI outbound most reliably degrades: fluent generated messages can hold reply rate steady while the composition of those replies shifts from interest to annoyance. If you track only reply rate, that shift is invisible until pipeline dries up.
Cost per qualified conversation. Not cost per lead. Total tooling spend divided by conversations worth having. It is the only figure that composes across a multi-tool stack with four different metering units, and it is usually an uncomfortable number the first time anyone calculates it.
If lead scoring is in the loop, add a fifth: the precision of the top scoring band. Of the leads your model ranked highest last quarter, what share actually converted? That is the question the model exists to answer, and it is answerable from data you already have.
A maturity ladder
If you are starting from nothing, the order matters more than the tooling.
- Write the ICP down. Not a persona document — a list of testable criteria. Most AI lead-gen failures are specification failures wearing a technology costume.
- Score by rules, manually. Apply your criteria by hand to fifty leads. If your rules do not separate good from bad on data you already have, no model will fix that.
- Automate the research, not the sending. Have an agent assemble the account brief, the contact, and the trigger. Keep the writing and the send decision with a human.
- Instrument deliverability before you scale volume. Authenticate the domain, set up Postmaster Tools, and know your spam rate before it matters.
- Learn a model only when you have the data for it. Enough conversions to hold out a test set, and a metric that is not raw accuracy.
Teams routinely attempt step 3 or 5 first. That is the single most common reason this does not work.
Where this is heading
The durable change is not that software writes emails. It is that the research half of lead generation — the reading, cross-referencing, and assembling that used to consume most of an SDR's day — has become cheap and fast, while the judgement half has not moved at all.
That asymmetry is likely to widen. Sourcing and enrichment will keep improving because they are retrieval problems with clear feedback. Autonomous outbound will stay hard, because its ceiling is set by recipients and spam filters rather than by model capability, and both of those adapt.
The teams who do well with this are the ones who point the cheap half at a target they have defined precisely, and keep a person on the expensive half.
How the work divides
Explore all AI agent use cases or start building on AI Agent.
ai lead generationai lead generation toolsai lead generation agentai agents for lead generationai lead scoringautomated lead generationcold email deliverabilityb2b prospecting