All Blogs
Discover how inbound AI SDRs automate lead qualification, real-time website engagement, and meeting scheduling. Learn setup, routing, integrations, and pricing.
Sachin Gupta
Co-founder & CEO
Published On
An inbound AI SDR is software that works the leads coming to you — website visitors, form fills, webinar attendees, event scans — the way a good SDR would: figure out who they are, whether they fit, how ready they are, and what should happen next. Decides whether the lead is worth a rep’s time, and either books the meeting/directs them to self sign-up, keeps working the lead, or hands it to a person with the context needed to continue.
This guide is for teams that have decided inbound is where AI belongs in their funnel, or are trying to work out whether it is. It covers what the work actually consists of, how to set the system up, what it needs to connect to, how to run it once it is live, and why pricing here follows different logic than the outbound tools.
If you are still deciding between inbound and outbound, or have not yet settled what an AI SDR is, start with our Guide on AI SDR agents and come back.
What an inbound AI SDR is, and how it differs from outbound
An outbound AI SDR starts with a list and tries to create an engagement that did not exist. An inbound AI SDR starts with a person who has already done something — visited the site, submitted a form, attended an event, downloaded an asset — and tries to turn that action into a qualified opportunity..
There are three differences that matter most.
The buyer has come to you. The inbound lead has done something first — visited, filled a form, attended, downloaded. You are not breaking into a cold channel. And while you don’t have the challenge of activating a new engagement, what you are up against is timing. A website visitor needs a response in seconds, a form fill in minutes, a webinar attendee within a few days, so the SDR needs to act fast and with context.
You already know something about them. In cold outreach, everything about the prospect has to be inferred, because they haven’t indicated any interest. But with inbound they always tell you something — which page they read, what they put in the form, which session they sat through, what they asked. Once you know who they are, combined with what your systems already hold: CRM history, prior touches, product usage, open deals, the context becomes even richer. Contrast that with outbound where the hardest challenge is to figure out who is likely to be in the market and why your solution is a fit for them.
You qualify as you go. Outbound has to qualify before contact – at list construction, because we don’t know if the prospect will engage so we have to be intentional about who we engage. Inbound, on the other hand, qualifies through the interaction itself — on the site, over email, in a follow-up thread. The agent updates what it assesses about the lead each time something new comes in.
That’s why inbound is where autonomous agents actually work today: there’s intent and context to act on. In outbound, AI is best as a layer under the rep — signals, research, drafts — while a person decides who to reach out to and when. Inbound is where the agent can run whole conversations on its own.
Inbound SDR isn’t one job
We talk to two kinds of marketing leaders who both say ‘our inbound SDR’ and mean completely different things. One runs several thousand form fills a month against a $12K product, and her SDRs do high volume lead qualification and routing. Qualification here is generally rule-based and speed-to-lead is the most important metric. The other gets forty hand-raisers a quarter against a $200K product, and her team’s real work is doing deep research on each lead, more nuanced qualification, and often engaging the lead(s) from a company across multiple channels – website visitors, webinar attendees, event badge scans and such to eventually convert them to hand-raisers. Both are describing inbound SDR work, but tey are two very different motions.
So “can an AI SDR handle inbound?” is the wrong question. Inbound is at least three jobs, split by what the lead is missing, and each one automates to a different degree.
Hand-raisers
Demo requests, pricing inquiries, contact form submissions all fall under hand-raisers. “They want to talk. The job is deciding whether they should, then booking them or turning them away.
This is the most automatable of the three. The qualification logic is explicit — you already know what makes someone a fit. The qualification process can vary in complexity, with a simple rule-based logic on the simpler end to something that requires deep account research, but in almost all cases it can be mapped neatly to an AI system, subject to the right data being available.
Two things make it harder than it looks.
The first is that the primary work is often turning people away. A high-volume funnel produces more people who want a demo than people who should get one, and the agent’s value is as much in what it keeps off a rep’s calendar as what it puts on it. So disqualificaton and handling the lead downstream from that is an important skill.
The second is the visitor who has questions but has not yet asked for a meeting. Answering well and engaging properly is what creates the hand-raiser. An agent that can only qualify people who arrive pre-qualified will miss most of these buyers.
Where it goes wrong:
Incomplete or Excess Qualification: There is a fine balance here, you need enough information (outside of what can be inferred from enrichment and signals) that you are able to qualify well, but you don’t want to ask too much, lest you push people away.
Enrichment: A lot of qualification can be done purely based on enrichment and third-party signals. But having a good source of data is foundational. A great agent with poor data will perform poorly.
Visitors with attention but no intent
Someone’s on the site reading and comparing, but not ready to talk. You have their attention, not their intent, and you may only know their company, or nothing at all.
The work here is provoking intent. The job is creating intent: showing the proof that matches the page they’re on, answering the question they haven’t asked yet, or just giving them what they came looking for. It’s closer to working a trade show floor than working a form fill.
This job barely existed before AI, because staffing a person to watch the site and catch browsers at the right moment never made economic sense. It is the part of inbound where the agent is doing something new rather than something a person used to do, and in our experience it’s where most of the new pipeline is coming from.
Two things about it are easy to get wrong.
Relevance is inferred here, from company resolution and behavior, rather than known — so the agent is often acting on account-level information rather than a confirmed contact, and the system should be built knowing that.
The failure mode is the opposite of hand-raisers. The risk is not that the wrong person reaches a rep, it is that the right person leaves without anyone noticing. So it’s important to have systems in place that catch missed engagement.
Leads with identity but no attention
Webinar attendees, asset downloads, event badge scans, etc. are what fall here. Identity is known and system context is rich, but these are not hand-raisers yet and there isn’t a live moment to work with.
The work is research and personalization to earn a reply that turns a lead into a hand-raiser. What a good SDR does with these leads is: figure out which of two hundred webinar attendees is worth pursuing, work out what would make that person care, and create a reason to talk that doesn’t exist out of the box.
This is the job that most resembles outbound, and it inherits outbound’s problems. There is no stated intent to qualify against, the message has to earn attention, and the reply rate is constrained by technical foundations like domain reputation, sending capacity etc.. The advantage over cold outbound is that the lead is already activated – they have had an engagement with you, it’s about how that activation can be converted to intent.
Why the three matter
Don’t look at them as stages of a funnel or levels of difficulty. They are different jobs with different failure modes, and every company has all three in different proportions. High-volume SMB motions are mostly hand-raisers, which is why those teams get fast results from automation and describe their SDRs as lead processors. Enterprise motions with few hand-raisers hold most of their addressable pipeline in the second and third jobs, which is harder work and automates less cleanly.
A meaningful share of the disappointment with AI SDRs comes from this. A team that bought a tool built for hand-raisers but directed it at a list of webinar attendees was bound to be a failed pilot as it was using the wrong tool for the problem at hand.
Before setting anything up, work out what your inbound is actually made of. If seventy percent of your volume is hand-raisers, the setup is mostly qualification and routing. If seventy percent is latent leads from events and content, the setup is mostly research and outreach. Even if the same system can cater to both these motions, the steup for each is different and requires you to be intentional about it.
Setting it up
Most of the setup isn’t technical. It’s deciding what the agent should do, in enough detail to configure it. The part that usually takes more time are the decisions your team needs to make about how you want your AI SDR to operate.
Qualification criteria
Start with what disqualifies, because it is easier to agree on and because the agent will apply it more often. Company size below a threshold, industries you do not serve, regions you cannot sell into, competitors, students, job seekers. Write these as rules the agent can check against enrichment data without asking anything.
Then write what qualifies, which is harder. Most teams have a fit definition that lives in a rep’s head and a scoring model that lives in the marketing automation platform, and the two often disagree. Reconcile them before configuration, because automation will amplify the problem. The best way is to take a representative set of leads that already exist in the system and work closely with the sales team to understand which leads qualify and more importantly why.
Once you have the qualification logic, decide what the agent is allowed to ask and what it can get from other sources. Every qualifying question costs a little goodwill, and every additional form field adds to drop-off. A good rule: the agent asks only what it cannot infer, and it never asks something the person already told you (not just in the conversation but anywhere in your CRM/MAP).
Routing: relevance and intent
Who handles a lead comes down to two things.
Relevance is how well the lead fits your ICP. It comes from firmographics — size, industry, geography, technology — plus signals that firmographics miss: recent funding, hiring patterns, leadership changes. Relevance is what research is for, and it is why an agent without good enrichment is just a good messaging tool.
Intent is how ready they are. Half of it comes from their actions: what they asked, what pages they browsed, did they fill a form, what did they say etc. The other half comes from their situation, and this is the half ordinary lead scoring misses. A VP of Marketing three weeks into a new role is high intent even before she visits your website, because she has a mandate, and propensity to drive change.Lead scoring that’s only based on behavior won’t see her until she clicks something, and that’s a missed opportunity.
Your routing rule depends on relevance and intent.
Low intent, at any relevance level, should go to the agent. This is where you are likely to have high volume, and nothing is lost by letting the agent work it. In fact, it’s highly inefficient to throw human labor at these categories of leads/accounts.
High intent and low relevance goes to the agent, and disqualification is a legitimate outcome here. The person wants to talk but is unlikely to buy. A rep’s time would be wasted. The conversation still has to be handled well — they may be wrong about their own fit, or may refer to someone who is not — but it does not need a person.
High intent and high relevance goes to a person, with the agent handing off.
Most vendors pitch AI SDRs one of two ways. Either the agent replaces reps outright, or the work gets split by task: the agent researches and drafts, a person reviews and sends. The first oversells what AI can do. The second puts a human in the path of every lead, which caps throughput.
We think about it differently. Based on intent and relevance, the agent runs some conversations end to end, and runs others up to a point before handing off to a rep..
Training the Agent: The knowledge base
The knowledge base is the input teams most often get wrong. More than the underlying language model, it’s the knowledge base that determines the quality and accuracy of the agent’s responses. It deserves the most attention during set up, as you not only decide what you are going to put in the knowledge base but also how you are going to keep it updated.
What goes in it: Product documentation, website content, positioning, use cases by segment, competitive comparisons, FAQs, integration details, and importantly what you are willing to say publicly about sensitive topics like pricing, security, and contracts. The last category is the boundary of the escalation rules below as the knowledge base defines what the agent can answer, and the escalation rules define what it must hand over to the team. It’s important to build them together.
What does not belong in it: Most AI SDR agents allow you to add your entire website to the knowledge base. It’s very tempting to do that as it’s easy, but that’s likely to degrade the agent rather than improve it. The website often has old content (like old blogs or news pages), duplicate content and information that is not relevant to the agent (like technical documentation for product users). More material in the knowledge base means more retrieval collisions, and more instances of the agent surfacing dated or conflicting information to the prospect.
To train the agent satisfactorily, you should curate the information that goes into. The website can be a good starting point but only add those pages that are required. Check if your tool allows you to add exclusion and inclusion lists for webpages, that way you can always include key pages and exclude irrelevant pages like blogs, or news. Any information that is not available on the website should be added as documents (ensure the platform allows you to do that) and finally some information like case studies may need special treatment so that the agent is able to pick the right case study for the prospect.
How to structure the knowledge base: As much as possible, have one topic per document/web page. Wherever information is split across multiple documents ensure there isn’t any conflicting information. Long documents are fine as long as the content is neatly structured and grouped according to topics. Most agents segment the knowledge base into smaller chunks, if unrelated information ends up in the same chunk it can result in poor information retrieval. Duplication doesn’t degrade quality but should be avoided whenever possible.
Whenever adding FAQs, the questions should be phrased the way buyers actually ask them, not the way product marketing typically writes them — “does this work with HubSpot” instead of “Integration Ecosystem Overview.”
Who maintains it. A knowledge base goes stale over time – product launches, change in functionality etc. It’s important to have a designated owner who will keep the knowledge base updated and the configuration up to speed. If your tool allows you, set up an evaluation query set that you run periodically against your agent. This helps you catch quality deviation.
Escalation rules
Routing and the knowledge base cover the normal cases. A few edge cases need explicit rules.
VIP visitors. A CRO from your target account needs special attention. If they are passively researching, they are fine interacting with an agent but if they exhibit intent you would want to escalate right away. The agent might handle it fine. You just don’t want to risk a conversation that is valuable. The way to handle this is to set a threshold — by title, by account size, by whatever your deal data predicts a large opportunity — above which the agent’s job is to hand off the conversation to the rep.
Non-public information. Pricing is the obvious one, but also roadmap, security specifics, contract terms, and anything the company has decided not to publish. An agent hallucinating here creates an issue of false commitment. Ensure that your tool allows you to set strict guard rails to prevent hallucination. If needed add documentation on how to handle such questions.
An explicit request for a human. If you have live reps available then ensure SLAs are followed and an explicit request for humans is honored immediately. There should be a smooth process for a live hand-off and proper setup to handle conversations that come during off business hours.
Handoff design
A handoff is how the lead comes to your team, and is an expensive failure point. There are two different handoff modes and a good AI SDR supports both.
Real time. The rep is pulled into the live conversation while the visitor is still there. This is the highest value but hardest to operate. It requires reps to be available, proper SLAs with the SDR team, a presence system that knows who is available, and a fallback when nobody is. This is the best handoff for the highest value prospects but noisy for all others.
Passive. The agent books the meeting and sends the rep a briefing. This is the lowest friction, and often the most common mode that most tools support.
The handoff sounds simple, but it’s where most setup time goes. It has to:
Needs to be aware of account ownerships – to decide how to route owned accounts
Understand routing logic – to decide how to route accounts/leads that are not owned.
Have awareness of business hours and team availabilities to route live conversations
Should be able to read calendars to book meetings – It should offer its own calendaring system or be able to integrate with existing ones.
Should be able to send emails as follow-ups.
The other critical piece about the handoff is passing on the rich context to the rep. It is not sufficient to just create a calendar entry; the rep also needs to know what the person asked, what the agent inferred about relevance and on what basis, what it inferred about the intent and why, what it has already told them and what information hasn’t been answered. All this allows the rep to go into the discussion well prepared and to seamlessly take over the conversation from where it ended.
What an Inbound AI SDR integrates with
An inbound AI SDR depends on what it can read from and write to. These six integrations matter most.
Bi-directional CRM Sync. Reading is the easy half. The agent should also write: conversations, enrichment, qualification outcomes, status changes, the reasoning behind decisions. All of it should be written back to the lead and contact records, so reps see a complete history and the next interaction starts with rich context.
Marketing automation. The MAP holds engagement history the CRM does not — email opens, webinar attendance, content downloads, scoring. For the latent-lead job this is the primary source of context, an agent that cannot read it is missing out on relevant past context that could be invaluable.
Enrichment. Firmographics and enrichment from a third-party source is critical for qualification. Look for an agent that handles this natively rather than expecting you to pipe it in, and budget for it separately, since enrichment credits are usually billed apart from the platform.
Identity resolution. This is specifically important for the browsing-visitor job. IP-based company resolution, session behavior, cookies matched against prior visits, and — where available — person-level identification. Quality here determines everything downstream, because an unidentified visitor cannot be scored against your ICP. Ask how the vendor does it and what match rate to expect on your traffic.
Calendar and routing. The agent needs to follow territory and ownership rules so escalations reach the right rep, and calendar access so bookings land correctly. Ownership logic is usually more complicated than teams remember — overlapping territories, named accounts, round-robin exceptions — and the tool should be able to capture your business logic. Many teams have an existing tool for lead routing, so make sure that the tool works well with existing systems and doesn’t result in double routing of leads.
Slack or Teams. For real-time handoff and for alerts. The agent needs a way to reach a rep in the seconds a visitor is still on the site, so a robust integration with your communication tool that ensures the right rep is alerted about leads in a way that is not missed becomes critical.
Running it
What to monitor
Escalation rate. A jump in escalations could mean more good leads are coming in, or that the agent is escalating things it should handle. A drop could mean it’s handling more conversations well, or that it’s missing leads. You have to read the conversations to know which one is the issue.
Answer quality on a sample. Either run evaluations against your baseline query set or pick real conversations and read them. While it may sound like a lot, it only takes 10-15 mins to go through conversations. They would often reveal interesting things about what your buyers are asking and how they are engaging. Or you could use an AI agent to analyze these conversations to look for insights. Look for wrong answers, stale information, and questions the agent should have escalated and did not. This is also how you find knowledge-base gaps and can inform your content planning.
Meeting show rate and downstream conversion. Booked meetings are not a good enough metric, measure the no-show rate and ultimately opportunity created – both by volume and dollars.
Rep feedback. Collect feedback from reps on the handoff process, the context they are getting for the leads, and how the overall process is for them.
The platform landscape
The inbound AI SDR category consolidated sharply in 2026. Drift, which created the conversational marketing category, was sunset following its acquisition by Salesloft, with 1mind designated as its successor. Key players like Qualified and Warmly were also acquired by Salesforce and HubSpot, respectively and absorbed into broader platform suites, diluting their pure inbound specialization. Most teams evaluating inbound tools today are either migrating off these platforms or building the function for the first time.
The main platforms, as of September 2026:
Qualified — acquired by Salesforce and absorbed into its broader suite, which diluted its pure inbound specialization. Still widely used in large Salesforce shops with established routing and account-based programs.
Warmly — acquired by HubSpot, integrating its person-level visitor identification and live handoff into HubSpot’s platform suite, though shifting focus away from dedicated inbound SDR.
Knock AI — consolidates several layers of the inbound stack (chat, forms, scheduling, event capture, de-anonymization) into one platform, and pitches on converting demand you have already paid for.
1mind — Drift’s designated successor, built around avatar-based AI sales conversations across lifecycle interactions. The most different in approach from the others on this list.
Conversica — the longest-tenured, strongest for high-volume email follow-up on latent leads rather than live website conversation. Custom-priced.
Fin (Intercom) — a single agent for both support and sales conversations. Best fit where product education and support questions are a large share of the buying process.
Chili Piper — routing and scheduling rather than a full agent, but frequently the routing layer underneath other tools.
Breakout — an inbound AI SDR focused on live demand capture: engaging visitors in real time, qualifying autonomously, and booking meetings. Categorized by third-party roundups alongside Qualified, Warmly, and Knock.
Most of these vendors publish comparison pages about each other, and none of those pages is independent. Read them for feature lists and ignore the rankings. And the category is moving fast enough that any list, including this one, is out of date within a quarter — check the date at the top of this section before relying on it.
How Inbound SDR pricing works
Almost everything written about AI SDR pricing assumes outbound: a list, a send volume, a conversion rate. Per-contact pricing, credit packs, cost per meeting. Inbound doesn’t work that way.
Send volume is not the primary metric, even if the SDR is following up over email. There is no list, because the population is whatever traffic/leads arrive. So per-contact and credit-based models don’t fit.
The natural unit on inbound is the conversation (whether on the site or over email), and more precisely the resolved conversation — one that ended in a booking, a qualified handoff, a disqualification, or a clean close. Models that price on conversations align with the work the agent actually does. Models that price on meetings booked align with only one of the four outcomes, and create an incentive to book meetings irrespective of the quality.
Per-seat pricing, which several vendors inherited from the conversational marketing era, is the worst fit of all. Seats made sense when a person was behind every chat. They don’t make sense when all conversations may not be routed to your team. It still makes sense to have some element of user seats in the pricing but it should not be the primary vector.
The practical questions to ask: what is the unit of work, what counts as valid, and how does pricing scale. If a vendor can’t answer these clearly, especially how cost scales, that tells you something. A full cross-vendor comparison is in this AI SDR pricing guide, which covers both sides of the funnel.
Frequently asked questions
What percentage of inbound can an AI SDR handle on its own? It depends entirely on what your inbound is made of. A funnel dominated by hand-raisers automates far more completely than one dominated by event and content leads.
Should the agent be allowed to disqualify leads? Yes, with a higher bar than it uses for escalation. A wrong escalation costs a rep 20 minutes. A wrong disqualification can cost you a $50,000 deal. Set the disqualification threshold conservatively, and read a sample of disqualifications in the first two months to ensure the agent is operating as expected.
How does it handle anonymous visitors? An AI SDR handles anonymous visitors through identity resolution — company is determined from IP, behavior from the session and is matched against prior visits and form fills. Match rates vary a lot by vendor and by traffic source. Ask for the expected rate on your traffic before you buy, and build the routing rules knowing that resolution may be incorrect and relevance on anonymous visitors is an inference and not 100% accurate.
What happens when someone asks about pricing? This depends on what you have configured, and what exists in the knowledge base. The right answer is usually that the agent shares what is public and hands off for anything that is not. This is the single most common escalation trigger and worth designing explicitly.
How long does setup take? Configuration takes days. If a vendor needs months, that’s too long. The slow part is the decisions that come first: what qualifies, what disqualifies, when to escalate, what reps receive, what goes in the knowledge base. Teams that have those written down go live in weeks. Teams that don’t spend the implementation figuring it out.
Can it replace our inbound SDR team? It can absorb a lot of the volume and most of the hand-raiser work, and can take almost all of your lead outreach work. It should not take up senior stakeholder conversations, complex evaluations, or anything requiring real discovery, and it cannot handle phone. The pattern that comes up most often is reallocation and not replacement: the agent takes up all conversations that can be defined through configurations and knowledge base, and reps move to full-cycle selling and the conversations that need a person.
How is this different from our website chatbot? A chatbot follows a decision tree. An inbound AI SDR answers unscripted questions, researches mid-conversation, and decides what should happen next — including deciding that a conversation should end or that a lead is not worth pursuing.
Does it work for outbound too? The same agent can usually support outbound as an automation layer — signal tracking, research, drafting — but running outbound conversations autonomously is a different product with different constraints.
Breakout builds an inbound AI SDR. The framework in this guide is how we think about the work. If you want to see how it behaves on your own traffic, start here.
