AI sales agents: what an AI BDR actually does in 2026
An AI sales agent is software that does a piece of sales work on its own: it researches an account, writes the first email, reads the reply, or updates the CRM without a rep doing it by hand. This page breaks the job down into the parts agents handle today, shows where a person should still check the work, and covers when to buy a packaged AI BDR and when to build your own.
Key takeaways
- "AI sales agent" covers any agent doing sales work. "AI SDR" and "AI BDR" are the subset that runs outbound up to a booked meeting.
- Agents are reliable at research, list building, scoring and CRM updates. They still need a person on ambiguous replies and on anything sent in a rep's name.
- Buying (11x, Artisan, AiSDR) gets you running fast. Building on Claude Code, your CRM and your sending tools gives you control over targeting and cost.
- Judge any agent on accepted meetings or qualified deals, never on emails sent.
What is an AI sales agent?
An AI sales agent pairs a language model with access to your data and tools, then lets it act. A chat window where a rep pastes in a company name is an assistant. It becomes an agent when it starts on its own trigger (a new lead, a reply, a job change) and acts in another system without someone pasting the result across.
The names overlap. AI SDR and AI BDR mean the same thing in practice: an agent that does a development rep's job, from finding prospects to booking a first meeting. Artisan calls Ava an AI BDR, and 11x calls Alice an AI SDR. AI sales agent is the wider term and also covers agents that answer inbound leads, route them, prep calls, or keep the CRM clean. If you want a list of AI SDR products with prices, that is on AI SDR tools. This page is about the work itself.
The jobs an AI sales agent does today
"AI BDR" sounds like one product, but it is really a chain of separate jobs. Agents are good at some of them and shaky at others, so it helps to look at each one and decide where a person signs off.
| Job | What the agent does | Human checkpoint |
|---|---|---|
| Account research | Reads the company site, news, job posts, funding and tech stack, then writes a short brief with sources. | Spot check a sample each week for wrong facts or stale news. |
| List building | Pulls companies and people that match your ICP filters from data providers, dedupes against the CRM, and drops current customers. | Approve the ICP filters and exclusion lists before the first pull. |
| Scoring | Scores each account and person on fit and recent signals, so only the top of the list gets contact data and outreach. | Review the scoring rules against deals you have won and lost. |
| Signal monitoring | Watches for hires, promotions, job posts, site visits or product events and queues the account when one fires. | Decide which signals count and how fast reps act on them. |
| Personalization | Writes a first email or LinkedIn note around one real signal, in your voice and inside your copy rules. | Approve the templates and rules, then read a sample of live sends every week. |
| Sending | Loads approved leads into Instantly or HeyReach sequences at a safe daily volume. | Someone owns domain health and send limits. |
| Reply handling | Reads each reply, labels it (interested, not now, wrong person, unsubscribe), and drafts the response. | A person approves every response before it goes out. |
| Booking | Offers times, sends the calendar link, and routes the meeting to the right rep. | The rep confirms the meeting is worth taking. |
| CRM updates | Writes the brief, score, signal, touches and outcome back to the CRM record. | Lock which fields the agent can write, and audit them monthly. |
The checkpoints are where OneGTM puts a person in the systems we build. Packaged AI SDRs let you set some of them and skip others.
Where AI sales agents work
Agents earn their keep on high volume work with a clear right answer, where a small mistake costs little. Researching 2,000 accounts before a campaign is a good example. So is scoring a list, or writing the same twelve CRM fields after every reply. A person doing that work gets bored and sloppy by account 200; an agent does account 2,000 the same way as account one.
They also work well on speed. An agent can read a reply the minute it lands and have a draft waiting in Slack, so a rep answers an interested prospect in minutes instead of the next morning. Inbound agents do the same for form fills and website visitors, which is why 11x sells a separate inbound agent (Julian) next to its outbound one.
And they make signal-based outbound practical. Watching hires, promotions and job posts across a few thousand accounts every day is impossible by hand. An agent can do it and only surface the accounts where something changed.
Where AI sales agents fail
- Confident wrong facts. An agent that misreads a funding round or names the wrong product in a first line burns the account. Research needs sources and spot checks.
- Copy that sounds like every other agent. Buyers now get many AI-written emails a week. Drafts built on a generic "I noticed you" template get ignored, and the fix is better inputs (a real signal, your own voice rules) plus a human read.
- Replies with nuance. "Not now, but talk to Sam in Q2" or an annoyed reply from a current customer needs a person. Fully autonomous reply handling is where most AI SDR horror stories come from.
- Burned domains. Agents that send at high volume without warmup, verification and send caps get your domains flagged, and that hurts your reps' email too. See cold email infrastructure.
- A messy CRM. An agent reading duplicate accounts and stale owners will route to the wrong rep and email current customers. Clean data first.
- No owner. Someone has to read outputs, tune prompts and watch the numbers. An agent nobody reviews drifts, and the team stops trusting it.
Buy vs build an AI sales agent
Buying means a packaged AI SDR. As of November 8, 2026, 11x sells Alice for outbound (account research and multichannel sequences) and Julian for inbound (qualifying leads and booking them with the right rep), and neither publishes pricing. Artisan sells Ava as an AI BDR that finds and enriches leads, writes and sends outreach, handles replies and books meetings, with data, deliverability and sequencing in one platform. AiSDR is the cheaper way to test the idea, priced per person prospected. Current prices for these and a dozen more are on AI SDR tools.
Building means a custom agent that runs on tools you already pay for. The pattern OneGTM uses: Claude Code writes and runs the agent, a data API like Deepline handles enrichment, your CRM (Attio, HubSpot or Salesforce) holds the record, Instantly and HeyReach send, and Slack is where a person approves anything that leaves the building. Every prompt and scoring rule is code your team can read and change.
| Question | Buy a packaged AI SDR | Build on Claude Code |
|---|---|---|
| Time to first send | Days to a few weeks | A few weeks for the first working version |
| Targeting | The vendor's database and filters | Any source and any signal you can pull, scored your way |
| Seeing why it acted | Varies by vendor; often a black box | Every prompt and score is in your repo |
| Cost at volume | Seat or contract pricing, usually annual | Model, data and sending costs you control |
| Who maintains it | The vendor | Your GTM engineer, or a partner who hands it over |
| Best fit | Broad ICP, small team, need volume now | Niche ICP, signals only you understand, an existing stack |
Plenty of teams do both: a packaged inbound agent on the website, and a custom outbound system underneath. If you have already tried an AI SDR and want the custom route, AI SDR alternatives walks through what to use instead.
How to evaluate an AI sales agent
Test it job by job, on accounts you already know, before it touches a live prospect.
- Research. Give it 20 accounts your team knows well. Count the wrong or stale facts in the briefs.
- List and score. Have it build and score a list, then compare the top 50 against your best customers and the bottom 50 against deals you lost.
- Drafts. Read 30 first emails. Flag any claim you could not defend on a call, and any line that could go to any company.
- Replies. Feed it a mix of real replies, including an angry one, a referral and an out of office. Check the label and the draft for each.
- Controls. Confirm you can require approval before sends, cap daily volume per mailbox, and limit which CRM fields it writes.
- The number that matters. Agree up front to judge it on accepted meetings or qualified deals over 60 to 90 days. Emails sent and open rates tell you nothing.
A vendor that will not let you run steps one to four on your own accounts before you sign is telling you something.
How OneGTM builds AI sales agents
We build agents as part of a client's outbound and RevOps system, on the client's own accounts, then hand them over. Some of what is running today:
- Reply desks. A Slack bot reads each reply, drafts the response, and waits for a person to approve it. Coastal Recruiting and Patlytics both ran one. At Coastal, an interested reply only counts once a recruiter accepts it, and accepted candidate intros went from 72 in April 2026 to 340 in August 2026. The full story is in the Coastal Recruiting case study.
- Signals, scoring and routing. For Antimetal we built the CRM, scoring, signals and routing behind their outbound, with leads routed through Default. Automated Outbound created 17 deals and $2.57M in pipeline from December 2025 to May 2026 (Antimetal case study).
- AI classification. SuperDial's market was scored with AI industry classification, and its AEs are assigned by call volume.
The build usually goes in this order:
- Pick the job. Start with the one that costs your team the most hours, usually reply handling or research, rather than the whole AI BDR at once.
- Wire it up. Connect the agent to your CRM, sending tools and Slack with only the permissions that job needs.
- Set the checkpoints. Decide where a person approves, using the table above as a starting point.
- Measure and hand over. Track accepted meetings and deals, then walk your team through the code and hand it over.
The longer version of the outbound build is on Claude for sales, and the wider automation work on AI automation consulting. For AI that helps reps rather than acting for them, see AI for sales.
AI sales agents FAQ
What is an AI sales agent?
What is an AI BDR?
What is the difference between an AI SDR and an AI sales agent?
Can an AI sales agent replace an SDR?
Should I buy an AI sales agent or build one?
How do I evaluate an AI sales agent?
Does OneGTM build custom AI sales agents?
Want an AI sales agent your team owns and can change?
Build an AI sales agent with OneGTM
Tell me what you're building. I reply the same day.
Pick a time
Nothing works? Write team@getaob.com and I'll find one.
Book a timeGrid not loading? Open the booking page
Rather do it the old way? team@getaob.com LinkedIn