AI agents for GTM and RevOps · Part 2 of 13
What an AI SDR can and can’t do today
The outbound jobs it does well, the ones to keep with people, what controls each, and four ways it goes wrong.
The short version
An AI SDR does the outbound work that repeats: research, first drafts, follow-ups, sorting replies and booking meetings. Strategic accounts, objections that need judgement and anything that commits the company stay with people. A setting or an instruction you choose controls each job. And its four failures start before the email is written: in the data, the brief, the volume, or what nobody approved.
The jobs, and what controls each
An AI SDR is strongest on outbound work that repeats. In Salesforce’s survey of 4,050 sales professionals, sellers expect agents, once fully implemented, to cut prospect research time by 34% and email drafting by 36%. For each job, a setting or an instruction decides whether the agent does it alone, with a person, or not at all.
Can be the agent’s
Research the account
- Who, today
- the AI SDR
- Controlled by
- what it can see: HubSpot’s agent considers the past year of engagements linked to contact records
Write first drafts
- Who, today
- the AI SDR
- Controlled by
- your brief: who to write to, what to offer, and what it must not claim (part 5)
Send
- Who, today
- a person approves each email, at first
- Controlled by
- HubSpot’s Automation tab: Review emails, or Send without review
Follow up
- Who, today
- the AI SDR
- Controlled by
- HubSpot: the maximum emails it sends, and the minimum days between them
Read replies
- Who, today
- the AI SDR sorts them
- Controlled by
- its instructions, as in Salesforce’s guide: on an opt-out request, perform the Opt Out Lead action
Book the meeting
- Who, today
- the AI SDR
- Controlled by
- each seller’s connected email and calendar, so it can offer their booking link (Salesforce)
Keep with people
Strategic accounts
- Who, today
- a named rep
- Controlled by
- left off the agent’s lists; the account has an owner
Objections that need judgement
- Who, today
- a rep
- Controlled by
- which replies the agent hands to a rep (part 9)
Anything that commits the company
- Who, today
- a person
- Controlled by
- prices, discounts and terms: the sales series, coming later
HubSpot’s email limits are a setting: when its agent picks its own sequence steps, it follows them “regardless of any custom instructions you provide”. Salesforce’s opt-out is an instruction, which works as well as the model follows it. So test instructions before launch (part 10), and keep the limits that matter in settings or in rules outside the agent (part 8).
Objections stay with a rep for a reason buyers give. In Gartner’s survey of 632 B2B buyers, “for buying tasks requiring contextual intelligence, such as determining whether a product or service fits their company’s needs, buyers prefer to seek seller input.”
One more job stays with a person: reading what the agent sent. SaaStr, whose AI SDR sent 4,495 messages in two weeks, sets this rule, at least at the start: “Spend 30-45 minutes every single day reading a sample of what your AI sent.”
Jason Lemkin, SaaStr, 10 July 2025. Results as SaaStr reports them.Four ways it goes wrong
The failures any AI agent can have are in support part 2. These four are outbound’s own.
The wrong person
One who left, or in another role.
- The cause
- old, duplicate or wrong contact data
- Fixed in
- part 4: contact data
A “fact” that isn’t true
About the prospect, in your name.
- The cause
- a thin brief, old data, or instructions planted in a bio
- Fixed in
- parts 4 and 5; planted text: part 10
Too much mail
Complaints that harm your domain.
- The cause
- volume with no limit counted across people and agents
- Fixed in
- parts 8 and 11: limits and volume
Saying what nobody approved
A price, a promise or a secret.
- The cause
- nothing written on what it may offer or reveal
- Fixed in
- part 5, and the sales series, coming later
The same Gartner survey found that “73% of B2B buyers actively avoid suppliers who send irrelevant outreach.” Robert Blaisdell, a VP Analyst at Gartner: “Bad prospecting actively damages relationships with potential customers.”
Gartner, survey of 632 B2B buyers, August to September 2024The wrong person starts with the data: a contact who left, or a duplicate of someone already talking to a rep. HubSpot’s agent considers the past year of engagements linked to contact records; an older conversation, or one not linked to the contact, is out of its view. Fix the data first (part 4).
A “fact” that isn’t true is a risk rather than a reported case: to personalise, the agent writes from what it found, and old data or a thin brief leaves it little that is true to say. An unverified Hacker News post from July 2025 describes worse: instructions planted in a LinkedIn or web bio, which AI SDRs then “personalize” into their emails. Parts 4 and 5 fix the data and the brief; part 10 tests for planted text.
Too much mail: Google asks everyone who sends to Gmail to keep the spam rate in its Postmaster Tools below 0.3%, and Salesforce’s guide gives the agent an address on your company’s domain, the one your team sends from too. Count limits across people and agents (part 8); set the volume in part 11.
Saying what nobody approved has a reported case, from a general AI agent, not an AI SDR product: a startup founder’s cold email to Zoho founder Sridhar Vembu, pitching an acquisition, disclosed “the name of a competing bidder and the specific price they had offered”. The founder’s “browser AI agent” then wrote: “I am sorry I disclosed confidential information about other discussions, it was my fault as the AI agent.” Write down what the agent must not offer or reveal (part 5); prices and terms are the sales series’ subject.
Is a segment ready for it?
Give the agent one segment at a time, once three things are written for it. As an example, take Example Software, an invented company selling to finance teams, and one of its segments: finance directors at logistics firms with 200 to 2,000 staff.
Is the message written for them?
Why a finance director in logistics would care, what to offer, the proof it may cite, and what it must not claim. Part 5.
Is the data checked?
Roles and emails checked after Clay enriches them, and every unsubscribe in HubSpot passed to the sending tool. Parts 4 and 6.
Are the limits written down?
In its playbook: no message to anyone who unsubscribed, and at most three in any 14 days, counted across the team. Part 8.
Next in this series
Part 3: build or buy an AI SDR. See all 13 parts.
Sources
- Salesforce, “Salesforce Announces State of Sales Report for 2026”, 3 February 2026, read on 21 October 2026.
- HubSpot Knowledge Base, “Set up and use the prospecting agent”, last updated 14 October 2026, read on 21 October 2026.
- Salesforce, “Agentforce for Sales Development Implementation Guide” (PDF), May 2025, read on 21 October 2026.
- Gartner, “Gartner Sales Survey Finds 61% of B2B Buyers Prefer a Rep-Free Buying Experience”, 25 June 2025, read on 21 October 2026.
- SaaStr, Jason Lemkin, “We Sent 4,495 AI SDR Emails in 2 Weeks and Got The #1 Response Rate”, 10 July 2025, read on 21 October 2026.
- Google, Gmail Help, “Email sender guidelines”, read on 21 October 2026.
- Hacker News, “You can turn ANY AI SDR into a hacker”, a poster’s claim, 28 July 2025, read on 21 October 2026.
- Deccan Herald (PTI), “’It was my fault’: AI agent leaks biz secrets to Zoho’s Vembu, then apologises”, 28 November 2025, read on 21 October 2026.