Agents can replace part of the SDR job today. They handle account research, enrichment, email verification, first drafts and reply triage well, once a person writes the rules. Judgment calls, live conversations and relationships still need a person. Few companies have made the switch: in a 2025 survey of 351 B2B companies, 1% named "AI SDRs" in their team model.
Which SDR tasks do agents do well today?
The research and admin tasks that come before the first call. We automate each of these in our own outbound, and each one still has a point where a person checks the work.
| Task | What the agent does | Where a person checks | Our example |
|---|---|---|---|
| Account research | Reads public feeds, job posts, funding news and on-chain data, and logs each event | A person writes the filter that decides which events count | 6,069 signals logged in ten weeks (our data, October 2026) |
| Enrichment | Matches the event to one company, then finds the right contact and title | A person writes the rules for namesakes and stale titles | Exact-name rule added after a namesake error |
| Verification | Checks every address before sending | A person decides how to treat catch-all addresses | 10% of a re-verified list came back catch-all (our data, October 2026) |
| First drafts | Writes an email that cites the signal | A person approves the template, the offer and the claims | Opening lines built only from the signal's own fields; variants ranked by positive replies per 100 sends |
| Reply triage | Labels each reply and alerts a person | A person answers every reply that needs one | Six labels, four of which alert a person |
The gain is coverage. The monitor reads every feed on every run, and the verification step runs on every send, busy days included.
Where do agents still make mistakes?
In the places where the data is ambiguous, and each mistake type needed a rule written by a person.
Signal quality is the first one. Our catalyst monitor logged 6,069 buying signals between 26 July and 7 October 2026. Of those, 5,857 (96.5%) were too imprecise to email about: noise, the wrong company in the headline, TVL drops, or job posts outside sales (our data, October 2026). Of the rest, 91 named a company that our contact database could not match to exactly one firm of 11 to 1,000 staff. Without a filter, an agent writes a confident email about every one of them.
Namesakes are the second. A hiring signal for one company matched a different company with a similar name in our data tool. A second signal did the same with another pair of firms. The fix was a rule: take the top-ranked exact-name match across all company sizes, then apply the size filter.
Titles are the third. We found a contact record whose job title named a different company from the one on the record. We now drop any title that names another firm.
Once a rule exists, an agent applies it to every record. A person had to spot each of these errors first, by reading a sample of the agent's output.
Which parts of the job still need a person?
Three parts, and our reply classifier shows where they sit. It sorts every reply into one of six labels: positive, referral, question, negative, out of office or unclear. Four of those labels go straight to a person, who answers from the same mailbox that sent the email.
Start with judgment calls. Whether a signal justifies an email to this buyer this week depends on context no filter captures, such as a recent conversation, a partner conflict or a deal already in progress. Unclear replies land here too.
Live conversations need a person too. A buyer's question often has several parts and a hidden objection. The classifier can label it "question". A person has to answer it, and a person runs every call that follows.
Relationships are the third part. A referral to a colleague, a follow-up months later, an introduction at a conference: each depends on trust between two people. The classifier labels a referral. A person makes the introduction land.
How many companies have replaced SDRs with agents so far?
Very few, on the best survey data we found. The Bridge Group's 2025 report surveyed 351 B2B companies. "AI SDRs" appeared for the first time as a distinct category in its question on SDR-to-AE ratios, at 1% of respondents (The Bridge Group, February 2025). The most common model was one SDR for every two AEs, at 31%.
The same report sets the benchmark an agent has to beat. Median SDR on-target pay was $80K, the average ramp was 3.0 months, and the median monthly quota was 10 first-stage meetings held. Measure any agent against meetings held at that bar, and ignore activity counts such as emails sent.
The survey data is from early 2025, and adoption has likely changed since.
How should you split the work between agents and people?
Give agents the steps a rule can decide, and keep people on the steps that need a reply or a relationship. This is the order we use:
- Write the buyer definition and the signal rules in plain words.
- Let agents collect signals, match each one to a single company and find the contact.
- Read a sample of matches by hand every week, and write a rule for each new error type.
- Verify every address in the week before sending.
- Let agents draft the emails from templates a person has approved.
- Let a classifier label replies, and route positive, referral, question and unclear replies to a person the same day.
- Keep a person on every call and every introduction.
With this split, the SDR spends less time building lists and more time in conversations. For the full system behind steps 1 to 6, read what GTM engineering is. If you are weighing a team against a partner, read outbound agency vs in-house SDR team.
To see which of your own outbound tasks an agent can take over, start with The Diagnostic, which maps your workflows and prices the agent stack in your numbers for $2,500.