type.com team guide
AI for SDRs: a practical guide
Set up one Sales Space that researches, routes, checks, and preps every morning, while reps write and send every message.
What does this guide set up for an SDR team?
This guide sets up one shared Sales Space where AI prepares an SDR team's research, routing, pre-send checks, call prep, and reporting every day, while each rep still writes and sends every message.
It's for SDR managers and reps who book meetings for account executives through cold email, calls, LinkedIn, and inbound. By the end you'll have seven workflows running in type.com, each with copyable instructions, a review rule, and a schedule.
AI does the reading and checking that eats a rep's morning: CRM history, transcripts, public signals, sequence status, and opt-outs. The rep does what earns replies: choosing who is worth a message, writing it, sending it, and running the call. Nothing here sends an email, enrolls a contact, or touches LinkedIn on its own.
The examples use a fictional company, Corvo Analytics, whose four SDRs sell ops reporting to logistics teams on HubSpot, Gong, and Gmail. Swap in your own stack. This guide is part of a series that also covers AI for marketing teams and AI for customer success.
| Workflow | When it runs | Rep time it replaces per week |
|---|---|---|
| Account research packet | On demand, and inside the morning queue | 3 to 5 hours |
| Inbound lead triage and routing | As form fills arrive | 1 to 2 hours for the rep on inbound duty |
| Morning outbound queue with a clear/flag check | Daily, 7:30 AM ET | 2 to 3 hours |
| First-touch brief (rep writes the email) | Inside the morning queue | 1 to 2 hours |
| Call prep and call debrief | Daily at 7:00 AM ET, and after each call | 1 to 3 hours |
| CRM and sequence hygiene | Mondays, 8:00 AM ET | 1 to 2 hours |
| Pipeline-generated review | Fridays, 3:00 PM ET | 1 to 2 hours for the manager |
How do you set up the SDR Space in 30 minutes?
Create one Sales Space, connect the CRM, call recorder, email, and calendar with the narrowest access that works, add five channels, and paste a starter set of team rules into the Space instructions.
A common pattern is one bot per job, plus a coordinator bot to manage the others. Each one then carries its own logins, instructions, and memory, and the manager can't see what they did. In type.com you create one Space for the department instead. Everyone in it shares the same connections, skills, memory, and AI setup. Channels hold the workflows, specialist jobs become shared skills, and routines become automations that post into a channel where the whole team can see the output.
Connect the CRM as an organization connection so research doesn't depend on one rep's login, and keep email and calendars personal so each rep's AI reads only their own mailbox. A Read-only setting tells the agent not to make changes, but it doesn't block changes at the connected service, so also connect with a CRM user whose role can't edit.
- 1
Create the Space from the Sales template
Choose Create, then Space, and pick Sales. Edit the prompt to describe your SDR team, choose Claude (a good default for research and writing), and create it. Approve only the proposed skills, automations, and connections you need now.
- 2
Decide who can see it
Make the Space private if pipeline and call recordings should stay with sales. Invite the SDRs, their manager, and the AEs who take handoffs.
- 3
Add connections at the narrowest level
In Space settings, Connections, add the CRM, the call recorder, and each rep's email and calendar at the access levels below. Assign the CRM only to this Space.
- 4
Create five channels
account-research for packets, inbound-leads for routed form fills, outbound-review for each rep's morning queue and briefs, calls for briefs and debriefs, and pipeline-review for the Monday hygiene list and Friday numbers.
- 5
Paste the Space instructions
Use the starter below. type.com includes the instructions with every message in the Space, so the rules apply to every workflow.
- 6
Map Slack if your team lives there
Connect Slack, invite the app to your SDR channel, then map it to outbound-review in Space settings, Automations. Choose Mentions only so the AI stays quiet until someone asks.
Starter Space instructions (Space settings, Details, Instructions)
You support the SDR team in this Space. Replace everything in [brackets]. What we sell: [one sentence]. Who we sell to: [industries], [employee range], [countries]. Tier A accounts: [definition]. Tier B: [definition]. Everything else is Tier C. Personas: [titles we target, and what each one cares about]. Systems: - HubSpot is the source of truth for owners, lifecycle stages, deals, sequences, and opt-outs. - Gong holds call transcripts. - Each rep's Gmail and calendar are personal. Only read the mailbox of the person who started the task. Ownership: an account belongs to its HubSpot company owner, and an open deal belongs to its deal owner. If someone else owns an account, say who on the first line. Rules for every task in this Space: 1. Never send, schedule, or draft an email, LinkedIn note, or message to a prospect. Reps write and send all outbound. 2. Never enroll or unenroll anyone in a sequence, change an owner, or edit, merge, or delete a record unless the rep approves that exact change in the thread. 3. Every fact about a prospect needs a source and a date. If you can't find something, write "not found". Label guesses "inference". 4. Anyone who unsubscribed, bounced, or asked not to be contacted is off-limits on every channel. 5. Keep pricing exceptions, internal deal notes, and credentials out of anything written for a prospect. Style: counts first, short lines, rep first names, times in America/New_York.
| Connection | Kind and access | Why | Upgrade later? |
|---|---|---|---|
| CRM (HubSpot or Salesforce) | Organization, read-only CRM user | Owners, deals, sequences, opt-outs, activity | Week 3, edit access for approved debrief updates only |
| Call recorder (Gong or similar) | Organization, read-only | Transcripts for briefs, debriefs, and warm history | No |
| Email (Gmail or Outlook) | Personal, one per rep | Who emailed whom in the last 90 days | No. The AI never needs to send |
| Calendar | Personal, one per rep | Today's external meetings for call briefs | No |
| Slack | Workspace, mapped channels only | Lets reps ask from Slack | Add channels as the team adopts it |
| Not connected | Reps look up and message people by hand | No |
How do you build an account research packet with AI?
A research packet pulls the account's history from your CRM and call notes first, adds up to three dated public reasons to reach out now, and says plainly what it couldn't find.
When it runs: on demand, when a rep types /account-research-packet and a company in account-research, and inside the morning queue. Owner: each rep. Inputs: the CRM, call transcripts, and public pages such as the newsroom and careers page.
Research is where AI saves the most time and does the most damage when it guesses. A packet that invents a title or email address is worse than none. So the order matters: ownership first, then your own history, then public signals. A closed-lost reason that has since changed beats any cold trigger.
In the example, Halden Freight lost a deal in March because of a warehouse system migration. On Sep 22 their newsroom said the migration was done, and VP of Operations Dana Ruiz had told Priya in February that reporting was "the next fire". That is a reason to write today, with a person and a quote attached.
Review rule: for the first two weeks, the rep opens every source link before using a packet. After that, spot-check two packets a day. Save the skill under Library, Skills, so every rep runs the same rules and anyone can suggest an edit for the owner to review.
Shortcut for agency and services SDRs: when a prospect has shared real data with you, the New Business Pitch Builder skill in the type.com Skills Library turns that audit into a prioritized gap list with sized opportunities. It won't invent numbers for a prospect who shared nothing.
Skill: /account-research-packet (SKILL.md)
--- name: account-research-packet description: Build a one-screen research packet for one target account before an SDR reaches out. Use when a rep names an account or domain, or when the morning queue needs a packet. --- Input: one account name or website domain. 1. Ownership. In the CRM, find the company, its owner, lifecycle stage, and any open deal with its owner. If someone other than the requesting rep owns it, say who on the first line and stop after the Snapshot. 2. Our history. List closed deals with their closed-lost reason, the last logged activity and its date, and anyone who replied to us in the last 12 months. 3. Call notes. If the call recorder has calls with this account, quote at most two lines that matter, each with the call date. 4. Why now. Up to three reasons to reach out now from the company's own site, press releases, or job posts. Each needs a link and a date in the last 90 days. Put older items under Gaps. 5. Who to contact. Only people in the CRM or named on a public page you link. Show title, the date it was last confirmed, and our history with them. Never guess an email address or profile URL. Mark titles not confirmed in 90 days "verify". 6. Angle. One sentence: why this account, why now, who first. Format: Snapshot (one line), Our history, Why now, Who to contact, Angle, Gaps. If a source fails or finds nothing, say so under Gaps. Do not write outreach copy. Do not change any record.
- Failure: signals are generic ("growing fast"). Fix: require a link and a date, and drop anything without both.
- Failure: a rep researches an account another rep owns. Fix: ownership is step one, and the packet stops there.
How should AI triage and route inbound leads?
Send every form notification to the Space's inbound email address, check ownership and opt-outs before scoring, and post a suggested tier, owner, and response deadline that a person accepts.
When it runs: as leads arrive. Create an inbound email address in the Space's Automations page with inbound-leads as its destination, then point your form notifications at it. A Space doesn't get an address automatically, and the destination is fixed once created. Owner: the SDR on inbound duty that week. Inputs: the form notification and the CRM.
Speed to lead is the metric here, and most of the delay isn't the reply. It's working out whether the lead is a customer, who owns the account, and whether anyone is already talking to them. The skill does those checks first, so the rep on duty sees an owner and a deadline.
Example output for a Tuesday 10:12 AM demo request: "A · High · Owner: Priya · respond by 10:27 AM ET. Operations manager at a 1,200-person third-party logistics company in Ohio. No CRM record. Matched the US East mid-market territory row. No conflicts." A second lead from a customer's domain gets "Customer, not new business · route to the account owner", not a sales sequence.
Review rule: the rep on duty accepts or changes the owner in the CRM themselves, then replies to the lead personally. The AI never assigns, enrolls, or replies.
Shortcut: the Lead Triage Router skill in the type.com Skills Library packages a scored version for higher volumes. It keeps weights in a config file, prints the signals behind every score, and only produces a work list.
Skill: /inbound-triage (SKILL.md)
--- name: inbound-triage description: Triage inbound leads from a form notification. Check ownership and opt-outs, score fit and intent, and suggest an owner and a response deadline. --- For each lead in the incoming message: 1. Pull name, work email, company, title, country, form name, page, and message. 2. Look up the email domain and company in the CRM: customer status, open deals, company owner, active sequences, and opt-out status. 3. Apply hard routes before scoring: - Existing customer: route to the account owner. Label "Customer, not new business". - Open deal: route to the deal owner. - Company already owned: route to that owner. - Unsubscribed or do-not-contact: label "Do not contact" and route to [ops owner]. - Personal email, student, job seeker, vendor pitch, or competitor: label "Review" and say why. Never label a lead "Junk". 4. Fit: A if it matches [industry], [employee range], and [countries]. B if it matches two. C otherwise. Intent: High for demo or pricing requests. Medium for trials and webinars. Low for content downloads. 5. Route new A and B leads with this territory table: [region + segment -> rep, with a backup]. If two rows match, or none, say so instead of choosing. 6. Deadline: High intent between 8 AM and 6 PM ET gets 15 minutes from the form time. Everything else gets end of the next business day. Output one block per lead. First line: "Tier · Intent · Owner · respond by [time]". Then the evidence for each score, then any conflicts. Write "not found" for anything you couldn't confirm. Never invent company size or industry. Never assign owners, enroll anyone in a sequence, or reply to the lead.
- Failure: more than about 15% of leads land in Tier A. Fix: tighten the fit definition until the team can actually work every A the same day.
- Failure: good leads hide in Review. Fix: the manager scans the Review list every Friday.
- Failure: nothing arrives. Fix: send a test form and confirm it lands in inbound-leads. A private editor test doesn't verify the email address.
How do you build a morning outbound queue with a clear/flag check?
A scheduled automation checks every contact on each rep's list against a written rule set before 7:30 AM, marks each one clear, flag, or skip with evidence, and posts the queue to outbound-review.
When it runs: daily at 7:30 AM ET, one automation per rep. Owner: each rep. Inputs: the rep's "this week" account list in the CRM, sequence status, and the rep's last 90 days of email.
The costliest outbound mistakes aren't bad copy. They're emailing a customer mid-renewal, pitching someone an AE is already working, writing to someone who unsubscribed, or burning a mailbox on addresses that bounce. The clear/flag check catches those before anyone writes, and its key design choice is the default: anything the AI can't confirm is a flag, not a pass.
Each rep creates their own copy because scheduled runs use the creator's identity and credentials. Priya's queue reads Priya's mailbox and her account list, and nobody else's. Tuesday's run checked 18 contacts at 6 accounts: 9 clear, 7 flagged, 2 skipped.
Review rule: the rep resolves every flag before 10 AM: keep, drop, or ask the owner. If a flag is wrong twice for the same reason, the rep suggests a skill edit so the whole team gets the fix.
- 1
Create the automation
Each rep chooses Create, then Automation, and picks the Sales Space. It opens in the Library, paused.
- 2
Paste the instructions
"Build my outbound queue for today from my CRM list '[name] this week'. Use /account-research-packet for each account, /clear-flag-check on every contact, and /first-touch-brief for each CLEAR contact. Post one message: counts, flags, then briefs. If today is Saturday or Sunday, post nothing. Never send email, enroll anyone in a sequence, edit the CRM, or do anything on LinkedIn."
- 3
Set the destination and schedule
Destination: outbound-review. Schedule: daily at 7:30 AM, America/New_York. Model: Agent default.
- 4
Save, then Test
Test runs privately, as you. Check five contacts against the CRM by hand, including every flag.
- 5
Enable it
Select Enable. Three failed runs in a row pause the schedule. Check the run history, fix the cause, test again, then enable it.
Skill: /clear-flag-check (SKILL.md)
--- name: clear-flag-check description: Check every contact on a rep's outbound list before any brief is written. Mark each one CLEAR, FLAG, or SKIP with evidence. --- Check each contact against the CRM (owners, deals, lifecycle, sequences, opt-outs) and the rep's last 90 days of email. SKIP when the contact is in an active sequence, or anyone here emailed them in the last 30 days. Name the owner. CLEAR only when every rule is confirmed: 1. The company is not a customer, open trial, partner, or do-not-contact account. 2. The contact has not unsubscribed, bounced, or asked not to be contacted. 3. No open deal, or the deal owner approved outreach in writing in the CRM. 4. Work email on the company domain, and a title confirmed in the last 90 days. 5. The account is Tier A or B. 6. The research packet has a linked signal from the last 90 days. 7. Our policy allows cold email to the contact's country. 8. The email address passed verification in the last 30 days. FLAG everything else. Show the first rule that failed and the evidence. If a source could not be read, FLAG every contact that depends on it and name the source. If more than [40] contacts are CLEAR, keep the top [40] by tier and signal date and list the rest under "Tomorrow". Output: counts first, then FLAG with rule and evidence, then CLEAR, then SKIP with owners. End with the sources you read. Never send, enroll, or edit anything.
How should AI help with first-touch emails without writing them?
For each clear contact, AI writes a short brief with the signal, why it matters to this person, one approved proof point, and an opening question, and the rep writes and sends the email themselves.
When it runs: inside the morning queue, under the flags. Owner: the rep. Inputs: the research packet, the clear/flag result, and your approved list of customer proof points.
Prospects spot generated outreach in a sentence, and a team sending it at volume trains its market to ignore the next email. So the brief contains no email: no subject line, no opener, no draft to tidy up. It gives the rep what a good manager would say first: the signal worth mentioning, why it matters to this person's job, a real customer result, and a question that starts a conversation.
Priya's email to Dana was four sentences, used the February call, and offered a way out. She sent it from her own Gmail at 9:12 AM. The brief made it fast. The words are hers.
Review rule: the manager reads five sent emails per rep each week alongside their briefs and coaches in the thread, not in a private DM. For more on placing review points, see how to design AI agent approval workflows.
Skill: /first-touch-brief (SKILL.md)
--- name: first-touch-brief description: For each CLEAR contact, give the rep the facts they need to write a first touch. Never write the message. --- For each CLEAR contact, write a brief of no more than 90 words: - Signal: the strongest reason to reach out now, with link and date. Warm history (a past call, a reply, a closed-lost reason that has changed) beats any public signal. - Why it matters to this person: one sentence about their role, not their company. - Proof point: one result from our approved list [link the list]. If none fits, write "none fits". Never invent or round up a result. - Opening question: one question the rep could ask. Not a pitch. - Avoid: inferences, internal notes, pricing, and anything personal. - Channel: email, call, or a LinkedIn request sent by hand, with a one-line reason. - Pre-send check: copy the CLEAR evidence from /clear-flag-check. Do not write a subject line, an email, a LinkedIn note, or any sentence meant to be sent. If a rep asks for a draft, say that reps write outbound here, and offer two more facts instead. Post the briefs under the queue in outbound-review.
- Failure: every brief uses the same proof point. Fix: keep an approved list of at least five results, tagged by industry and persona.
- Failure: briefs lean on inferences. Fix: anything labeled "inference" in the packet goes under Avoid.
- Failure: reps paste the brief into the email. Fix: that's a coaching conversation, and the five-email review catches it early.
How do you prep and debrief SDR calls with AI?
A 7:00 AM brief covers every external meeting on the rep's calendar, and a debrief after each call turns the transcript into quotes, a proposed CRM change table, and recap points the rep approves.
When they run: the brief is an automation, daily at 7:00 AM ET into calls, created by each rep so it reads their own calendar. For same-day bookings, the rep runs /call-prep-brief by hand. The rep runs /call-debrief in calls after each meeting. Inputs: the calendar, the CRM, and the call recorder or a pasted transcript.
A qualification call either earns an AE meeting or it doesn't, and the handoff is where deals leak. The brief means the rep knows who's new on the invite and what was promised. The debrief means the AE walks into Thursday knowing what Nadia actually said, with timestamps.
The debrief also fixes a common sequence error: a prospect books a meeting and still gets step four. Unenrolling is one of the proposed rows.
Review rule: nothing reaches the CRM until the rep replies "approve" or names the rows to change. Budget and amount stay blank when they weren't discussed. The recap email is the rep's to write.
Shortcut: the Pre-Call Briefing skill in the type.com Skills Library builds a fuller one-page brief from calendar, CRM, email, Slack, and call notes and is read-only. The Call Debrief skill confirms CRM changes before writing by default. Set its follow-up option to draft or none, never send.
Two skills: /call-prep-brief and /call-debrief
--- /call-prep-brief --- For each external meeting on my calendar today (skip internal-only invites): 1. Who is coming: name, title, and whether the CRM has them. Flag anyone new. 2. Our history: deal stage and owner, the last three logged touches, and what they replied to. 3. What they told us: up to two quotes from past calls, with dates. 4. Open items: questions we didn't answer and anything we promised. 5. Plan: three qualifying questions on [need, timeline, decision process], and our criteria for booking an AE meeting. Under 200 words per meeting. Write "not found" rather than guess. --- /call-debrief --- Using the call's transcript from the recorder, or one pasted in the thread: 1. Outcome in one line: qualified, not qualified, or follow-up needed, with the next meeting if one was booked. 2. Quotes with timestamps on pain, current process, timeline, budget, and decision process. If a topic wasn't discussed, say so and leave the field blank. 3. Proposed CRM changes as a table: field, current value, proposed value. If a meeting was booked, include unenrolling the contact from any active sequence. 4. Points for the rep's recap email. Do not write the email. Write nothing to the CRM until the rep replies "approve" or names the rows to apply.
- Failure: the debrief fills in a budget nobody said. Fix: quotes or blank, never a summary of a number.
- Failure: a meeting was booked and the sequence kept sending. Fix: the unenroll row in every booked-meeting debrief, and the Monday hygiene check as a backstop.
How do you keep the CRM and sequences clean with AI?
Every Monday at 8:00 AM ET, a read-only check lists broken contacts, sequences, and early deals by owner, worst first, and each rep fixes their own records.
When it runs: weekly, Mondays at 8:00 AM ET into pipeline-review. Owner: each rep fixes their list. The manager checks that the critical items are gone by noon. Inputs: the CRM, including sequence enrollment.
SDR data decays quietly, and most of it isn't cosmetic. A contact who unsubscribed but is still enrolled is a compliance problem. A prospect in two sequences gets two emails a day. A cadence that keeps running after a booked meeting makes the rep look careless. That's the critical tier.
Example: Jon's Monday list has 2 critical items (a contact who bounced Sep 30 but is still enrolled, and one who booked a meeting Oct 1 but is still in step three) and 6 for this week.
Review rule: the AI changes nothing. Reps fix their own records, because a person can tell a slow deal from a neglected one.
Shortcut: the Pipeline Hygiene Audit skill in the type.com Skills Library audits a deal export for stale deals, past close dates, missing amounts or owners, skipped stages, and duplicates, with a ranked fix list per owner. It never writes back. Use it alongside this check if SDRs own early-stage deals.
Skill: /crm-hygiene-check (SKILL.md)
--- name: crm-hygiene-check description: Weekly check of SDR-owned contacts, sequences, and early-stage deals. Lists problems by owner. Changes nothing. --- Check records owned by [SDR names]. List each problem once, under its owner, most serious first. Critical, fix today: 1. Unsubscribed, bounced, or do-not-contact contacts still in an active sequence. 2. Contacts with a meeting booked in the last 14 days who are still in a sequence. 3. Contacts in two or more active sequences. Fix this week: 4. Contacts in a sequence meant for another persona or segment. Use this map: [sequence -> persona]. 5. Sequences that ended more than 7 days ago with no next task. 6. Possible duplicate contacts or companies (same email, or same domain on two company records). 7. SDR-sourced deals in the first two stages with no activity in 21 days, or no next step. 8. Leads assigned more than 3 business days ago with no logged touch. For each item: record link, rule number, evidence, and suggested fix. At the top, show counts per owner. If a check can't run because a field is missing, name the field. Change nothing.
- Failure: the list is 200 lines long. Fix: cap each owner at the ten worst items and show the total.
- Failure: critical items reappear every week. Fix: that's a process gap, such as an unsubscribe not syncing to the sequencer. Escalate it to sales ops.
How do you review SDR pipeline generated each week with AI?
Every Friday at 3:00 PM ET, a read-only review totals the week's first touches, replies, meetings, accepted SQOs, and pipeline by rep, compares them with the last four weeks, and names data gaps instead of guessing.
When it runs: weekly, Fridays at 3:00 PM ET into pipeline-review. Owner: the SDR manager, who creates the automation so it runs with the manager's CRM access. Inputs: the CRM and the call recorder.
The number that matters is pipeline the AEs accepted, not emails sent. The review leads with accepted SQOs, then shows the activity behind them. In the example, 380 first touches produced 19 positive replies, 14 meetings booked, 11 held, and 7 SQOs worth $186K. Four of the seven came from warm or closed-lost accounts, a reason to put more of next week's queue there.
Small numbers lie. Three meetings instead of four isn't a bad week, so the skill only calls a trend at counts of 10 or more moving over 20%, and it never ranks reps.
Review rule: the manager reads it before Monday 1:1s, chases the data gaps, and adds the spam complaint rate from Google Postmaster Tools by hand. For the general pattern, see how to automate a weekly team report.
Skill: /weekly-pipeline-review (SKILL.md)
--- name: weekly-pipeline-review description: Friday review of SDR pipeline generated this week, by rep, with data gaps named. Read-only. --- Cover Monday to Friday of this week in America/New_York. The CRM is the source of truth. 1. For each rep and the team: first touches (emails and calls logged), positive replies, meetings booked, meetings held, SQOs accepted by an AE, and the pipeline amount on those SQOs. 2. Compare with the average of the previous 4 weeks. Only call a change a trend if the count is at least 10 and it moved more than 20%. Otherwise write "too small to call". 3. Source of each SQO: warm or closed-lost, inbound, or cold. 4. Bounce rate from sequence email stats. Flag anything above 2%. 5. No-shows, with the days between booking and meeting. 6. Data gaps: meetings without an outcome, deals without an amount or source. Name the owner and ask them in this thread. Lead with pipeline accepted by AEs. Do not rank reps or comment on effort. End with one question per rep the manager could ask in Monday's 1:1. Change nothing.
- Failure: "meetings booked" counts meetings the AE later rejected. Fix: report booked, held, and accepted separately, and celebrate accepted.
- Failure: the numbers don't match the CRM dashboard. Fix: write your definitions (what counts as a first touch, an SQO, a held meeting) into the skill.
What does the weekly operating rhythm look like?
The week runs on two daily automations, one event trigger, two weekly reviews, and a clear human action for each.
Pin this table in the Space so new reps know what arrives when and what to do with it.
| When | Workflow | Channel | Owner | What a person does |
|---|---|---|---|---|
| Daily, 7:00 AM | Call briefs | calls | Each rep | Reads briefs before the first call |
| Daily, 7:30 AM | Morning outbound queue | outbound-review | Each rep | Resolves flags, then writes and sends first touches by 10 AM |
| As form fills arrive | Inbound triage | inbound-leads | Rep on inbound duty | Accepts or changes the owner and responds within the deadline |
| After each call | Call debrief | calls | Rep who ran the call | Approves CRM rows and writes the recap email |
| Mondays, 8:00 AM | CRM and sequence hygiene | pipeline-review | Each rep; manager checks | Fixes critical items by noon, the rest by Friday |
| Fridays, 3:00 PM | Pipeline-generated review | pipeline-review | SDR manager | Chases data gaps and takes one question per rep into 1:1s |
| On demand | Account research packet | account-research | Any rep | Checks sources before using a packet |
What guardrails should an AI SDR workflow have?
AI may read and prepare, a rep approves any CRM change, and AI never sends outbound, enrolls contacts, acts on LinkedIn, or overrides an owner.
Write these rules into the instructions and skills, then enforce them with permissions in the tools themselves. Instructions alone are not a security control.
Who runs what matters. In type.com, AI acts with the effective access of the person who started the work. A Test run uses the access of the person pressing Test, while scheduled runs use the creator's identity and credentials. When a rep leaves, archive their automations and remove their personal connections. Memory carries context forward but isn't a source of truth, which is why every skill reads the CRM directly. Keep credentials out of prompts and skills, and handle call transcripts and contact data under your privacy policy and retention rules.
The SDR-specific rules are where most teams get hurt, so here is how an experienced SDR leader would set them:
| Tool | AI may read | AI may prepare | Only with a rep's approval | Never |
|---|---|---|---|---|
| CRM | Accounts, contacts, deals, activity, sequences, opt-outs | Packets, routing suggestions, change tables, hygiene lists | Apply approved debrief rows (from week 3) | Change owners, merge, delete, bulk edit, enroll |
| The rep's own last 90 days | Briefs and recap points | Nothing | Send, schedule, or draft a message | |
| Calendar | The rep's own external meetings | Call briefs | Nothing | Accept, move, or create events |
| Call recorder | The team's call transcripts | Quotes and debriefs | Nothing | Share recordings outside the Space |
| Sequencer | Enrollment and email stats | Hygiene and deliverability flags | Unenroll after a booked meeting | Enroll anyone or change cadence steps |
| Nothing (not connected) | A channel suggestion in the brief | Nothing | Requests, messages, or scraping | |
| Slack | Mapped channels | Answers when mentioned | Nothing | Post in channels shared with prospects |
- Deliverability: authenticate sending domains with SPF, DKIM, and DMARC, warm new mailboxes, verify addresses first (rule 8), cap new contacts per mailbox per day, keep bounces under 2%, and keep spam complaints under 0.1% in Google Postmaster Tools, well clear of Gmail's 0.3% bulk-sender limit.
- Opt-outs: an unsubscribe applies to the person on every channel. U.S. law allows up to 10 business days to honor one; set a same-day standard. Rule 2 and the critical hygiene tier enforce it.
- LinkedIn: it limits weekly invitations, and its User Agreement prohibits bots and automated activity. Reps connect and message by hand. An unverified profile URL is a flag, not a guess.
- Territory and ownership: the account and deal owners win. The AI names conflicts and never resolves them. The SDR manager settles disputes and updates the territory table.
- Consent regions: rule 7 flags countries where your policy requires consent, such as many in the EU or Canada. Legal sets that policy.
How do you roll this out over 30 days?
Start with one rep and read-only research in week one, add the morning queue in week two, add calls and the whole team in week three, and turn on the weekly reviews and any CRM writes in week four.
Start with one rep, ideally a skeptic, whose flags will tighten the rules fastest. Widen only when the previous step has proved safe.
| Week | Turn on | Review | Widen when |
|---|---|---|---|
| 1 | Space, read-only connections, instructions, /account-research-packet and /inbound-triage on demand, inbound email address | Every packet against its sources. Time 10 packets against manual research | Spot checks find no invented contacts or titles |
| 2 | Morning queue automation for the pilot rep, with /clear-flag-check and /first-touch-brief | Every flag, and every CLEAR contact for 3 days. Five sent emails | Zero opted-out or owned contacts marked CLEAR for 5 straight days |
| 3 | Morning queue for every rep (each creates their own), call briefs, /call-debrief | Every debrief change table before approval | Debrief tables are right without edits on most calls |
| 4 | Monday hygiene check, Friday pipeline review, CRM edit access for approved debrief rows | Compare week 4 with the baseline below | Map more Slack channels; keep sending human |
How do you measure whether AI is working for your SDR team?
Measure research time, speed to lead, pre-send misses, positive reply rate, meetings held, accepted pipeline, and deliverability against a baseline you take in the first week.
Take the baseline before enabling anything: four weeks of CRM data, plus two reps timing ten accounts of manual research. The 30-day goal is more selling time at the same or better quality, not more emails.
When a workflow drifts, fix the skill rather than the thread, so every rep gets the correction the next morning. Read how shared AI memory works for teams, and for a full list of skills, see the best AI skills for B2B teams.
| Metric | How to get the baseline | Good at day 30 |
|---|---|---|
| Research time per account | Two reps time ten accounts by hand | Under five minutes to review a packet |
| Speed to lead (high intent, business hours) | Median minutes from form time to first human reply, last 4 weeks | Under 15 minutes |
| Pre-send misses | Count of opted-out, owned, or customer contacts emailed last quarter | Zero |
| Positive reply rate | Positive replies divided by first touches, last 4 weeks | Flat or up. A drop means quality slipped |
| Meetings held and SQOs accepted | Held and accepted counts per rep, last 4 weeks | Up, with no rise in AE rejections |
| Bounce and spam complaint rates | Sequencer stats and Google Postmaster Tools | Bounce under 2%, complaints under 0.1% |
| Packet accuracy | Share of spot-checked claims that held up | 95% or better, with misses fixed in the skill |
Frequently asked questions
Will AI replace SDRs?
No. AI takes over the preparation: research, routing, pre-send checks, call prep, CRM cleanup, and reporting. The SDR still chooses who deserves a message, writes and sends it, runs the call, and handles the reply.
Should AI send cold emails or LinkedIn connection requests automatically?
No. Mass AI-sent outbound reads as generic, raises spam complaints, and can burn a sending domain. LinkedIn's User Agreement prohibits bots and automated activity, so automated requests also risk the rep's account. Let AI prepare the facts, and have the rep write and send.
How do I set up AI for an SDR team?
Create one Sales Space, connect the CRM read-only as an organization connection and each rep's email and calendar as personal connections, and add five channels. Write the team's rules into the Space instructions, add a skill per workflow, and enable automations one at a time after a private test.
What CRM access should an AI agent have?
Read-only to start. In type.com, a Read-only setting tells the agent not to make changes but does not block them at the connected service, so connect with a CRM user whose role cannot edit. Add edit access later only for approved debrief updates.
How do you protect email deliverability when using AI for outbound?
Keep sending human: authenticate domains with SPF, DKIM, and DMARC, verify addresses before they reach a rep, cap new contacts per mailbox per day, keep bounces under 2% and spam complaints under 0.1%, and honor opt-outs the same day. AI helps by checking these rules, not by sending.
Is it safe to put prospect data into an AI workspace?
It can be if access stays narrow. Assign the CRM connection only to the Sales Space, make the Space private if pipeline data is sensitive, keep credentials out of prompts, and follow your privacy and consent rules for transcripts and contact data.