Build it as a chain of nine stages: ICP and list, account research, signal triggers, first-touch drafting, human review, sending within deliverability limits, reply handling, CRM logging and a weekly review of reply quality. AI does the research and drafting. A person approves every first touch before it goes out.

Key points

• An AI SDR workflow is a pipeline, not one tool, and it works best inside an AI operating system for a B2B team.

• AI is strongest at research and first drafts. People are still better at judging fit, tone and timing.

• Human-reviewed drafts protect your domain, your brand and your reply rate while the system learns.

• Gmail and Yahoo publish hard rules for bulk senders. Authentication, a low spam rate and easy unsubscribes are table stakes.

• A weekly review of reply quality, not open rates, tells you whether the system books meetings.

What is an AI SDR workflow?

An AI SDR workflow is a set of connected steps where AI finds, researches and drafts outreach to target accounts, and people approve and own the conversations. It replaces the manual research and writing an SDR does, not the judgment.

Vendors package this in different ways:

• All-in-one AI SDR products. AiSDR says it "finds in-market buyers using real-time signals" and then hands off, "having humans step in to close and expand accounts."

• CRM-native agents. HubSpot's prospecting agent "researches accounts, finds the right contacts, and drafts personalized outreach."

• Build-your-own stacks. A data tool, a research agent, a sending tool and your CRM, wired together.

The stages below apply to all three. The difference is how much of the wiring you own.

What are the stages of an AI outbound system?

There are nine stages, and each one should produce something a person can check. If a stage has no visible output, you cannot debug it.

How do you define the ICP and build the list?

Write the ICP down before you pull a single contact. A model will happily research and draft for the wrong accounts at scale.

A usable ICP for outbound has four parts:

1. Firmographics. Industry, headcount band, region and tech used.

2. Buyer role. The title who owns the problem, plus one backup title.

3. The problem. One sentence on what hurts, in the buyer's words.

4. Disqualifiers. Who you will not contact, such as current customers or open deals.

Then build the list. LeadFuze says it "builds your lists automatically" and lists 269M+ profiles in its database. Clay pulls from many data providers into one table. Either way, export a sample of 20 rows and read them before you enrich the full list.

Verify every email before it enters a sequence. Emailable states that "no more than 1% of emails verified as Deliverable will bounce upon sending," which is the kind of guarantee you want in writing from any verifier. Bounces count against you with mailbox providers, so clean lists are a deliverability step, not admin.

How does AI research accounts and spot signals?

AI reads public information about each account and writes a short brief, then flags the events that make outreach timely. This is where AI saves the most hours, because manual research is slow and repetitive.

Account research

Clay describes its agent as one you can "point Claygent at any manual GTM task like company research, lead qualification, signal detection." A research agent in Claude or Relevance AI can do the same job if you give it a fixed output format.

Keep the brief short and structured:

• What they sell and to whom. Two sentences.

• One recent event. With the source link.

• Likely pain. Tied to your ICP problem statement.

• Confidence. High, medium or low, with the reason.

Signal triggers

A signal is a reason to write now instead of next quarter. Common ones are a new leader in the buyer role, a funding round, a job post for the role your product supports, or a visit to your pricing page.

Start with three signals, not ten. Track replies by signal for a month, then keep the ones that produce conversations. Keep only what earns replies.

Should AI write and send first touches on its own?

No. Let AI write the first draft and let a person approve it before it sends. Fully autonomous sending saves a few minutes per email and risks your domain, your brand and every future reply.

Four reasons to review:

• Errors compound at volume. A wrong fact in one prompt becomes a wrong fact in 500 emails before anyone sees it.

• First impressions are permanent. A prospect who gets a clumsy AI email rarely opens the next one.

• Spam complaints are shared. A spam rate problem hurts every email from that domain, including sales replies and invoices.

• Review produces training data. Every edit a person makes shows you what the prompt gets wrong.

Vendors build review in for these reasons. HubSpot's FAQ says that "every time you enroll contacts, you have the option to review and edit AI-drafted emails before they are sent." Relevance AI lists the ability to "maintain control with human oversight and approval workflows" as a core feature.

Review stays fast when each draft shows the signal, the source and the claim to check beside it. Send drafts to a Slack channel or a CRM task queue with approve, edit and skip buttons.

When can you loosen review? Only after weeks of drafts that need no edits, and only for follow-ups to people who already replied. The L1 to L4 AI automation maturity ladder explains how to move a workflow up one level at a time.

A simple way to set this up is three doors: the agent writes when every condition passes, asks a human when the signal is ambiguous, and stops when a rule says no.

What deliverability rules does an AI outbound system have to follow?

Gmail and Yahoo both publish rules for bulk senders, and outbound tools that send at volume fall under them. Meet the authentication, spam rate and unsubscribe rules before you send your first sequence.

Google's email sender guidelines apply extra requirements to anyone sending 5,000 or more messages a day to Gmail accounts. They include:

• Authentication. "Set up SPF and DKIM email authentication for your domain," plus DMARC for your sending domain.

• Alignment. "The domain in the sender's From: header must be aligned with either the SPF domain or the DKIM domain."

• Spam rate. "Keep spam rates reported in Postmaster Tools below 0.30%." Google also advises staying below 0.10%.

• Unsubscribe. "Marketing messages and subscribed messages must support one-click unsubscribe," with a visible link in the body too.

Yahoo's sender best practices set similar bulk sender rules:

• Authentication. "Implement both SPF & DKIM" and "publish a valid DMARC policy with at least p=none."

• Spam rate. "Keep your spam rate below 0.3%."

• Unsubscribe. "Honor unsubscribes within 2 days."

Deliverability checklist

1. Use a separate sending domain so outbound problems do not touch your main domain.

2. Set up SPF, DKIM and DMARC and confirm alignment before any sequence goes live.

3. Verify every address and pause if bounces climb.

4. Cap daily volume per inbox and raise it slowly on new domains.

5. Add a working unsubscribe and process opt outs within two days.

6. Watch Google Postmaster Tools weekly and treat 0.10% as your ceiling, not 0.30%.

Deliverability limits feed the middle door of the three-door model above, where the agent asks a person instead of sending. Here is an illustrative threshold map for that door.

How should replies be handled and logged?

AI should sort replies and draft responses, and a person should send every response to a real reply. Every touch, signal and outcome should land in the CRM the same way each time.

Reply handling

Sort replies into five buckets: interested, not now, wrong person, unsubscribe and out of office. AI is good at this sort. Unsubscribes should be processed automatically and immediately, because that rule is not optional.

For interested and wrong-person replies, AI drafts the response and a person sends it. These replies are where meetings come from, so they get the most human attention.

CRM logging

Log five fields for every contact in a sequence:

• Signal used. Which trigger started the outreach.

• Draft version. Which prompt or template produced it.

• Edited by reviewer. Yes or no.

• Reply type. One of the five buckets.

• Outcome. Meeting booked, nurture, or closed out.

HubSpot and Salesforce both handle this with custom properties. A connector layer such as Composio keeps the agent's CRM permissions scoped instead of sharing one admin key. The guide to rolling out Claude to a team securely covers how to set those access rules.

How do you run a weekly eval of reply quality?

Once a week, read the replies, not the dashboards. Count meetings and positive replies by signal and by draft version, then change one thing for the next week.

A simple weekly eval:

1. Pull last week's replies grouped by bucket.

2. Read every interested and wrong-person reply. Note what the prospect responded to.

3. Read ten drafts the reviewer edited. List the most common fix.

4. Compare signals. Which one produced the most positive replies per hundred sends?

5. Update one input. The ICP, a signal, or the drafting prompt. Never all three at once.

Open rates are a weak measure now, because privacy features on many mail apps distort them. Replies and meetings are what the system exists to produce.

If reviewers keep cutting the same opening line, fix the prompt. If they keep skipping a segment, fix the ICP.

What are the common mistakes when building an AI SDR workflow?

The biggest mistake is turning on autonomous sending before anyone has read the drafts. Most of the others come from skipping setup work that feels slow.

1. No written ICP. The model researches and drafts for anyone the list returns.

2. Sending from the main domain. One bad week can hurt every email the company sends.

3. Skipping verification. Bounces push you toward the spam folder.

4. Too many signals at once. You cannot tell which one works.

5. Long, generic drafts. A short email tied to one real event beats three paragraphs about your product.

6. No unsubscribe. It breaks the published bulk sender rules from Gmail and Yahoo.

7. Logging nothing. Without signal and draft version in the CRM, the weekly eval has nothing to compare.

What does an AI outbound system cost in time saved?

That depends on how many hours your team spends on research and first drafts today. Measure that before you buy anything.

Here is the worked example from my AI Workflow Install page: a 12-person team where each person spends 8 hours a week on manual work, at a $60 loaded hourly rate. That is $299,520 a year, and in the example one workflow accounts for half of it, $149,760.

If outbound is not where your hours go, start elsewhere. The manual-work leak audit guide shows how to time each workflow and rank them.

Run the math for your own team: people, times weekly manual hours, times the loaded hourly rate, times 52.

FAQ

What is an AI SDR?
An AI SDR is software that finds prospects, researches them and drafts outreach the way a sales development rep would. Most products still hand the live conversation to a person.

Can an AI SDR book meetings on its own?
Some products send and book without review. I recommend a person approves first touches and replies, because errors repeat at volume.

Which tools do I need to start?
A list source such as LeadFuze or Clay, a research and drafting model such as Claude, an email verifier, a sending tool and your CRM.

Do the Gmail and Yahoo rules apply to cold outreach?
Google's extra requirements apply to anyone sending 5,000 or more messages a day to Gmail accounts, and Yahoo publishes similar bulk sender rules. Following them at any volume protects your domain.

What spam rate should I aim for?
Google requires below 0.30% in Postmaster Tools and advises staying below 0.10%. Treat 0.10% as your limit.

How long does it take to build an AI SDR workflow?
It depends on how clean your ICP, list and CRM fields are. Start with one segment and three signals, and expand once the weekly eval shows replies.

Find out whether outbound is your first workflow to automate

If you want help deciding whether an AI outbound system is the right first build, book a free 30-minute Leak Call. We estimate what manual work in one area costs per year, and you get a costed workflow list within 48 hours. You can also see the tools I build on in my stack page.

About the author: Kevin Payne is an AI engineer and GTM operator who installs AI workflows for B2B teams. He has 10+ years in B2B SaaS demand generation and RevOps across Techstars, a16z and YC portfolio companies, and his writing has appeared in HubSpot, GetResponse, Outbrain and more.