Most guides on prospecting with AI describe a stack: one tool to find contacts, another to enrich them, a third to score fit, a fourth to write and send messages. You still have to connect them, keep the connections working, and carry the context from one tool to the next yourself. That is not a small detail. It is the actual work, and it is the reason so many “AI prospecting” setups get built once and abandoned within a month.
I walk the process a different way: as one continuous conversation with an agent that carries the context for you, from the first brief to the first reply you have to answer yourself. Ask most people how AI prospecting actually works and they describe a list of use cases; here it is a single path, in order, at a level of detail you can execute today rather than a set of concepts to assemble on your own. I will point to deeper guides on specific steps (defining an ICP, calibrating a lead score, enriching a profile) where the detail matters more than a single walkthrough can cover, and mark clearly where each detour leads back into this path.
Two things are true about how to use AI for sales prospecting at every step below: nothing sends without a channel connected (LinkedIn, email, or both), and nothing in the workflow requires you to touch a spreadsheet or a separate sequencer at any point. If a channel is not connected yet, I still define strategy, find and score prospects, and draft messages for you to send manually, so the workflow degrades gracefully instead of stalling.
Step 1: Turn Your Offer and Target Into a Brief I Can Act On
Everything downstream depends on this step, and it is where most first cycles go wrong before they even start. A vague brief (“B2B companies that might need consulting”) gives me nothing to score against, so I either surface a broad, low-relevance list or ask for more detail before doing anything useful. A specific one gives me a real filter: target roles, industries, company size, and the actual problem your offer solves for that target.
If you already know your ideal customer profile, this step is quick. You describe the offer, the target roles, and the commercial objective (starting a conversation, booking a meeting, or driving a specific signup), and I turn that into a working persona with its own targeting criteria and messaging preferences: tone, formality, message length per channel, and any proof or content worth referencing when it fits naturally. If you do not yet know who your best-fit customer actually is, that is a separate problem worth solving properly before you prospect at volume, not something to guess at inside a single brief. I cover the full method, including how to pull it from your closed-won accounts instead of guessing, in how to define your ICP with AI.
One offer can support several personas if you sell more than one thing to more than one audience, and you do not need to get any of it perfect on the first pass. You can edit and refine a persona as replies start telling you what actually resonates, which matters more than getting the wording right before you have sent a single message.
This is also the moment to set the boundaries I should never cross regardless of automation level: a blacklist of companies that should never be proposed or contacted (a current client, a competitor, a company that already said no through another channel), and, if relevant, a target-company list that gets priority during discovery. Both apply the same way whether you are reviewing every action manually or running Auto later.
Step 2: Let Me Find and Score Prospects
Once a persona exists, I search for matching profiles one at a time and score each one from 1 to 5 stars against both the persona and the offer, showing the reasoning behind the score alongside it rather than a bare number. A profile that clears the threshold gets held for your review before I do anything else with it; a profile that does not still stays visible, and you can rescue it manually if the reasoning does not convince you.
The mechanism behind that score (how the criteria are weighted, how to set a threshold, and what to do once reply rates start suggesting the threshold is wrong) is its own topic, deep enough that folding it into this walkthrough would bury the rest of the process. I go through it in how AI lead scoring actually works. For this walkthrough, the default recommended threshold is a minimum of 3 out of 5 stars: strict enough to filter out clear non-fits, loose enough not to starve your own pipeline before you have any data on what actually converts. For the full picture of this discovery and scoring phase on its own, see how to find and qualify B2B leads with an AI agent.
Step 3: Enrich the Ones Worth Contacting
I enrich a validated prospect automatically: current role, company profile, responsibilities, expertise, interests, identified challenges, and any relevant signal I can find. This step is what makes the next one, writing a message, possible at all. A message written from a name and a job title reads like every other cold message a prospect gets that week. A message written from what the person actually does, what their company is dealing with right now, and a signal that suggests the timing is right reads like it was written by someone who did homework, because it was.
I go deeper on what to enrich and in what order, and why sequencing matters more than the tool you pick, in B2B lead enrichment: what to enrich before any outreach. One detail worth flagging here since it changes how you should think about cost: enrichment itself does not consume credits. Finding an email or phone number does, and only when a usable one is actually found; a search that comes back empty costs nothing. This is also why a merged first name in a template stops working: without this step, the message below has nothing real to draw from.
Step 4: Generate a Message From the Prospect’s Actual Context
This is the step where the previous three either pay off or get wasted. I write the message using the offer, the persona’s tone and objective, the prospect and company profile, and, for a follow-up, the prior conversation history, so the angle varies instead of repeating itself message after message. You can edit or regenerate before anything goes out, and I will not send anything without your review unless you have already turned on Auto.
Take a concrete case, because this is where a use-case list stops being useful and a real example matters more. Say the persona is “VP Sales at a 50 to 200 person B2B SaaS company, commercial objective: book a meeting.” I find a VP Sales at a 140-person SaaS company that posted two new AE roles in the last month, score the profile 5 stars (role match, company size match, hiring signal), and enrich it with the fact that the company just announced an EU expansion. The message I generate for LinkedIn does not open with “I hope this finds you well.” It opens with the hiring signal and the EU expansion, ties both to the specific problem the offer solves for a sales team scaling into a new region, and closes with the meeting ask the persona is configured to make. Change the persona’s objective to “start a conversation” instead of “book a meeting,” and the same enriched profile produces a shorter, lower-commitment message rather than a calendar link.
That is the part a tool stack cannot do on its own: score, enrichment, and message generation drawing from the exact same context in one pass, instead of three separate outputs someone has to stitch together by hand between three different logins. If you want the full picture of what an AI prospecting agent does across the entire chain, not just this one message step, that comparison lives here. I go deeper on this specific step, from company and person understanding to the finished subject line and follow-up angle, in how to personalize cold email with AI.
Step 5: Send, Then Decide How Much You Validate
By default, I prepare the action and wait for your approval before anything reaches the prospect. That default is not a limitation to work around; it is where most people should stay for their first cycle, so they can see exactly what gets sent and why before handing off any part of the AI prospecting workflow.
Once you trust the output, two automation levels are available, and the difference between them is what I actually execute without you in the loop, not the quality of what gets produced. In Semi-Auto, I build your prospect database automatically (discovering, scoring, and enriching prospects on my own) but never contact anyone; you still manage every message yourself, so it works even before you have connected LinkedIn or email. In Auto, I go further: I calculate the next best action for each prospect, send LinkedIn invitations, messages, and emails, follow up within the limits you configure, and stop the moment a prospect actually replies, handing control back to you immediately. In neither mode do I run a fixed campaign on a timer; I work prospect by prospect from that prospect’s real state and history. Measured data across 2,000+ founders shows most land on Semi-Auto once they stop trying to review everything by hand.
If this is your first cycle, you can run it entirely with manual approval on a 14-day trial before deciding whether Semi-Auto or Auto fits how much oversight you actually want to keep once volume goes up.
Step 6: Follow Up and Qualify the Reply
I do not run a fixed sequence of follow-ups on a timer. I recalculate the best next action once a prospect becomes eligible again, using their current state and history, up to a configurable maximum (5 messages across LinkedIn and email by default, spaced at least 5 days apart, is the recommended starting point, and it is editable before you activate anything).
When a reply comes in, I detect it, attach it to the right prospect, and classify it: hot, warm, cold, or stop. Going back to the VP Sales example from step 4: a reply along the lines of “interesting timing, happy to grab 15 minutes next week” gets classified hot, triggers a notification, and moves the prospect to Responded, because that is the moment a person needs to take the conversation over. A reply I cannot classify with enough confidence gets marked for manual review instead of guessed at, and a hard no gets tagged stop so it never gets contacted again. This is also the point where the cycle stops being about volume and starts being about the actual conversation, which is exactly where I should hand things back to you rather than keep going on my own.
Once a handful of prospects have moved through this cycle, the operational view worth checking daily is the pipeline itself: how many sit at To prospect with no action taken yet, how many are in Prospecting waiting on a follow-up, and how many have moved to Responded. That breakdown, paired with the LinkedIn and email reply rates calculated from prospects actually contacted, tells you within a week whether the persona and the message are working, without needing a separate reporting tool to find out.
Where a Generalist AI Stops and an Agent Takes Over
A fair question at this point: why not just run steps 1 through 6 with ChatGPT? A generalist assistant can draft a decent cold message if you paste in a job title and a company name. What it cannot do is remember your offer and every prospect’s history across sessions, find and score new prospects on its own, or send and track anything once the draft is done. Every step above stays a copy-paste job between a chat window and wherever your prospects actually live, which is a different workflow from the one I describe here, not a shortcut through it.
I have written the full feature-by-feature comparison in LEO vs. ChatGPT and Claude for prospecting, and the broader case for what makes an agent structurally different from a chatbot or a sequencer in what an AI prospecting agent actually does. The short version for this guide: a generalist AI helps with one message at a time when you ask it to. I carry the brief, the score, the enrichment, and the reply history forward through every step above without you re-explaining any of it in a new chat tomorrow.
Common Mistakes That Break a First AI Prospecting Cycle
The most common failure is not the AI producing bad output. It is running the cycle with inputs that were never going to work, then blaming the result on the technology instead of the brief that fed it.
Sending before enrichment finishes is one version of this. A message generated from a name and a job title, with no signal or company context behind it, reads exactly like the generic outreach the prospect already ignores a dozen times a week. The fix is not writing harder or adding a clever opening line; it is not skipping step 3 to get to sending faster.
Ignoring the score threshold entirely is another. Contacting every prospect a search surfaces, including the ones scored below your own cutoff because the list “still looks fine,” dilutes your reply rate and makes it impossible to tell afterward whether the targeting or the message itself was the actual problem when replies stay flat.
Turning on Auto before a manual cycle has proven the message quality is a third, and probably the costliest one, because it scales a mistake before you know it is one. Auto is a scale decision, not a starting point. Validate a batch of messages by hand first, confirm the tone and the offer framing land the way you expect them to, then decide how much of the sending you actually want to hand off. The same caution applies to Semi-Auto: it only builds the database, but a persona fed a threshold that is too loose will still quietly fill your pipeline with prospects nobody should be spending a message on.
The last one is treating a first cycle’s results as final. A 3-star threshold, a persona’s tone, and a message angle are all starting points meant to move once real replies start telling you what works and what does not. Running a second cycle with those adjustments takes minutes inside the same conversation, not a rebuilt campaign from scratch, which is the entire point of keeping the context in one place instead of across five tools.
