Open your laptop tomorrow morning. Which window do you go to first?
For most sales people the honest answer is the CRM or the inbox. Those are the two tabs that never close. Everything else is a visit.
Mine is neither, and has not been for months.
I am a founder doing founder-led sales at Big Sister AI. There is no SDR to hand the routine to. I am the pipeline, the follow-up and the CRM hygiene, on top of building the product. On most days I do not open Pipedrive and I do not open my mailbox. I work in Claude Cowork. The routine happens there. The decisions are still mine.
This is not automation in the old sense. Nobody wrote a Zap. There is no workflow builder anywhere in this. It is an AI-first sales stack built one step at a time, where each step is something you already do by hand, taught once and then handed over.
Five steps. Each one takes a day or two. You do not skip ahead.
Step 1: Connect everything into one place
The test is simple. Open your browser. Every tab you keep open to do your job is a connector you need.
For me that is:
- CRM: Pipedrive
- Email: Spark, sitting on top of four work mailboxes. This matters more than it sounds. Founders accumulate addresses, and controlling them all from one place is the difference between a system and a mess.
- Notetakers: Granola and Fireflies. Two, not one, because they cover different meetings.
- Tasks and projects: ClickUp
- Calendar: Google Calendar
- Files: Google Drive, mounted as a disk on my laptop, so whole folders connect as folders in the chat
That last one is the trick most people miss. Drive as a local disk means an agent reads a client folder the same way it reads a CRM record. No upload step, no copy-paste.
What is missing from that list is as interesting as what is on it. No dialer. No enrichment tool. No sequencer. You do not need a complete stack to start. You need the tabs you actually live in.
Step 2: Make it the prime agent, not a tool you visit
Some people run this through Telegram. Some build a Slack bot. I use Claude Cowork, and the reason is not the model. It is that Cowork writes markdown files and folders on my laptop, and those files survive the session.
That is the whole difference between a chat and a system.
A chat forgets. A system accumulates. Every session I run adds to a set of markdown files that separate my projects, my task types, my technical work, my customer work. Next session starts where the last one ended. Knowledge grows instead of resetting.
The practical version: it is my Jarvis. One prime agent. I talk to it on my laptop all day, and when I leave, I open the same session on my phone and keep going. Not a different assistant on a different surface. The same conversation.
If your AI lives in a tab you visit when you remember to, it will never become the place the work happens.
Step 3: Run it manually first, deal by deal
This is the step everyone wants to skip, and skipping it is why most people's AI setup quietly dies in week two.
Do not write a prompt. Do not build a workflow. Open a session and work.
The pipeline session
Start here:
Help me prioritise which deals in my pipeline I need to grab first. Ask me anything about the deals and the pipeline you need to understand, and tell me where you are not sure that is the right way.
It comes back with a read: this stage matters, these deals are hottest. Some of it is wrong. You say so. That correction is the point.
Then:
Go one deal at a time. Give me a short story about the deal, then your advice on what to do: what to send, what follow-up to make, which fields to fill.
Deal by deal, you approve or you redirect. What looks like a slow afternoon is actually you teaching your AI coworker how you read a pipeline. And while you do it, you update the playbook and the rules files in the same session.
The inbox session
Same shape. You do not open Spark. You ask:
Look at all my active emails. Tell me what is here, where I need to respond first, group the rest, and tell me what you think I should send.
Then, and this is the part that makes it work:
Give me three drafts for the top three.
Three at a time. You read them, you approve, and it sends. Spark lets it send. That is a real handover, not a suggestion box.
The rule underneath all of it
One instruction does more work than everything else I have written. Read everything before you write anything.
Before drafting a single reply to anyone, the agent:
- Reads every earlier thread with that address, and looks for other addresses belonging to the same person
- Opens Pipedrive and finds the contact, the open deal, and every closed deal we have had
- Reads the notes and the conversation history on the record
- Pulls what the notetakers captured from any call with them, Fireflies or Granola
- Only then tells me what to send, and what the CRM is now missing
A human SDR has never once done all five before replying. That is the actual advantage, and it has nothing to do with writing better sentences.
Step 4: Turn the session into a skill
Work like that for a day or two. Handle real deals, real email, without opening the underlying tools. Then say:
Create a skill based on this session.
Now tomorrow is different. New session, you say "let's handle my inbox," and it loads the skill and starts working the way you taught it. At the end of the day:
Update my skill with what we changed today.
That is the loop. Session teaches skill. Skill runs session. Skill gets updated. Run it for a week and the thing knows your pipeline stages, your tone, your buyers, your exceptions.
You can also get the AI to build the first draft of a skill for you. What I ask for:
Analyse the tools I need for this process and build me a skill. Ask me whatever you need to about how the process actually runs. Look at how I have used these tools before. Ask me where each type of data should be kept, and keep the rules separate from the steps.
Then I hand it the playbook I already had and explain it. Most sales leaders have a playbook nobody reads. This is the first time it becomes executable.
Step 5: Schedule it, then widen the gate slowly
After three to five days of the manual loop, the skill is good enough to run without you. Now it goes on a schedule.
Every source becomes a trigger. Calendar, notetakers, calls, email. Something happens out there, something runs in here. My daily schedule reads the day's transcripts, updates the deals, prepares what needs to be on the calendar, and sends me one message with the decisions it could not make alone.
Autonomy is not a switch. It is a gate you open one notch at a time. My inbox routine spent its first phase in what I call shadow mode, and the instruction was explicit:
You have exactly two write permissions: apply triage labels, and create drafts. You must not send any email, archive anything, move anything to spam or trash, or create anything in the task manager. Tasks are proposed in the report only. These restrictions get lifted stage by stage once accuracy is proven. If you are ever unsure whether an action is permitted, do not take it. Report it instead.
Three more rules earn their place in every skill I run.
The safety rail. Never classify as quiet anything touching legal, overdue or failed payments, investors, a hard deadline, or a top-tier contact. When torn between a quiet class and a visible one, always pick visible. A missed important email costs far more than a noisy brief.
Never guess in the CRM. One open deal, link it. Zero open deals, skip it. More than one, do not link, report it as ambiguous with the deal names. Never guess. Never touch a won or lost deal.
Live data comes from the system, not from a document. A specific deal, a specific task, a specific call gets queried live. A document is where the standard lives, never where the current state lives.
What still breaks
If a guide has no failure section, someone is selling you something.
Notetaker filters lie. A "my meetings" filter silently returned zero results for a two-week backlog, and zero looked exactly like nothing happened. The standing rule now: before an empty result becomes a finding, prove the instrument reached the thing.
CRM pagination has hard ceilings. Pull activities above a modest limit and the response overflows. Cap it and iterate.
Non-English calls degrade badly. Our non-English transcripts garble often enough that the rule is to file them as low confidence rather than let anything reconstruct what was probably said.
Long runs die silently. Twice a scheduled job finished all its analysis and never wrote its output. Any routine that runs unattended needs to be able to detect and finish an incomplete previous run.
None of this is a reason not to build it. All of it is a reason to keep a human at the gate.
Where this ends up
Do this for a month and you arrive somewhere uncomfortable.
Your agents are now writing to customers. Drafting the follow-up. Answering at 11pm. Doing the first pass on inbound. And you have no idea whether any of it is any good, because the only thing you ever checked was whether it sounded fine when you skimmed it.
That is the same problem sales leaders already have with people. You review a handful of calls a month and manage everyone else on faith. AI does not create that problem. It multiplies it, because now the thing you are trusting on faith scales to a thousand touches a week.
The fix is not to slow down. It is to have a standard that applies to both.
At Big Sister we score every customer conversation against one written standard: every closing skill graded one to five against a rubric, rolled into a single Sales Score from 0 to 100 per rep and per team. Every score cites the moment in the conversation that earned it. Any rep can dispute any score, and the dispute goes to a trained human referee, not back to the model.
We built it to score humans and agents from day one. That was the point. If you cannot score a closer, you cannot score an agent, and you will put both in front of buyers on faith.
CRM is the record. Your recorder is the tape. Big Sister is the method and the referee that scores every game.
Build the stack in this article. Then decide what you are willing to let it do unsupervised, and how you will know it earned that.
Build your version
There is nothing to download at the end of this. Every prompt I run is printed above, in full, including the ones I would rather have cleaned up first.
So do the obvious thing with them. Open Claude Cowork and hand it this:
Read https://www.big-sister.ai/blog/ai-first-sales-stack. Then look at the tools I actually have connected and the way my pipeline is shaped, and tell me what my version of step one looks like. Ask me whatever you need to about how I work and where my data lives. Then build it with me, starting today, one step at a time.
It will get something wrong about your pipeline. Correct it. That correction is exactly the work step three asks for, and you just did the first hour of it for free.
Then score what you build. Send us your closers' recorded conversations from Fireflies, Zoom, Meet or your CRM. Within days you get a Sales Score per rep with the evidence behind every number. Start with Solo if you want your own score this week. Scoring a team is the second path: book a demo and we onboard you personally.
Val Yaromenko is CEO and co-founder of Big Sister AI. He has been building B2B sales teams since 2009. Read The Closer's Manifesto.