How to set up two AI agents for one sales rep: a CRM updater and a coach

The short answer. Write each agent up the way you would a hire: a job description, a goal, the capabilities it needs, and the rules it has to follow. Christopher O'Donnell's example is one rep with two agents. The first keeps Salesforce current from her meetings, emails and customer Slack, following a written spec for when a deal may change stage. The second sends her a ranked list through the day, with emails already drafted, deals that are slipping, and call prep built from what has worked in similar deals.

What the video covers

• 2:36 Engineering teams can share one set of agents because everyone works the same way. Reps each run a different book and motion, and each answers for a number of their own.
• 3:37 The org-chart question: ask who you would hire, then write the job description, set the goals, and decide what the agent needs to be able to do and to know.
• 4:09 An agent defined like a hire needs what a hire needs: guardrails, security, oversight and evaluation, which can't be left to each rep.
• 6:44 Before any agent runs, RevOps cross-references the sales process with the Salesforce setup and writes one spec for when a deal moves stage and when it must not, so agents move deals only on evidence.
• 7:47 Agent one works from every meeting, email, customer Slack thread and API event, updates fields when a meeting ends, fills the MEDDPICC fields, leaves the fields the CRO wants typed by hand, and emails the rep a weekly summary. RevOps owns its skills as managed skills the rep can't change.
• 9:20 Agent two messages the rep through the day with a ranked list: emails to send with drafts attached, deals falling through the cracks, prospecting touches, and an urgent tier saved for time pressure that can be verified.

Excerpts

[0:00] I turned Claude into a sales rep that works 24 hours a day. If your team already runs Claude Code over a company brain repo of some kind, with customer docs and sales process, meeting feedback, you are already halfway there. But you may also feel right now like your stack feels worse every month. Why does every tool that you add make your days and your week feel heavier? Now, there is an elite RevOps approach […]. And it goes the other direction. It compounds in a good way. Each problem these teams solve leads to multiple other problems getting solved for free.

[0:40] Why should you listen to me? I spent a decade at HubSpot leading product before starting our company now, Day AI. Today my job pulls me inside the AI stacks of RevOps teams all day, every day. What I see everywhere is this: everyone has some company brain. These ICP definitions, sales process docs, feedback from meetings. It's technically a harness, but you can think of it just as whatever folder RevOps folks are opening Claude Code in. That part is great.

[1:11] Then […] teams get stuck, because the plan is that everyone will sit in front of Claude all day and Claude will be this perfect work surface. In practice, it's slow and messy. Sales reps are told to clone a repo from GitHub. You ever tried telling a sales rep to go clone a repo from GitHub? Okay. Then everybody open Claude Desktop in that folder that you just cloned. Now always remember to git pull, but never git push. You can't do that. Okay. Now tell Claude what you want to put in Salesforce. Wait a few minutes for that to happen. Okay, then open Salesforce to make sure it happened. It sort of did. And then finally retype everything into Salesforce by hand. […] It's not just Salesforce, it's basically all these tools that we've had around forever.

[…]

[2:18] But every month RevOps is adding info, adding capabilities to this company brain they have. Now, the old world and the new world are just not converging for these teams. They're diverging. And the point is not necessarily that you bought the wrong tools. So what do some folks know that you don't? Here's a way to think about it. They know how software engineers work. If you talk to an elite engineer today, they will tell you, "I don't even write code. The vast majority of my work happens for me in the background 24/7." […]

[3:03] So should RevOps just copy what the engineers are doing with AI? Yes, but there's a catch. Engineering teams do the same activities the same way across the whole team, so they only need one shared set of agents and skills. Go-to-market is messier. Every seller is running a different book of business, maybe a different motion. They have their own style they bring into it. And each one of them, and this is super important, is held personally accountable for a financial result. It's about outcomes, not activity.

[3:35] I have the answer for you. Go-to-market teams just need a different metaphor. Just use the mental model of the org chart. When something needs doing, ask the question you would ask about a human: who would I add to my team? Write the job description, set the goals, figure out what capabilities it needs, what it needs to be trained on. That is all you need today to produce a modern AI agent that works autonomously in the background and then interactively with you when you want to talk to it. Don't ask what software you'd buy, ask who you'd hire.

[4:12] But the moment you define an agent like a hire, you inherit everything else a hire needs. And this is where the do-it-yourself path turns from frustrating into dangerous.

[…]

[4:50] Imagine agents like that running around your whole sales team, sending email, updating CRM, messaging your customers. This stuff needs to be guardrailed, secured, watched, evaluated. And on a go-to-market team, that can't be expected of each individual sales rep. But at the same time, it has to have that feel of a moldable, coachable personal agent that the rep can talk to and ask to behave a certain way. This is one of two fundamental unlocks that Day AI gives you: the ability to do everything I just described. The second is that everything the agents do, they're reading and writing from one secure shared context graph. We call it the customer memory layer. Let's build some agents in Day AI so you can check it out.

[5:36] Here I am in Claude Code. I'm pointed at a company brain repo, probably very similar to the one you already have. I have ICP, I have pipeline, sales process definition with a bunch of detail. I have meeting notes from internal planning meetings, ops meetings. I have feedback and directives from the CEO, finance. I have everything. My goal is I want my sellers effectively working around the clock. Now, a human can only work the hours she's awake, rested, available, not on vacation, but agents don't sleep. So for 16 hours a day, it's going to be just the agents.

[6:13] Let's make it concrete with a rep. We'll call her Margo. Margo may not be real, but like any good sales rep, she swears her pipeline is absolutely all real. She has 17 opportunities in Salesforce. Great. Let's take a peek at the calendar. Okay, she blocks these 90-minute blocks on Tuesday and Thursday in the afternoon. That's for CRM data entry, super common. That's 3 hours a week of typing facts into fields. Joyous. Okay. And what about the meetings that were already recorded? Got to put those in. That's pretty dumb. It looks like she's getting two, maybe three high-value calls done a day, plus another two or three high-value activities of some kind: really good email threads, hot prospect responding, whatever it is. Let's make our goal to double all of that.

[7:04] Now, one prep step before we do the agents. Let's take the sales process definition and cross-reference our Salesforce implementation. Let's write one canonical record, one doc and spec of when a deal should move from stage to stage and when it absolutely shouldn't. And we're going to have an emphasis on verifiability and specificity, because I'm a good student of Mark Roberge. Now, these are some of the guardrails that I'm talking about here. We have the behavior, and it's going to follow that behavior. The agents are going to work from this and only move deals when it has evidence that it absolutely should.

[7:40] Okay, now, I'm thinking two agents per seller. The first one, super straightforward, is going to be a CRM data entry specialist. Job description: do Margo's Salesforce stuff. Do all the data entry and work off of absolutely everything: all of the meetings, the emails, customer Slacks, integration events that come over the API, notes that get added in by other agents, everything. And it can also run on events, so that update happens right when a meeting ends. It all kicks off, however I want this to work. Margo's running MEDDPICC. That's great. Let's fill out those MEDDPICC fields in Salesforce.

[8:18] Okay, but internally, our CRO feels really strongly about this: a couple of these fields need to be filled in by hand. It forces discipline and thought process. Cool. We'll play along. So what it's going to do is, every week it emails her a state of the union on her Salesforce data. I can push this live, and on day one, the data entry blocks are just gone from her calendar. We're done. All of them. 100%. And an important note: because this is such touchy stuff with Salesforce, I'm actually going to do these as managed skills, meaning I own these in RevOps, and it doesn't matter what Margo does, she's not going to be able to change the rules. Again, guardrails, super important.

[9:02] Now, I could stop there, and I've given Margo a ton of time and energy back in her day. I've gotten the Salesforce data to be super clean for forecasting, for communicating with the CEO. That's great. And we could just stop right there. […] But why stop? Let's add a second agent. And this is going to be sort of a coach, a prioritizer, a task manager, somebody to keep Margo spending every minute and every hour of her work day in the best possible way. Job description for this one: research around the clock, make sure Margo spends every working hour in exactly the right way. Every, let's say, 4 hours, it's going to Slack her. It's going to send an ordered list, a prioritized list: emails to send. And by the way, those are going to come with a pre-written email draft that's super well written for her to review.

[9:56] Let's also make sure to send her a section of details about deals, places where stuff is falling through the cracks. And then later in the day, let's prioritize and send those prospecting touches, those prioritized follow-up mental notes that we have: to connect with this person on LinkedIn, to send a follow-up note to that person, etc. Oh, and at the top, I want an urgent tier. Like, drop what you're doing, leave the meeting you're in. This is absolutely reserved for cases where there's huge time pressure and it's verifiably real.

[10:28] Now, because every recorded word of every deal, every customer conversation lives in this memory layer I mentioned, her prep for these high-value calls comes with what has worked in this exact situation before. It's not just regurgitating public information about the people on the call that you've met with six times already. It's like: what do I actually do? This agent is going to bring Margo the answers to every test. So what is this all worth? Like, what's the ROI here? Replace calendar data entry hours with selling time, […]

[11:17] This is a big lift, but then it starts to compound. These are just much more enjoyable work days for Margo, so she's more likely to stay at our company. She's more likely to make P-Club, more likely to close happy customers in the right way. If all this AI stuff is stressing you out these days, if it's confusing your sales team, there is an easier way, and we're happy to set it up with you. Click the link below and let's get you some 24-hour-a-day sellers. This thing is real. You got to see this.