Chat is the wrong primary interface for a sales rep. Every CRO I know wants the same basic thing out of their GTM harness and tools: your middle and bottom performers taking the actions your very best sellers already take. Make chat your main surface, or run everything through Slack, and the workflow turns into ask and receive. A rep types a question, gets an output, and everybody hopes they asked for the right thing. What if they didn't?
Picture a normal morning for an AE. They've got 20 mid-funnel tasks sitting in front of them, plus a big presentation coming up that should take precedence over almost everything else. That prep looks like a 45-minute block until they hit a snag and it turns into a 2-hour proposal that pulls in solutions engineering. Dropping somebody into chat and saying, "what's going on? What should we do? Text me. Tell me like, tell me what you need," is not directive enough for that order of operations. I want reps working out of an inbox, a queue, or a cockpit that suggests what to do next and in what order. That holds in SMB, where you want to be very directive, and in enterprise, where the advantage of suggesting a higher-value move is even greater.
I jumped straight into that argument on Brendan Short's The Signal Live Show before we'd even done introductions. Brendan had brought together Tim Geisenheimer (CRO at Hatch), Parth Gujare (PM on Ramp's revenue team), and me to debate a question a lot of revenue leaders are stuck on: should AI systems in GTM be centralized or decentralized? Spoiler alert: we didn’t really disagree or debate ¯\_(ツ)_/¯
Your reps are software users
Parth told a great story that sums things up (and shows that even companies like Ramp make mistakes in this AI journey). At Ramp's company kickoff in February, co-founder Karim Atiyeh went on stage and told everyone they needed to be a builder and use AI. They gave the whole team basically unlimited tokens. For the first month people were fired up, and pretty quickly it turned into about 3 months of chaos: 17 versions of the same dashboard, things breaking, and builds people were excited to demo once and then barely used. When Brendan asked how many pre-call research agents they ended up with, Parth said, "we had a ton."
Making decisions about your GTM tech stack, build vs buy and rep experience becomes a lot easier when you wrap your head around the fact that your reps are software users and your GTM stack is a B2C product you're taking to market (a very difficult market lol). As I said on the show, "You wouldn't let your users write code and build your product." Some users are sharp and bring you real problems. Others show up with N-of-1 ideas that only make sense for their own workflow.
Traditional RevOps hasn't worked this way. You sit in a silo with the CRO, build a vision for the future, lock down a bunch of fields so reps have to do what you say, and lob a huge launch into the team. You wouldn't build a product by forcing users into actions they don't really want to take. You spend time with users, ship an MVP to an alpha group, watch the metrics, and build low-friction flows that pull people through. Parth summed up how Ramp runs this: "we're looking at the metrics, we're talking to the users, and then we're continuously building the thing."
I got an up-close look at that culture back at Shopify, where the unofficial stack rank was engineering, product, design, growth, janitorial services, and then sales. Humbling if you were in sales, and also a useful reminder of who actually knew how to build. Ramp and Cursor got ahead here because they put real PMs on the GTM stack. Most companies aren't Ramp and can't afford a PM for go-to-market, so RevOps has to become that, and sales leaders have to start thinking about the job the same way.
Centralize the foundation so the right people can build
Tim Geisenheimer had the best caveat of the panel. He agreed with Parth and me about 90% of the way. Hatch has about 100 salespeople, and you can't have everyone shipping their own tools as the only strategy, so they centralize a context layer and core tools for things like sales deck generation and warehouse access. Tim still leaves about 10% room for individual experimentation: "it's kind of okay if some people are building at the last mile for themselves or doing some things for themselves that help them win." When a rep builds something that genuinely works, you can pull it back into the core stack for everyone else.
Brendan put a 4-level ladder of GTM AI sophistication on screen (a framework he joked he stole from me, or maybe I stole from him):
Level 0: exploratory, like a rep poking around in ChatGPT
Level 1: reusable tools
Level 2: workflow automation
Level 3: infrastructure and leverage
Level 4: a recursively improving system
Turning everyone loose to build at levels 0, 1, and 2 is how you get 17 versions of the same dashboard. Tim's 10% starts working at level 3, once governance, permissions, a shared context layer, and baseline infrastructure are in place. Ramp built that layer into how their team touches data. Reads pull from a governed data layer they maintain (a unified Snowflake or Postgres table), and they blocked most writes from native Salesforce or Gong MCPs. Write actions route through Ramp's own MCP and APIs, where everything is logged, committed, and can be rolled back.
Inside a revenue org, my expectations climb fast at manager, director, and up. Give leaders a GTM repo, shared skills, automated checks and tests, and hosting behind SSO, and they should be building and pushing to prod, including the occasional throwaway Vercel app.
The night before the panel, I had an itch I couldn't scratch about why a set of metrics wasn't reconciling. I sent Claude at it, and it went through Snowflake and Salesforce and handed me the answer. In a prior life that was a 4-day biz ops or data science project, the exact kind of question Tim said used to sit "at the bottom of the list, there's 10 other things ahead of it." Claude could pull the right answer because somebody had already built the semantic layer underneath, and everyone running teams needs to get comfortable doing that kind of work themselves.
Start from an outcome back, then bank the momentum
Listening to Parth walk through Glass and Inspect, Ramp's internal AI tools, is a lot (in a good way). When Brendan asked what he'd tell an earlier-stage company that isn't Ramp, Parth said his advice might sound dumb: start with the metric. Hire a data scientist if you need one, map the core flows, and "just start with the problems you actually have and build the, like, smallest thing that you think will, like, move the metric." On the panel I called it building "an outcome back, not infrastructure out."
Ramp's own starting point was modest. After they deprecated their automated SDR function (OATS), the first thing Parth's team built was a plain Retool app, and it hit 90% adoption because it actually helped reps do their jobs. Then they followed the data. Hybrid sequences (email plus calls) beat email-only and call-only all the way through the funnel, SDRs were burning TAM by dialing everybody, and emailing ate 2 hours a day.
So they built an internal sequencer that cut that email time to 30 minutes. Email volume doubled over the last year, about 95% of threads now run through it, and when stale contacts started pushing bounce rates up, they added a validator at send time. They just kept following the next constraint in a workflow reps already used.
We're all very AI-pilled people, so we forget how much of the mid-market and enterprise market still has no AI momentum. I've spoken at a few events lately and I'm still surprised by that. Find a 10% performance or capacity improvement without disappearing into a giant infrastructure project, and you start to build belief. You build momentum with your reps so they don't think AI is going to take their jobs, they pay attention the next time you ship, and you can get the money you need for the next thing. The building of internal momentum across all of those stakeholders is undervalued in a big way.
If you want something to run with on Monday, skip the 6-month roadmap. Pick one bottleneck your reps already complain about, pull the baseline metric, and ship the smallest version that could move it. Brendan Short's bar is even sharper: figure out what can drive a metric next week, even if you throw the whole thing away in 2 months.
Huge thanks to Brendan Short for herding us through the live-stream chaos, and to Tim Geisenheimer and Parth Gujare for shooting straight. Hope this helps.
Our segment starts at 41:03 on the replay, and the chat UX rant kicks off around 43:58: https://x.com/TheSignalclub/status/2122174331664883833

