From $0 to $400M: The Playbook I'd Run Differently Today (With AI)

In August 2010, I walked into Groupon with a mandate and no team. Two years later, we had 400+ sellers generating $415M in revenue across 23 markets.

Looking back, I can divide that build into three distinct phases: the first ten hires, the scale to 100+, and the machine at 400. Each phase had different challenges, different failure modes, and very different decisions I'd make today if I were doing it with the AI tools that exist right now.

Phase 1: The First Ten Hires (0–30 Sellers)

What I Did Then

The first hires were everything. At that stage, you're not building a process - you're finding the people who will define what the process becomes. I hired for raw sales instinct and intellectual horsepower over pedigree. I wanted people who could figure it out in a market we were inventing in real time, not people who already knew how to run a mature playbook.

ICP development in those days was brute force. We called everyone. Merchants of every size, category, and geography. We made thousands of calls and built the pattern recognition manually - learning who bought, who didn't, what made the difference, what pitch landed. It took six months of calls to develop the intuition that told us where to focus.

Forecasting was gut. I called the number based on my read of the team's pipeline and my sense of the market. Sometimes I was right. Often I wasn't. But at 10–30 people, the cost of forecast error is recoverable.

What I'd Do Differently

The ICP development alone would be transformed. Today, you feed your closed-won data into an AI model after the first 30–50 deals and it identifies the firmographic and behavioral patterns that correlate with fast closes and high retention. We would have known in week three what we figured out in month six. That's not a marginal improvement - it changes your entire ramp trajectory.

I'd also use AI-powered candidate sourcing and assessment from day one. The signal-to-noise ratio in early hiring is brutally low. Tools that can analyze communication patterns, cognitive ability markers, and situational judgment responses at scale would have improved my first-ten-hire batting average meaningfully. Instead, I relied on my gut and a lot of phone screens.

Phase 2: The Scale (30–150 Sellers)

What I Did Then

This is where the build gets genuinely hard. At 30 sellers, you still know everyone personally. At 100, you don't. The transition requires you to move from being the player-coach who is in every deal to being the system designer who builds the process that gets deals done without your direct involvement.

The enablement challenge at this phase nearly broke us. You're hiring faster than you can train. Onboarding that worked for the first 20 people doesn't scale to 100 new hires a year. I built content, ran trainings, and hoped enough of it stuck. The reps who ramped fastest were the ones who happened to sit near a high performer and got informal coaching through proximity. That was random and inequitable.

Hiring quality varied wildly. When you're hiring 10 people a month, some months you nail it and some months you don't. The cost of a bad hire at this stage - in management time, in team morale, in the deals that don't get worked - is substantial.

What I'd Do Differently

Enablement would look fundamentally different. Every rep would have an AI deal coaching tool that gives them immediate feedback on their qualification - the same feedback I would have given them if I had infinite time, applied to every deal in their pipeline, 24 hours a day. The Deal Coach on this site is exactly what I would have wanted at Groupon in 2011.

The reps who ramped fastest wouldn't have been determined by who sat near whom. They'd have access to the same AI coaching tools regardless of desk location. The performance distribution across a 100-person team compresses when every rep has access to good coaching, not just the ones lucky enough to get manager attention.

Forecasting would use behavioral signal scoring from month one. I wouldn't wait until the quarter was two-thirds done and the slipping deals were obvious. I'd have a system that flags the deals decelerating 3 weeks before the rep does.

Phase 3: The Machine (150–400+ Sellers)

What I Did Then

At 400 people, the build stops being about individual rep performance and becomes about system design. Every decision - hiring profiles, compensation structure, territory allocation, quota setting, promotion criteria - is now a policy that affects hundreds of people and has second and third-order effects you can't fully predict.

SDR research and list building at this scale consumed enormous resources. We had people whose primary job was finding phone numbers. That's not an exaggeration. The data quality across 23 markets was inconsistent, the lists were always stale, and the reps spent meaningful hours every week on prep work instead of conversations.

Competitive intelligence was informal. Smart managers tracked what they heard in deals and shared it in team meetings. But there was no systematic mechanism for capturing what competitors were doing, pricing, winning on, and losing on across 400 deals a day.

What I'd Do Differently

The SDR productivity math would be unrecognizable. With a Clay + AI stack, one operations person does what required a team of researchers in 2010. Your SDRs spend their entire day on conversations, not prep. The unit economics of outbound change fundamentally - and at scale, that difference compounds.

Competitive intelligence would be autonomous. An agent monitoring competitor positioning, pricing signals, product announcements, and review patterns across 23 markets - flagging changes in real time and delivering synthesized intelligence to the managers who need it - that's infrastructure. At Groupon's scale and speed, competitive blind spots cost money. I had too many of them.

The One Regret I Haven't Talked About Publicly

Here it is: I scaled headcount too fast in markets where we hadn't proven the motion yet.

At Groupon's growth pace in 2010–2012, there was enormous pressure to expand. New markets, new teams, new managers. We were adding cities weekly at our peak. And in the pressure of that growth, I made the mistake every scaling executive makes: I moved people into leadership roles because of tenure, not demonstrated readiness.

I promoted sellers who were excellent individual contributors into management roles they weren't prepared for, because the org needed managers and they were available and loyal. Some of them made it. Some of them didn't. The ones who didn't took good reps with them on their way out - through turnover, performance issues, and team culture damage that took months to repair.

With AI-powered performance analytics and better assessment tools, I would have had a clearer signal earlier about who was actually ready for the transition. Not a guarantee - management readiness is still fundamentally a human judgment - but better data would have produced better decisions.

The Playbook That Doesn't Change

After all of that - the wins, the mistakes, the things I'd do differently - here's what I've concluded is genuinely durable:

The quality of your first ten hires determines the ceiling of your culture. You can overcome a lot with the right early team. You cannot overcome the cultural debt of too many wrong early hires, regardless of what tools you have.

Clarity of ICP, message, and motion is more valuable at every stage than any tool. AI amplifies direction. It doesn't create it. If you're going in the wrong direction, better tools get you there faster.

Managers who develop their people outperform managers who direct their people - at every scale I've seen. The coaching culture you build at 30 people determines what's possible at 300.

The playbook hasn't changed. The speed at which you can execute it has changed by 10×.

If I were building Groupon's sales org today, I'd hire fewer people, arm them with significantly better tools, set a higher bar for what "ready" looks like at every stage, and move faster on the markets where the motion is proven while being more disciplined about the ones where it isn't.

The output would be the same. The efficiency - and the number of mistakes I'd avoid - would be unrecognizable.