The modern sales org is not designed to learn. It is designed to report what happened after the fact.

That made sense when humans had to inspect the work, summarize the pattern, and decide what should change next. It makes less sense now when AI can turn sales management into a feedback loop itself: a system that observes what happened, advises on the next best move, measures the result, and improves the motion without waiting for the next forecast call.

A few days ago I watched a Y Combinator talk by Tom Blomfield, the founder of Monzo and now a YC group partner. It put language around something I had been feeling for weeks: the future sales org will not be a bigger hierarchy with better tools. It will be a smaller, sharper set of self-improving loops.

Blomfield was synthesizing a body of ideas pulled from several places: a thread from Jack Dorsey on how hierarchical companies are organized, work from his YC colleague Diana Hu on what she calls "AI loops," and an essay from Pete Koomen on why most companies are still using AI the wrong way. I kept translating every example into sales, because if there is one corner of the modern company still committed to humans-as-information-conduits, it is the sales organization.

This piece is my synthesis of that argument, what I think it means specifically for sales leaders, and the experiment I'm about to run to find out whether it actually works.

The sales org is the most Roman thing we still build

The Roman legions were engineered to project power across two continents. The mechanism was a nested hierarchy with consistent spans of control: named individuals at each level passed orders downward and reported observations upward. The chain was the information system. Jack Dorsey's framing, picked up by Blomfield in the talk, is that almost every modern company is still built on this assumption: that human beings are the conduit through which information flows up and down.

Look at a modern sales org. SDRs report to SDR managers. AEs and BDMs report to regional managers. Regional managers report to a VP. The VP reports to the CRO. Information moves up the chain in the form of CRM updates, forecast calls, weekly pipeline reviews, and Friday commit emails. Targets, playbooks, and rules of engagement move down. The humans are the bus, and the bus is slow. CRM updates arrive late. Each SDR and BDM has a different level of hygiene. Deal context gets fragmented across notes, calls, Teams messages, and memory. By the time the data is clean enough to trust, you often need weeks or months of observations just to make sure the signal is real and not background noise.

We have never seriously questioned this structure. We've layered tooling on top of it, from Salesforce and HubSpot to Gong, Apollo, and Outreach, but the underlying assumption that humans are the conduit has stayed intact.

That assumption is the thing AI breaks.

The wrong way to do AI in sales

A year ago, if you asked sales leaders what AI was doing for them, you'd get the same set of answers. Gong summaries. Apollo enrichment. CRM copilots that auto-fill activity logs. Sequence generators. AI SDRs that send the same cold email a human SDR would send, just at higher volume.

This is what Pete Koomen called the "horseless carriage" framing. We take the existing motion and bolt a faster engine onto it. The reps ship more activity. In a landmark 2023 study of more than 5,000 customer support agents, Brynjolfsson, Li, and Raymond found generative AI tools delivered a 14% productivity gain on average, climbing to 34% for novice workers. Useful, but the shape of the org doesn't change.

That framing is the thing to abandon. Not because the tools are bad, most are genuinely useful, but because they leave the architecture untouched. You still have the Roman legion. It just runs on caffeine instead of water.

Goldman Sachs' Jim Covello tracked exactly this dynamic across two years of enterprise AI deployments and concluded that 95% of organizations were getting zero return on their AI pilots despite $30–40 billion in spend. That isn't an AI problem. It's an architecture problem.

The interesting move is to stop bolting AI onto the sales org and start asking what the sales org would look like if it were redesigned around AI from the ground up. The answer, I'm increasingly convinced, is: a set of recursive, self-improving loops, with humans only at the edges where they actually add value.

The shape of a self-improving sales loop

Every self-improving loop has the same skeleton. Diana Hu has written about the canonical structure; translated into sales terms it looks like this:

Sensor layer. Call recordings, email threads, CRM stage changes, win/loss notes, prospect website visits, intent signals, opened proposals, response latency. Anything that tells the system what happened.

Policy layer. Rules about what the system can do on its own versus what it must escalate. Update the CRM autonomously. Draft a follow-up, always. Send it unsupervised below €1k MRR; flag for human review above. Never offer a discount without a human signoff.

Tool layer. Deterministic APIs the agent can call. Query the CRM. Score a lead. Look up a contact on LinkedIn. Read the calendar. Book a meeting. Push a deal stage forward.

Quality gate. Evals on email tone, factual accuracy, brand voice. Human review on anything above a price threshold. Sampling reviews of autonomous decisions to catch drift.

Learning mechanism. Did the meeting book? Did the deal progress? Did the prospect respond? Did we win? When we lost, why? Every outcome feeds back to the top and shapes the next iteration.

The part most sales tooling misses is the last one. The loop has to close without a human in the middle of every cycle. The human supervises at the edges. The loop runs on its own most of the time, and improves itself while the team is asleep.

The flagship loop: SDR → BDM → close-rate self-improvement

This is the loop I'd start with, and it's the one I'm planning to test next.

The SDR-to-BDM handoff is the single highest-leverage moment in most B2B sales orgs. SDRs generate opportunities. Some of those opportunities convert into closed-won deals; most don't. The standard explanation, "the SDR didn't qualify well enough" or "the BDM didn't work it hard enough," is almost always too coarse to act on.

What if you ran the handoff as a self-improving loop?

Sensor. Every SDR-sourced opportunity, with the qualification call recording, the discovery notes, the prospect's stated pain, the SDR's MEDDIC scoring, the company firmographics, and the eventual close-rate outcome.

Policy. The agent can update SDR playbooks, propose new qualification questions, rewrite the ICP definition, and flag SDRs whose opportunities consistently underperform. It cannot fire anyone or change comp.

Tools. Query the CRM. Pull the call transcript. Pull the BDM's discovery notes. Score the alignment between SDR-claimed pain and BDM-discovered pain.

Quality gate. A weekly human review of any proposed playbook changes before they ship. Sample-check 10% of autonomous CRM updates.

Learning. For every closed deal, the agent runs a post-mortem: which signals at the SDR stage actually predicted the win? For every loss, the same question in reverse: what did the SDR claim that turned out to be wrong? Over time, the agent updates the SDR's qualification framework, not the high-level MEDDIC template the entire industry copies, but the specific version that works for this product, this market, this BDM team.

The compounding effect is the interesting part. After a month, every SDR is qualifying with a sharper framework than the one they had a month ago. After three months, the framework is custom-fitted to your actual win patterns. After a year, the SDR team's productivity has moved less because of any individual rep and more because the playbook itself has become smarter every night.

This is the loop I want to run. I'm starting in the next few weeks, and I'll write up what worked and what didn't.

Other sales loops worth running

Once you see the pattern, you see it everywhere.

Discovery quality loop. Which discovery questions correlate with deals that close vs. deals that stall? Update the discovery framework weekly.

Objection handling loop. Which responses to "you're too expensive" lead to closed-won outcomes? Which kill the deal? Build a living objection library that reorders itself based on what actually works.

Outbound sequence loop. A/B test sequences continuously. Pick the winner. Replace the loser. Repeat. No quarterly sequence reviews; the system runs the experiment for you.

ICP loop. Every closed-won and closed-lost deal feeds the ICP definition. Companies that look like your actual best customers, not the ones in your founder's head, get more outbound attention.

Deal-risk loop. An agent watches every open deal for early warning signs (champion went quiet, no proposal opened, calendar slip on the next meeting) and surfaces the at-risk list every morning, with suggested recovery moves.

None of these are theoretically novel. Sales leaders have talked about all of them for years. The thing that's new is that you can now run them without a full-time analyst, without a project, without a quarterly review cycle. The loop runs itself.

What this means for sales leaders

A few implications I think are non-negotiable.

Burn tokens, not SDR headcount. Blomfield's talk noted that YC companies are now reaching demo day with roughly 5x the revenue per employee they had eighteen months ago. CNBC reported the broader shift in March 2025: YC's current crop is the fastest-growing and most profitable in fund history, with companies hitting $10M in revenue inside twelve months on teams of fewer than ten people, a quarter of them having 95% of their code written by AI. The binding constraint for sales orgs is about to shift from how many SDRs you can hire to how many tokens you can spend. Hire fewer, equip them with more compute, and measure who in your org is exploring the frontier of what's possible.

Every outcome should belong to a named person. Not a committee. Not "the SDR team." A single human directly responsible for each loop, each experiment, each metric. The org chart flattens. The accountability sharpens.

The CRO becomes a context engineer. The most valuable thing a sales leader can do in the next twelve months is make the organization legible to AI. The leader who does this fastest will compound.

How to make your sales org legible

This is the prerequisite for everything above, and most sales orgs are nowhere close.

Record everything. Every call. Every demo. Every discovery session. Every internal pipeline review. Every Teams message about a deal. If it wasn't recorded, it didn't happen, as far as the AI is concerned. Most sales orgs already have Gong or Chorus and don't realize they're sitting on the most valuable training data in the company.

Force CRM hygiene. Not because the CRM is a beautiful artifact, but because a deal whose state isn't written down is invisible to every loop. The rep who keeps deal context in their head is the rep whose pipeline can't be learned from.

Synthesize, don't dump. You can't feed 10,000 hours of call recordings into a model. You diarize them. You categorize them by deal stage, by objection type, by industry. You compress them into a living playbook, one that regenerates itself monthly from the new recordings, the same way YC just rebuilt its founder user manual from 2,000 hours of office hours.

What humans are for in this new shape

Think of the sales org as a brain. The data, the call recordings, the playbooks, the deal history, the qualification frameworks, that's the brain. The humans sit around the edges, interfacing with reality.

Humans go where the models can't yet, and where they probably won't for a long time:

• The high-stakes closing meeting where the customer wants to feel a person across the table.
• The executive sponsor relationship where trust is the currency.
• The dinner where the deal actually gets closed.
• The champion management call when the deal is wobbling and the champion is wavering.
• The negotiation when the customer asks for something the policy layer doesn't cover.
• The on-site visit. The conference floor. The kickoff.

That work stays human for the next twenty years. The institutional view is starting to confirm this. In August 2025, Gartner published a forecast that by 2030, 75% of B2B buyers will prefer sales experiences that prioritize human interaction over AI, a reversal of the digital-first narrative Gartner itself was pushing five years earlier.

The job of the sales leader is to make sure the humans are spending their time on exactly those moments, rather than on CRM hygiene, weekly forecasting spreadsheets, or coaching reps through the eighth iteration of the same objection-handling drill.

The replacement myth

Most sales leaders think AI is here to replace salespeople. I think that's the wrong fear. AI doesn't replace the people; it removes the people from the connective tissue between tasks, so the ones who remain spend all their time on the decisions that actually close deals.

The roles that disappear first aren't the closers. They're the coordinators, the routers, the summarizers, the people who exist because information needed a human to carry it from one task to the next. Those roles are the Roman legion's signal corps, and AI is a much better signal corps than humans have ever been.

The closer stays. The discovery rep stays. The executive sponsor stays.

The clearest evidence is already on the table. 11x.ai, the most heavily funded AI-SDR company in the category and backed by $74M from Andreessen Horowitz and Benchmark, claimed $14M in ARR. After trial conversions, the actual figure was closer to $3M. ZoomInfo, one of its largest reference customers, publicly stated that 11x "performed significantly worse than their SDR employees." Customer churn ran 70–80% within months. The category has not produced a single durable example of AI replacing the closer or the qualifier, only examples of AI replacing the coordinator between them.

What collapses is everything that sat in between them, slowing the loop down. The sales org doesn't get smaller because AI does the selling. It gets smaller because AI does the connecting, and the selling, the part that actually requires a human, becomes a larger share of every salesperson's week.

That's the bet under everything else in this piece. Get it right and you don't downsize your team. You concentrate it.

The bet I'm about to make

I'm starting the SDR-to-BDM self-improving loop on my own team in the next few weeks. I'll instrument the handoff, wire the post-mortem agent into the close-rate data, and let it propose playbook updates weekly. The control group is the current playbook. The treatment group is whatever the loop produces.

I'll report back on what worked, what broke, and where the loop drifted off-strategy, because I'd be lying if I said I expected it to work cleanly on the first try. The interesting question isn't whether the loop produces a better playbook on day one. It's whether it produces a better playbook every week, and whether that compounding effect actually shows up in the numbers.

If it does, the implication for sales leaders is straightforward: stop hiring your way to the next number. Start building loops.

If it doesn't, I'll tell you exactly where it failed.

Either way, this is the direction the field is moving. The sales orgs being built right now in this new shape will outproduce the legions. The only question for the rest of us is how quickly we're willing to rebuild.

---

Sources

• Tom Blomfield, How to Build a Self-Improving Company with AI, Y Combinator — YouTube
• Pete Koomen, AI Horseless Carriages — koomen.dev
• Erik Brynjolfsson, Danielle Li, Lindsey Raymond, Generative AI at Work, NBER Working Paper 31161 — nber.org
• Goldman Sachs, Gen AI: Too Much Spend, Too Little Benefit? — goldmansachs.com
• CNBC, Y Combinator startups are fastest growing, most profitable in fund history because of AI (March 2025) — cnbc.com
• Gartner, By 2030, 75% of B2B Buyers Will Prefer Sales Experiences That Prioritize Human Interaction Over AI (August 2025) — gartner.com
• The AI SDR Wars: 11x.ai competitive teardown — useanterion.com