Notes from the climb.

The honest version of how this started: Alliance let me go in March. Nothing personal, just arithmetic. Rate hikes were coming, and the investors who bought what we raised for, single-tenant net-lease real estate, were pulling back fast. I had spent a year raising capital for buildings, and the market for buildings blinked first. So instead of hunting for another seat, I built the first thing I had ever built for myself, with Gabe Newman, a technologist who could make hardware, networks, and render pipelines talk to each other. We called it Super Luminal, a high-performance edge-computing project. Gabe handled the systems. I handled the structure. The pitch, when I said it out loud in early 2022, made most people's eyes glaze: compute is becoming a commodity, and a commodity is a financeable, asset-backed thing.

The architecture was edge computing, processing close to the source of the data, where latency is the enemy: exchanges, machine vision, and the workloads I cared most about, training and running AI. The financial structure was the real idea. Debt was still historically cheap, and everyone could feel that window starting to close. You could finance racks of GPUs the way you'd finance a building, service the debt with the fixed income the hardware threw off, rendering, mining, leased cycles, and own an appreciating, cash-flowing asset underneath it all.

Compute was quietly becoming an asset class. I wanted to finance it like one.

Most people heard "crypto" and stopped listening. What I actually saw was the substrate. The same parallel hardware that mined a token today would train and serve the models that, I was convinced, were about to eat the world. Mining and rendering were just the yield that paid for the position while we waited for the real demand to arrive.

I want this on the record at the very bottom of the mountain, because everything above it grows from here: I was betting on the raw material of AI months before ChatGPT made it an obvious thing to bet on. The form of the bet changed many times after this. The thesis never did.

It took about three months to hit the wall. The thesis was right. The engineering was sound. We had the talent, the architecture, and a model that penciled. What we could not get was the one thing the entire structure depended on: the capital.

The irony was not lost on me. I had just spent a year raising capital for other people. I knew the rooms, the instruments, the language. It did not matter. Just about every notable alternatives investor passed, because the market simply did not know how to underwrite a rack of GPUs as collateral. And underneath that sat the deeper miss: nobody fully understood the value of inference yet, the idea that this hardware would not just train models but run the day-to-day thinking of every business that used them. I will be honest: we did not fully understand it either. We could model the rendering and mining yield to the dollar. The part that turned out to matter most, we felt more than we could prove.

The full story has a twist I still think about. One institution did lean in: HSBC took a genuine interest in the play. And we turned it down. A core piece of the roadmap was providing edge services to government agencies, and walking into those rooms carrying money from a bank that deep in Beijing's orbit was a contradiction we could not structure around. We said no to our only yes. It was the right call, and it still stung.

I had the right asset and the right thesis. What I did not have was a market ready to underwrite either.

It was a hard, formative failure. Being able to raise is not the same as a market being ready to fund. When the collateral is new, you have to build the lender a bridge, the structure, the data, the story, and we ran out of runway before the bridge was finished. Access to capital is not a reward for being right. It is the precondition for getting to find out whether you are right at all.

So I let it go and started again. That same month I set up Green Valley Partners. The first idea was modest and familiar: acquire industrial outdoor storage sites, the kind of unglamorous real estate I knew how to underwrite in my sleep. It did not stay that narrow for long. The firm quickly became the flexible vehicle for everything that came after. And everything I have done since, raising and syndicating to acquire companies, structuring debt and equity, comes straight out of this failure. Super Luminal did not summit. It taught me which bridges need building between new assets and old money, and I have been building them ever since.

Here is what the buildout actually looked like, set against the 480-GPU cluster in our deck: three GPUs in a Miami apartment. Then twelve. The model was still beautiful on paper, buy the cards, point them at the highest-yielding work, RNDR on the render network, mining elsewhere, and watch fixed income roll off an appreciating asset. Demand was so hot you couldn't find a high-end card. The reality was a living room doing double duty as a data center, a dozen cards turning a Miami apartment into a sauna the air conditioning never beat.

Even at that scale, the pro-forma lessons landed. The grace period is real: the outlay comes months before the cash does. Supply slipped, power was its own line item, and every assumption sat exposed to a rate cycle we had already written into our own risk column. A spreadsheet scales frictionlessly. Twelve hot cards in an apartment do not.

A spreadsheet earns its return instantly. Real infrastructure makes you bleed through a buildout first.

By the end of that summer we were honest with ourselves. Without institutional capital there was no path from twelve GPUs to the cluster the thesis needed, and we were not going to apartment-mine our way there. We unplugged, in every sense, and started looking forward. That clear-eyed letting go was its own education: the elegance of a thesis is worth nothing without the capital to execute its middle, and knowing when to stop climbing a route is part of mountaineering too. The energy went somewhere better. It went into Green Valley Partners.

2022 came for the thesis from two directions, and only one of them surprised us. The rate hikes we saw coming. We had written them into our own risk column from day one, and locking long, fixed-rate term debt was half the reason the raise had mattered so much. What we did not see coming was September: Ethereum's move to proof-of-stake, the Merge, ended GPU mining on the network overnight. By then we were already standing down, but the revenue stream at the center of our old pro-forma evaporated all the same, for everyone still in the trade.

That hurts in a way a spreadsheet can't prepare you for. But it clarified something I'd half-believed and now knew in my bones: mining was never the point. It was a yield bridge to keep the asset fed. The durable demand for this hardware was always going to be compute for artificial intelligence.

Mining was the bridge. AI was the destination, we just got there ahead of the traffic.

The deeper lesson was about asset selection under uncertainty. An asset can survive the death of its first use case if it has a bigger one coming. A pile of GPUs whose mining revenue just vanished is a disaster, unless the world is about to need every FLOP it can find to train models. I believed it was. And there is a humbling corollary I can only admit from a distance: every investor who told us no that spring accidentally saved us from being levered into this exact moment. The structure was already gone; the conviction survived. Conviction, it turns out, is the thing you carry to the next climb when the current one collapses under you.

On the last day of November, a chatbot went live and the world changed its mind about AI in roughly a week. Here is the honest part: I had been following OpenAI for years as a startup, the papers, the API, GPT-3. I still did not grasp what a productized LLM would do. My wow moment was not code or a deal memo. I asked it to explain Hesiod to me, the Theogony and Works and Days, as if I were five years old. And it did, instantly and charmingly, ancient Greek cosmology turned into a bedtime story. A machine that had read everything and could meet anyone exactly where they stood. That caught me flat-footed too. ChatGPT did what months of my explaining could not: it made the demand for compute obvious to everyone at once.

I felt two things at the same time, and I'll be honest about both. The first was vindication, every argument I'd made about compute being the substrate of an AI wave stopped being a pitch and became the consensus, almost overnight, with GPU and data-center economics suddenly repriced to match. The second was the particular ache of being right and early. Super Luminal's structure had taken its hits in the rate shock and the Merge; I wasn't positioned to ride the exact wave I'd been forecasting.

I'd spent the year arguing compute was the substrate of AI. In a week, the whole world agreed.

That combination, right thesis, wrong vehicle, reshaped what I wanted to do next. I'd been investing in the picks and shovels of an AI boom before the boom. The boom arrived. And I realized the more interesting, more durable game wasn't owning the compute; it was operating the companies that would turn cheap, abundant intelligence into products, revenue, and outcomes. That's the pivot. Everything from here up the mountain is me playing that hand instead.

2023 was the year compute became a household anxiety. A single chipmaker became one of the most valuable companies on earth, every board wanted an AI story, and the scramble for GPUs looked exactly like the demand curve I'd drawn on a slide a year earlier. But the shift that actually changed my trajectory was smaller and more personal.

I started using these models to augment my own output at Green Valley Partners. First drafts of memos and market maps. Diligence summaries. The first pass of a model's assumptions page. The connective tissue of advisory work, the part that quietly eats your week, started taking an hour instead of a day. I treated prompting like a craft and my own job like the test bench, and the results were not subtle.

The fastest way to understand a technology's business impact is to point it at your own job first.

Doing that taught me the Super Luminal lesson from the other side. When an input commoditizes, the value moves to whoever turns it into an outcome, and for a year that outcome was mine: more deals analyzed, more structures drafted, more conversations walked into prepared. I was my own proof of concept.

It also sharpened the appetite. Augmenting one advisor was interesting. What I wanted to know was what this does to a whole revenue engine, a sales team, a company. I went looking for a place to find out first hand. That place found me a few months later, and it came with a founder named Alex.

I met Alex through Unpopular VC, Sergii Zhuk's shop, one of those introductions that looks small in the moment and turns out to be a hinge. He is the Ukrainian founder of EVE, a voice-AI company he had built and proven in Europe years before the current wave, patented, with hundreds of millions of calls behind it. The war forced him out, and he came to the United States to relaunch the company from something close to scratch. I took the job of Chief Revenue Officer and, in practice, the entire go-to-market and sales function.

The thesis was clean and, by late 2023, finally credible. The models were good enough that a voice agent could carry the high-volume, repetitive phone conversations a business can never quite staff: the missed call, the after-hours lead, the reminder, the follow-up. Years of betting that AI would find real work to do, and here was a product doing it, on the phone, that day.

The mountain in front of me was never the technology. It was trust and distribution, in a market that had never heard our name.

The honest version of the beginning is that it was slow and cold. No US logos, no pipeline, a founder rebuilding his life and his company at the same time, and a product that worked beautifully selling into a market with no idea we existed. I had spent the prior stretch on capital and advisory work, learning how money and businesses actually get assembled. This was the chance to sit fully in the operator's chair and build a revenue engine from nothing.

Meeting Alex did more than hand me a title. It locked in the focus I had been circling all year: learning the business impact of AI first hand, from inside a quota, not from a slide. I believed the thesis. I also knew, from Super Luminal, that believing a thesis is the easy part. The climb is whether you keep moving when nothing has moved yet. So I started dialing.

Two models shipped this spring that changed the tenor of every operating conversation I'm in. GPT-4o made multimodal feel casual; Claude 3.5 Sonnet quietly out-ran models twice its supposed weight. Suddenly every founder I work with wants to know the same thing: how much of my company can this run?

My answer, for now, is: none of it, and a meaningful slice of every person in it. This generation is a copilot, not a colleague. It collapses the distance between a question and a competent first draft, which sounds modest until you count how much of knowledge work is first drafts. At EVE, the voice-AI company where I was running go-to-market, we put an assistant in front of the revenue team for proposal drafting, call summarization, and CRM hygiene. The wins were immediate and unglamorous.

The model is brilliant at the first 80% and dangerous in the last 20%, exactly inverted from where accountability lives.

So the rule we wrote down early: AI drafts, a named human signs. Anything that touches a customer, a number, or a contract gets a person's name on it. That single line is what let us move fast without moving recklessly.

The investor in me is taking notes too. If a year of capability gain turns every analyst into a team of analysts, the companies that win won't be the ones with the fanciest model, they'll be the ones whose workflows were clean enough to absorb the leverage. Most aren't. That gap is the opportunity.

Two years ago I bet that abundant, cheap compute would find a killer application. Watching a revenue team get its afternoons back, I realized this was that application arriving, pointed, finally, straight at the work itself.

I spent August looking at AI pilots across companies I touch, and the pattern was depressingly consistent: lots of motion, almost no margin. Everyone had a pilot. Almost no one could tell me what line of the P&L it moved.

This is efficiency theater, AI adopted to be seen adopting it. A chatbot bolted to a help center. A "co-pilot" nobody opens twice. A pilot that exists so the board slide has a logo on it. It feels like progress and costs like progress, but it doesn't compound.

So we ran a cull. Of four AI initiatives at one company, we killed three. The one we kept was the least exciting: automating the assembly of renewal quotes, a task that ate a senior rep two days a month. We tied it to a number, hours returned, cycle time, win rate on renewals, and we watched the number.

If you can't name the P&L line it moves, it's not a strategy. It's a souvenir.

The discipline here isn't technical, it's financial. I came up structuring deals and standing up operations after the close; you learn fast that initiatives without an owner and a metric die quietly and expensively. AI doesn't change that law. It just tempts you to forget it, because the demos are so good.

The companies that will look smart in two years are the ones being unglamorous right now, picking one workflow, instrumenting it, and refusing to celebrate until the number moves.

For a stretch of at least four months I was in front of five to ten CEOs of small and mid-sized businesses every single day. Demos, discovery calls, listening. It was the hardest and most clarifying work of my career, and it is the closest thing to a superpower I have found in go-to-market: talk to the market, in volume, until the market teaches you its own language.

The lesson came fast. The product was never the pitch. Their problem was the pitch. Every call was a short tutorial in where a business actually bleeds: the calls that go unanswered, the leads that arrive at midnight, the collections nobody has time to chase, the front desk that cannot scale. Our voice solution was only interesting in the exact shape of someone's specific pain, so I stopped demoing features and started naming problems.

The call I will never forget was a remediation contractor in South Carolina who rebuilds after hurricanes. Over a million dollars of his revenue sat trapped in insurance adjusters' queues, and his fix was brute force: a team of ten people whose entire job was to call the carriers, wait on hold, and ask for a status update. Every single day. We built him a bot that sounded human, dialed, sat on hold for hours without complaint, and came back with the update. Ten salaries of hold music, automated. Nobody buys voice AI. They buy their million dollars back.

Some wins came with an asterisk that taught me more than the win. We built and sold KFC's Ukrainian business a system that used our voice engine to monitor how politely cashiers treated customers. It worked, it shipped, it got paid for. When we tried to bring it to the United States, privacy standards stopped it cold. I logged that one carefully: in this category, what the technology can do and what a market will let you do are two different products. It would not be the last time that lesson showed up.

The product was never the pitch. Their problem was the pitch.

Do that five to ten times a day and something compounds. I could predict objections before they arrived. We sharpened the targeting, rewrote the script, and cut the demo to the two minutes that mattered. The pipeline went from a trickle to genuinely large, a real book of small and mid-sized businesses that needed labor leverage and could not hire their way out of it.

This was also the year the broader market woke up to agents and automation. I was not theorizing about it on a panel. I was selling the actual thing, automated conversations, into companies desperate for exactly that, learning in real time which promises the technology could keep and which it could not.

Last week a model used a computer. Not metaphorically, it moved a cursor, clicked, filled a form, navigated software built for humans. The capability is rough and slow and I wouldn't trust it with anything that matters yet. But I sat up, because the unit changed.

Until now, AI automated text: it gave you words you still had to act on. The moment a model can operate the tools your team operates, it can automate tasks, the click-through-five-systems-to-close-a-ticket kind of work that no API ever bothered to expose. That's most of what actually happens inside a company.

"Text" automation makes a person faster. "Task" automation changes how many people you need.

For the first time I could draw a credible line from copilot to genuine headcount leverage, not this quarter, but on a horizon a CFO should already be modeling. I started asking each operating team a new question: which of your workflows is just a human ferrying data between screens? Because that work is now on the clock.

It also reframes how I underwrite. A SaaS tool whose entire value was "we have an integration" should be nervous; if a model can drive the UI, the integration moat gets shallow. I'd rather own the system of record than the connector between systems. The pitch deck of the future has a slide titled "what happens when the agent can just do this itself." Most don't. They should.

Two things happened this fall that the headlines under-rated. Reasoning models started "thinking" before answering, and Anthropic open-sourced a protocol, MCP, for connecting models to the systems where work actually lives. The first got the press. The second will matter more to operators.

Here's what a year of deploying this stuff taught me: the binding constraint was never raw intelligence. It was context. A genius that can't see your data, your tickets, your pipeline, your contracts is just an eloquent stranger. Most failed AI projects I've seen didn't fail on the model. They failed on the plumbing.

The highest-ROI "AI work" we did all year was cleaning our revenue data. None of it was AI.

At one company we spent a quarter de-duping accounts, fixing ownership, and giving every record a clean lineage before we pointed anything intelligent at it. Unsexy. It was also the unlock, once the context was trustworthy, the same model that had been useless became dependable. A standard like MCP matters because it turns that plumbing from a custom project into a fitting you can buy. It standardizes the trailhead so the climb is about altitude, not about re-paving the path each time.

If you're an operator wondering where to spend in the new year: spend on the boring substrate. Clean data and clean integrations are the capital expenditure of the AI era, and they appreciate.

Somewhere between the CEO calls, a different door opened sideways. Back in the summer I took what was supposed to be a tidy consulting engagement: help Synapse Virtual Production stand up its first studio, a flagship headquarters in Los Angeles. Virtual production, LED volumes, real-time engines, final pixels in camera, could not be further from voice AI. That was the appeal.

The launch outran the script. Work with Netflix and Disney landed almost immediately and drove a large part of the initial runway. Ben Affleck's Artists Equity got involved, and the stage shot two of their films, The Ruffian and The Accountant 3. Eminem's Houdini video came through the volume and ended up Grammy nominated. In an industry that runs on proof, the studio built a highlight reel before it built a long payroll.

The same toolkit travels: structure the capital, sharpen the story, build the engine that sells it.

And the consulting piece refused to stay small. It grew into a real finance and strategy seat, shaping the growth story and the financing plan behind it, and working shoulder to shoulder with the sales team on the go-to-market motion, how a premium services business scales efficiently without burning its margin. Different industry, same muscles.

People ask how the seats fit together. Honestly, the same way they always have. Voice AI was teaching me what selling a brand-new category feels like. Synapse was teaching me what scaling a premium services brand feels like. Both kept feeding the conviction this journal circles again and again: technology compresses the cost of making things, and the value moves to whoever turns that compression into outcomes, on a soundstage as much as in a sales pipeline.

In January an open-weight model from DeepSeek reached the neighborhood of the frontier at a fraction of the cost, and a general-purpose "operator" agent shipped that would actually go do things on the web. Markets wobbled. The takeaway I keep repeating to founders: capability is commoditizing, and it's getting cheap faster than anyone budgeted for.

That's not bad news. It's just clarifying. If frontier-grade intelligence is becoming a utility, then your durable advantage was never going to be the model. It's the things models can't copy: proprietary data, distribution you already own, regulatory trust, and the unglamorous excellence of execution.

When the engine gets cheap, the moat moves to everything around the engine.

I was running go-to-market at a voice-AI company at the time, and the cost collapse cut both ways. It made our own product cheaper to run, and it made the giants able to fold the same capability into their platforms for next to nothing. The constraint flipped from can the model do it to can a startup still own it.

As an investor I got more skeptical of thin wrappers and more interested in businesses with a real data asset and a distribution edge. If your only advantage is access to a model, you have the same advantage as everyone with a credit card. The companies I want to own are the ones where cheaper intelligence makes their existing moat deeper, not the ones whose moat was the intelligence.

By early 2025 two things were undeniable, and together they made the decision for me. The first was consolidation. The capabilities we had worked so hard to build were being commoditized in real time by the largest players in the field, OpenAI and ElevenLabs among them. The same voice quality that was a startup's edge in 2023 was becoming a checkbox feature inside someone else's platform. I had written, a month earlier, that the moat is never the model. Now I was living the other side of that sentence.

The second was trust, and it was the harder wall. AI voice calling sits squarely against a real regulatory and cultural line. TCPA and FTC rules around automated and AI-driven calling are not a technicality, they reflect a genuine public discomfort with a machine on the other end of the phone. We could engineer the product. We could not, by ourselves, engineer away the credibility tax that every AI call started with.

And beneath both sat the operating math nobody puts in a demo. The industry we sold into, outbound calling, was getting hammered by the carriers and the FTC at the same time; cold outreach simply is not the channel it once was. Compelling voice agents were costly to stand up, every deployment meant weeks of setup and testing, and conversion came in lower than the demos promised. I ran the unit economics over and over, hoping they would say something different. They did not.

A good operator does not keep climbing a face that is crumbling underneath them.

So I made the call to leave. Not because the work failed, we had built real pipeline and real revenue, but because the durable business was not there for us, not against that consolidation and that trust barrier at the same time. Conviction is knowing when to climb. Judgment is knowing which mountain.

What voice taught me pointed straight at what came next. The constraint was never capability, it was distribution and trust. Mobile engagement and broadcast, SMS, MMS, and RCS, were channels people already trusted, with consent and compliance built into the rails. That was the better mountain: the same operating problem with the trust question already answered.

The door opened the way doors actually open: through people. Frederik Roikjer, a former Morgan Stanley investment banker who became my business partner, knew Jed Alpert, the founder who built Mobile Commons in the first place. Jed was working on buying his company back out of Upland Software. A founder reclaiming a trusted, compliant messaging platform, the exact inverse of the trust problem I was walking away from. We all jumped on board.

We put our first real agent into production this month. Not a chatbot, an agent: it reads inbound messages, classifies intent, pulls the relevant account context, drafts a response, and routes or sends it under a human's supervision. It's been live for three weeks. It has been humbling and thrilling in roughly equal measure.

The thrill is obvious. The humbling part is what I want on the record: the model was the easy part. What consumed us was everything around the autonomy, permissions, logging, the line between what it can send and what it must escalate, and the evaluation harness that tells us when it's drifting. We spent more time on guardrails than on prompts, and that was correct.

You don't measure an agent in tokens. You measure it in tasks closed, and in mistakes caught before they reached a customer.

The mental shift for the team was the hard currency. We stopped asking "what can the AI tell us" and started asking "what can the AI do, and who is accountable when it does." That second clause is the whole discipline. An agent without a clear owner and a kill switch isn't innovation; it's unmanaged risk wearing a nice demo.

Where this points: the org of the near future has humans setting intent and reviewing exceptions, with agents handling the long flat middle of a process. We're not there. But for the first time, standing at this altitude, I can see the ridge line.

The surveys this summer tell a funny story if you read them honestly: a large majority of companies say they've "adopted AI agents," while only a sliver have one genuinely running a process in production. I believe both numbers. I've sat in both rooms.

The gap between an impressive pilot and a deployed system is the entire game, and almost no one prices it correctly. The hard part isn't intelligence anymore, it's secure, reliable access to the systems where real work happens, plus the monitoring to trust it unattended. That's integration work, change management, and permissions. It's a slog. It's also the moat, precisely because it's a slog.

The demo earns the meeting. The integration earns the margin.

We went deliberately hybrid: off-the-shelf agents where the problem is generic, custom builds where the workflow is proprietary and the data is ours. That blend lets us move quickly without handing the crown jewels to a tool we don't control. The instinct to build everything is ego; the instinct to buy everything is laziness. The operator's job is to know which is which, function by function.

My advice to the founders I back has gotten blunter: stop counting AI features and start counting processes that now run with a human only on the exceptions. One genuinely deployed agent beats ten pilots in a slide. The market is about to start telling the difference, and so are the people writing checks.

August's marquee model launch landed with a thud relative to the hype, a rocky rollout, a demo with mislabeled charts, expectations the product couldn't meet on day one. The discourse treated it as a referendum on whether AI was slowing down. I think that's the wrong frame entirely.

What actually happened is that progress moved from spectacle to plumbing. The gains that matter now aren't a jaw-dropping leap on a benchmark; they're the steady, boring accumulation of reliability, tooling, and integration that turns a capable model into a dependable system. That work doesn't trend. It compounds.

We stopped chasing model releases and started compounding workflow wins. Our best quarter for AI-driven margin came without changing the model once.

For operators this is liberating. You no longer have to refactor your roadmap every time a lab ships. Pick the workflows, deploy, instrument, and let the underlying model improve under you for free. The frontier raising the floor is a tailwind you don't have to chase.

For investors it's a sharper filter. I'd rather back a team with one agent in production and a credible margin story than a team with a dazzling demo and a thesis that depends on the next model being magic. The hype cycle is sorting operators from tourists. That sorting is where returns get made, on the way down from a peak of expectations, not on the way up.

The frontier labs spent the back half of the year shipping models that were, frankly, hard to tell apart on a spec sheet, all very good, all reliable enough to trust with real work. When the technology stops being the bottleneck, the bottleneck becomes you: your processes, your incentives, your org chart.

So this fall we stopped bolting agents onto roles designed for humans and started redesigning the work first. You don't get the gain by handing a person an agent and hoping. You get it by asking what the process would look like if it had been designed for humans and agents from the start, then staffing that design.

An agent dropped into a human-shaped org gives you a faster version of your old bottleneck. Redesign the process and you remove it.

New roles appeared that didn't exist eighteen months ago. Someone owns the evaluation harness, the tests that decide whether an agent is fit to run. Someone supervises a fleet of agents the way a floor manager once supervised a shift. Old roles changed shape: our reps spend less time doing and more time deciding, handling the exceptions the agents escalate. Headcount didn't simply fall; it recomposed, toward judgment and away from rote.

This is the part of the climb most companies will get wrong, because it's organizational, not technical, and the technical part is what everyone trained for. The advantage now belongs to operators willing to redraw the chart, and that's a discipline you can't download.

At Mobile Commons the question stopped being whether AI helps the sales team and became which parts of each role it should carry. The answer, in practice, is Claude. Not a chatbot bolted onto the side of the org, but a working part of the go-to-market function with what amounts to a job description: enablement, demo production, and the material that moves a customer conversation from one stage to the next.

Start with enablement, because it is the least glamorous and the most transformed. Battlecards, vertical one-pagers, objection libraries, talk tracks tuned for healthcare versus government versus nonprofit, the content that used to take a quarter to produce and went stale before anyone used it. That now gets drafted, refreshed, and tailored in days. Every rep walks into a conversation with material built for that prospect's industry, use case, and compliance reality, not a generic deck with the logo swapped.

The bigger unlock is live demos and presentation material built mid-cycle and tailored to the exact customer in the room. Everything we know feeds it: the full account history if they are a returning customer, every scrap of discovery if they are new. A prospect describes their workflow on a Tuesday call. By the follow-up they are looking at their own use case running: their segments, their message flows, their numbers. What used to be possible only for a strategic account now happens for every serious conversation.

The new sales motion is show, not tell. The asset that closes is the one built for exactly one customer.

What this does to the roles matters more than the tooling. It does not remove the rep, it promotes the rep. Less time assembling decks and chasing collateral, more time in actual conversations. The job tilts toward discovery, judgment, and relationships, the parts a model cannot carry, while Claude handles the connective tissue between conversations. This is the org redesign I wrote about in December, playing out one function at a time.

The investor takeaway is the same one I keep arriving at from different trails: point AI at the material between conversations and deals move measurably faster. Pipeline velocity is a compounding number. Augment the people who create it and the whole engine climbs.