AI Conference HumanX Takeaways and Frameworks

6,714 AI leaders from all over the world simply gathered for a large AI convention in San Francisco. There have been individuals from 79 nations and 500+ periods.

From over 35,000 minutes of content material on the occasion, these are the 9 takeaways to know. There are additionally frameworks and psychological fashions after the takeaways – one you completely must know is the three waves of AI by Jensen Huang.

Each gross sales chief is aware of the story: your finest reps carry the quantity whereas everybody else tries to get fortunate. HockeyStack closes the hole with Income Brokers for the enterprise.

Each deal will get its personal devoted agent, surfacing the best play based mostly on how your organization really wins. Reps deal with judgment, relationships, and the conversations that shut offers. Brokers deal with all the pieces else.

That’s why groups like RingCentral, Outreach, ActiveCampaign, and Fortune 100 corporations depend on HockeyStack. Study extra at hockeystack.com

Outbound gross sales at this time continues to be constructed round guide work: writing emails, constructing lists, and managing sequences throughout disconnected instruments.

Nooks is breaking the mildew. With their new AI Sequencing product, Nooks is an agent workspace the place AI works alongside reps, serving to them perceive accounts, prioritize the best prospects, and generate context-rich outreach based mostly on precise first-party interactions.

No extra losing time in a number of instruments – it’s one unified workspace for all of your outbound channels. Arm your gross sales workforce with clever outbound. Study extra at https://www.nooks.ai/gtmfund

Takeaways

1. Spatial intelligence is the subsequent frontier.

Dr. Fei-Fei Li (World Labs founder, Stanford AI Institute) delivered a thesis that can outline the subsequent 5 years of AI funding: all the pieces we’ve constructed to this point is “wordsmiths in the dead of night.” Language fashions can generate textual content brilliantly. However they’ve zero understanding of the 3D world, physics, motion, or causality. They’re solely summary.

Spatial intelligence – the power to understand, purpose, and generate in precise 3D house – is the lacking layer separating AI that may do actual work from AI that may solely theorize about work. A language mannequin can’t train a robotic to maneuver, a textual content generator can’t construct a dynamic recreation world, and an summary medical algorithm can’t diagnose from 3D scans with out spatial reasoning. You may’t construct intelligence on language alone.

This creates an information constraint. Spatial coaching information is exponentially scarcer than language information. Which implies the businesses that crack the artificial information flywheel first – constructing fashions that generate practical 3D worlds that then prepare robots and different fashions, creating new coaching information – will personal the subsequent decade the best way basis fashions owned this one.

In case your software doesn’t work together with the bodily world, language fashions are most likely enough. If it does (robotics, provide chain, manufacturing, design, healthcare), you want spatial grounding. You want companions constructing spatial intelligence. The distributors dominating spatial intelligence within the subsequent 5 years would be the infrastructure layer of the 2030s.

2. There are 3 waves: Generative → Reasoning → Agentic.

Jensen Huang (NVIDIA) mapped the precise trajectory of AI that most individuals misunderstand.

Wave 1 (2023-2024) was generative: are you able to make textual content, photos, code?
Wave 2 (2024-2025) is reasoning: are you able to suppose by means of issues, situation outputs, do analysis?
Wave 3 (2025-2026+) is agentic: can the mannequin take a enterprise downside in plain English and execute end-to-end throughout your precise programs?

The sample is within the prompts.

12 months 1: “What’s?” “The place is?” “How does?” (extraction).
12 months 2: “Cause about this.” “Summarize this.” (considering).
12 months 3: “Construct this for me.” “Execute this.” “Determine it out and report again.” (autonomous work).

His prediction flips the org chart: inside 1-2 years, a non-technical CEO will have the ability to describe a enterprise downside and have an agent execute the answer. Not draft an e-mail – really combine programs, learn documentation, iterate, resolve. You don’t want engineers to orchestrate anymore, you want an individual who understands the result and may articulate it to the agent.

This collapses all the talent ladder. Historically, engineers have been the bottleneck. They join programs, pull information, construct workflows. In 12-18 months, somebody describes the result and brokers construct the workflow. This can be a categorical shift.

You’re shouldn’t be ready for brokers to be “prepared,” as a result of they’re already prepared for 70% of your workflows. You’re betting on orchestration and adoption, not expertise. The groups that transfer quickest now – that begin orchestrating brokers into their core workflows within the subsequent 3 months – will set the sample for the way their complete business operates. Everybody else might be copying.

3. The talent turning into much less priceless is execution, and the talent turning into priceless is style.

Inside OpenAI, Srinivas Narayanan’s engineers don’t write code anymore. They information brokers that write code. 80% % of their time is spent on judgment – understanding what the AI did improper, what to attempt subsequent, whether or not the output is correct. A full repository is generated by CodeX, however the human decides if it’s proper.

Open Proof’s 70 engineers constructed what would have taken 500 as a result of the execution layer collapsed.

What’s left? Style.

This inverts your hiring and promotion calculus. You may’t rent for “quick executor” anymore. That’s a commodity. You want individuals with sturdy opinions in regards to the area, the power to see flaws in AI output earlier than they ship, and judgment at pace.

In case you’re nonetheless optimizing for pace to execution, you’re hiring for yesterday’s bottleneck.

4. “Human within the lead, not within the loop” is how enterprises scale.

When enterprises say “human oversight,” they often imply “people continuously validating each determination.” That doesn’t scale. The aggressive benefit is human within the lead. That is the place people set imaginative and prescient, technique, what to measure. Then, brokers deal with execution at scale.

This may affect who you’re promoting to and the way. In case you have been focusing on IT or operations heads who look to manage variables, you could now be pitching enterprise leaders who wish to set route and transfer at enterprise pace. Totally different purchaser, totally different message, utterly totally different gross sales cycle. Be aware of this distinction.

5. The mannequin is free now. What’s really scarce is downside perception.

Open Proof shouldn’t exist. 70 staff, 35 engineers, $12+ billion valuation, 60%+ of US physicians utilizing it every day. If proudly owning the mannequin was the defensible asset, a startup with no customized fashions ought to be unattainable. However they don’t personal the mannequin, they personal the issue.

All the narrative round “mannequin superiority” has inverted. Mark Terbeek (Greycroft) and Hans Tung (Notable Capital) see the identical sample throughout winners: the businesses profitable aren’t those with the very best expertise, they’re those closest to the shopper downside. Those that constructed backward from “what do medical doctors really want” as a substitute of “what can our mannequin do?”

Pace kills all the pieces else now. When the mannequin improves weekly, delivery quick and staying near buyer issues beats characteristic parity each single time. A smaller workforce shifting at enterprise pace will at all times outrun an enterprise that should deploy modifications by means of committee.

This kills the normal moat-building playbook. You may’t win on expertise possession. Fashions are commodities bettering on a one-week cycle. What’s defensible is downside perception – the founders and workforce who perceive the area so deeply that they know what to construct earlier than clients can articulate it. And the organizational pace to ship it earlier than the mannequin catches up and makes it free.

6. Unit economics are what separate sturdy corporations from developments.

Salil Deshpande (solo GP managing $750M) named the disaster everybody’s avoiding: most AI infrastructure has damaging gross margins. Fashions enhance weekly, closed-source fashions keep 9 months forward. The mathematics doesn’t work. That is dot-com infrastructure yet again – huge overbuilding adopted by a collapse that wipes out corporations constructed on developments, not unit economics.

When you may’t maintain your CAC, enhance margins each quarter, or scale profitably, you’re not a enterprise. You’re a characteristic ready to be copied or a vendor ready to vanish.

For operators evaluating distributors: if a vendor has sub-70% gross margin, they’re not fascinated by sturdiness. They’re fascinated by development at any price. Which implies inside 18 months they might pivot, get acquired, shut down, or be compelled to lift at a down spherical and minimize assist. You’ll be caught with a software that turned orphaned. Ask about CAC payback and gross margin trajectory. In the event that they hedge or refuse to reply, that’s your sign.

7. Brokers don’t change your workforce, they change your processes.

Madhav Thattai (Salesforce, Agentforce) and Rob Seaman (Slack) uncovered a sample that kills the “AI will displace staff” narrative: Agentforce is at $800M ARR with 25,000 clients working billions of agent transactions. Firms aren’t shrinking, they’re shifting quicker and turning into extra bold.

The concern is misplaced. Brokers don’t change your workforce, they expose what your workforce ought to really be doing. When an agent handles the 80% of customer support that’s rote, the remaining 20% turns into seen: the circumstances requiring judgment, empathy, problem-solving. That human? They’re value extra now, not much less. Engineers don’t disappear when brokers write code. They change into the individuals who transient brokers on what to construct subsequent, assessment structure, make strategic calls.

8. The place brokers sit issues greater than what they do.

Slack hit a billion messages a day. 1,000% enhance in AI apps being constructed. And the brokers didn’t get smarter, they simply moved. That’s all the story!

The hole between an agent individuals use and an agent individuals ignore is placement. An agent on a separate web site? Adoption dies. That very same agent, invisible in Slack the place work really occurs? Adoption jumps 25% instantly.

That is the place most AI investments fail: they’re constructing good brokers after which burying them three clicks away in a separate software. Or worse, they’re anticipating gross sales groups to “uncover” the agent, log in to a brand new system, be taught a brand new interface. That’s by no means going to compete with the agent that simply seems in Slack when somebody says “I don’t know what to do.”

Rob Seaman defined how Slackbot turns into the invisible router: when somebody wants advantages information, Slackbot surfaces the advantages agent. When they should file a ticket, it routes to Linear. The agent by no means must be named or found, it simply emerges contextually from the circulate of labor.

For product builders: your agent’s functionality might be positive. Possibly it’s 70/10/20. But when it’s not within the circulate of labor – if it requires an additional step to entry – you’re betting towards human friction. People are lazy. They use what’s in entrance of them. Construct the agent, then architect the location.

9. Day 2 is more durable than day 1…observability is the actual battle.

Each founder thinks Day 1 is the exhausting half: construct the agent, get it reside, ship it. That’s when the actual work begins, and most groups aren’t ready for it.

Firms dash to construct an agent in 2-3 weeks, have fun the launch, after which uncover they don’t have any visibility into whether or not it’s really working. The agent is reside. Is it performing? Is it dealing with the best circumstances? Is it drifting? In healthcare and monetary providers, “largely works” is code for “about to create a lawsuit.”

The distinction between a high-growth firm and a stalled one is observability. Most distributors obsess over agent functionality. Those profitable obsess over measurement. Firms want to trace agentic work items (precise accomplished duties, not tokens), monitor KPI supply, and spot drift the second enterprise circumstances change however the agent nonetheless follows directions from three months in the past.

Salesforce itself realized this the exhausting method: constructing the agent took 1.5 months. Truly working it – refining, measuring, optimizing – took one other 2+ months. And it’s steady. When a KPI modifications, the agent will get coached and re-measured. When enterprise guidelines shift, the agent doesn’t auto-adapt. Somebody has to note and intervene.

The distributors pitching “construct brokers in hours” are promoting you the primary mile. The distributors speaking about observability, monitoring, and Day 2 operations are promoting you the opposite 99.

Frameworks & Psychological Fashions

These are the reusable patterns beneath the takeaways. Use these to consider your individual state of affairs.

The 5-Layer AI Stack (Jensen Huang, NVIDIA)

Energy → Chips → Infrastructure → Fashions → Purposes

Each layer has its personal ecosystem, margins, and aggressive dynamics. Most enterprise capital flows to functions, however functions are nugatory with out the layer beneath them. The constraint shifts over time. Proper now, energy and chip design are bottlenecks. In 18 months, it is perhaps fashions. In 3 years, it is perhaps functions.

A important perception: an important layer is functions. Not as a result of it’s the sexiest, however as a result of it’s the place worth accrues for purchasers. Chips are commodities until they permit new functions. Fashions are infrastructure until they unlock new work. Power is ineffective until it powers one thing individuals need.

When evaluating an organization or a vendor, ask: which layer are they really profitable in? Are they defending a layer they’re good at or pretending to compete all over the place?

The Three Waves of AI (Jensen Huang)

Wave 1: Generative (2023-2024)

What it does: Generate textual content, photos, code from language
Immediate sample: “What’s?” “The place is?” “How does?”
Consumer habits: I ask, AI solutions
Enterprise affect: New content material, new advertising, new coding velocity

Wave 2: Reasoning (2024-2025)

What it does: Cause by means of issues, do analysis, situation outputs
Immediate sample: “Summarize this.” “Cause about this.” “Why did X occur?”
Consumer habits: I ask advanced questions, AI causes aloud
Enterprise affect: Higher accuracy, grounded outputs, belief will increase

Wave 3: Agentic (2025-2026+)

What it does: Take enterprise issues in pure language and execute end-to-end
Immediate sample: “Create X for me.” “Construct Y.” “Execute this process and report.”
Consumer habits: I describe end result, AI does the work autonomously
Enterprise affect: Discontinuous productiveness, new classes of labor, talent compression

In case you’re nonetheless working in Wave 1 or Wave 2 considering, it is advisable to get to Wave 3.

Spatial Intelligence vs. Language Intelligence (Fei-Fei Li)

Language Intelligence (at this time)

What it understands: Textual content, patterns in textual content, the way to generate new textual content
What it’s blind to: 3D house, physics, motion, causality within the bodily world
Metaphor: “Wordsmiths in the dead of night”—good at language however not grounded in actuality

Spatial Intelligence (rising)

What it understands: 3D house, geometry, physics, dynamics, motion, interplay
What it permits: Robotics, autonomous programs, medical imaging, recreation worlds, design instruments
Output: Not textual content or photos, however 3D worlds, simulations, predictions of subsequent states

The convergence: Language fashions + spatial fashions collectively = AI that may purpose about work and execute it within the bodily world. Language alone is idea. Spatial alone is mechanics. Collectively, they’re intelligence.

The Maturity Curve: Automation → Discovery → Actual Work (Salesforce/Slack)

Each agent deployment follows this development:

Month 1: Job Automation

Agent handles easy, repetitive duties
Standing checks, e-mail drafting, coverage lookup
Consumer sees: “The system can do that for me”
Enterprise worth: Time financial savings on low-impact work

Month 2: Data Discovery

Agent can reply questions, synthesize info, discover context
Buyer historical past, coverage explanations, information lookups
Consumer sees: “The system understands my area”
Enterprise worth: Quicker decision-making, much less context-switching

Month 3+: Actual Work

Agent executes consequential duties: orders, service resolutions, transactions
Full workflow autonomy with guardrails
Consumer sees: “The system really strikes enterprise ahead”
Enterprise worth: Income affect, buyer expertise transformation

Don’t choose brokers on Month 1 functionality. All of them look mediocre. The businesses profitable are those which have the infrastructure and self-discipline to ship Month 3 work.

Agentic Work Items vs. Tokens (Salesforce)

Tokens = enter mechanism

What number of phrases/items of textual content does the mannequin eat?
Like measuring a truck by how a lot gas it burns, not how a lot it delivers

Agentic Work Items = output mechanism

What number of significant duties did the agent full?
Precise enterprise outcomes: orders processed, clients resolved, choices made

The error: optimizing for token effectivity as a substitute of labor unit affect. You may burn tokens perpetually with out delivering worth.

Three Constraints on Frontier Fashions (Fei-Fei Li)

For any frontier AI firm, three issues constrain development:

Compute: Entry to GPUs, information facilities, coaching infrastructure
Fashions: Analysis functionality to innovate new architectures
Information: Coaching information that teaches the mannequin what really issues

Most corporations obsess over compute and fashions. Information is the hidden constraint. Spatial information is even scarcer than language information. The information flywheel—the place output turns into enter for the subsequent era—is the place defensibility lives.

Two Methods to Assume About “Busy” (Jensen Huang)

Prescriptive: “Right here’s precisely what you need to do with AI”

Professionals: Clear, measurable, predictable
Cons: Misses a lot of the alternative, forces inflexible considering

Inspiring: “Right here’s why this issues. I can’t predict precisely the way you’ll use it, however that is the route.”

Professionals: Captures optionality, lets groups discover their very own breakthroughs
Cons: Tougher to measure, requires belief in distributed judgment

The very best leaders let a thousand flowers bloom all over the place besides the important core (coding, provide chain, chip design). In these domains, you go to the frontier and don’t fail.

HumanX

Following a landmark second version on the Moscone Heart in San Francisco, HumanX 2027 will happen from March 7–10, 2027, at Mandalay Bay in Las Vegas.

Tag @GTMnow so we will see your takeaways and assist amplify them.

Anthropic hit a $3B income run price and is doubling down on infrastructure, partnering with Google and Broadcom to construct customized AI chips. The guess: proudly owning silicon is the way you management cost-per-token at scale, and at this development price, that math issues extra each quarter.

Trimble is buying Doc Crunch, bringing AI-powered contract danger evaluation into its building challenge administration ecosystem. Building is likely one of the final industries to modernize round doc compliance, and purpose-built AI that catches cost disputes and notification failures earlier than they escalate is an actual wedge. Vertical AI discovering its method into incumbent platforms is the sample to look at.

GTM 185: How One Hackathon Took Zapier’s AI Utilization From 10% to 97% | CEO of Zapier

For the complete factor, hear on Apple, Spotify, YouTube or wherever you get your podcasts by looking out “The GTMnow Podcast.”

Armada – named amongst Quick Firm’s Most Revolutionary Firms in computing, alongside NVIDIA and Google. They’re constructing distributed AI infrastructure for the environments most platforms ignore. Value watching as enterprises and governments push AI into the sector, not simply the cloud. To be taught extra you may also take a look at the GTMnow episode with CEO Dan Wright.

Examen – raised $4.3M in whole funding and launched. Examen is constructing an autonomous analyst for industrial actual property.