What you missed at 1682
Ten talks and fireside chats plus three Venture Award pitches kept landing on one idea: anyone can build with AI now, and the hard part is everything around it. That means process, clean context, governance and human judgment, not which model or vendor you pick.
Mike Pitone's PB&J video, Ethan Mollick on “YOLO mode,” and Ophir Samson's AI interviewer refusing a live jailbreak.
HubSpot: up to 1,850% more AI-referred leads. Lovable: one company built 295 apps in 6 months and needed 9.
Gadsby's keynote framed the day. Dialed ($127 vs ~$35K/yr) got quoted back. Lazar named James “the Lazar of O3.”
BleedIO took the win with wireless mesh networks for physical AI.
Candid, funny, and very Philly
It felt less like a vendor pitch and more like “here's what we're figuring out.” Speakers owned their mistakes on stage.
“Nobody knows what's going on.”Ethan Mollick, summing up the mood
- Everyone admitted mistakes. ServiceNow said it moved too fast on employees. Guru burned its whole 2026 token budget in 3 months. Lazar's “holy grail” internal tool got zero users.
- Speakers talked to each other. Gadsby's $127-vs-$35K stat, his three-step model and his Hugging Face story were quoted back all day. “Every 11 days” (how often a new model ships) came up three times. Lazar opened by joking he had “planted a small recording device” to write his talk from everyone else's.
- Practical over hype. The big moments were real operator numbers: Buffalo Construction cut closeouts from 120 days to 35, West Shore installs bathrooms about 7 days after sale, and the Barnes cut its art-app costs 97%.
- Very Philly. Wawa in-jokes, a Shorty Hoagie dig from the emcee, a Sheetz jab (“Wawa's gonna go hack Sheets… don't put that out there”), an Eagles end-zone line, and Rick's Schuylkill River water analogy.
- Community first. Keith and Gadsby both said the audience could have replaced the speakers. Gadsby closed on his first-grade report card (“a vivid imagination”) and “find value in the relationship with one another.”
- The energy curve. Packed and buzzy all morning, peaking at the Mollick fireside after lunch. The room thinned a little by the last pitch, then Lazar and cocktails in Annenberg Court carried it home.
Six ideas that ran through nearly every session
The builder has changed
The people closest to a problem now build the fix themselves, with no IT team in the middle. O3's PMs ship Dialed updates mid-call, Buffalo Construction's president vibe-codes “toothpick” fixes, and Barnes staff work beside agents in a tool called Switchboard.
Process before prompts
AI exposes broken workflows instead of fixing them. Danny Fisher: AI doesn't need a use case, it needs a process.
Marjón Dean calls it “organizational debt.” Mollick says prompting classes are useless now; a good prompt is just a good SOP.
Context is the moat, not the model
Models are good enough for nearly anything, but they know nothing about your company. Rick Nucci: the real problem is accuracy. Not access.
Christina Clark applied the same logic to marketing: brands have to become the answer AI engines cite.
Let it sprawl, then consolidate
One Lovable customer gave every employee a license, built 295 apps in 6 months, and needed just 9. Experiment widely, then fold the winners into one governed “mothership” with a named owner. Rick: pick one workflow, … one outcome, … one squad.
Humans manage the loop
No one reads 20 hours of agent work line by line, so people shift to setting direction and owning outcomes. Mollick: I don't think you could be in the loop anymore… I think you could manage the loop.
Gadsby: We can't outsource governance.
Build vs. buy is back
Cheap building is pulling SaaS spend in-house. The Barnes cut its app's operating cost 97%, Dialed replaced a ~$35K/yr tool, and Lovable's internal sales tool saved about $1M. Rick's caveat: It's not the V1, it's the V4.
Only build what's truly yours.
The lines people will be repeating
Lightly trimmed from the session recordings. Filler words are cut and gaps marked with “…”.
If you're feeling imposter syndrome, that is exactly the right place to be. We are all making this up… Nobody knows what's going on.
If you only take one thing from this section, please know that AI doesn't need a use case, it needs a process.
We came into 2026 like token maxing… 3 months later, we're like, that was a horrible idea… We blew out our whole budget for the year in 3 months.
The main risk today is no longer the AI slop era… the bigger challenge today is AI sprawl.
The trust is the only currency in the change, period. Communicate some more and communicate again.
It feels a bit like 1980s dating videos, like, hi, I'm Taylor, I work at Blockbuster.
AI for all is a great idea, but AI for all and being responsible is a better idea.
So we're just all YOLOing everything. And that is the technical term, by the way… It actually is YOLO mode.
It's less about learning prompt design. And it's more about just learning to be descriptive with your context.
Your website is not dead, but it's not the growth model that it used to be because it's not one surface.
It's like needing a toothpick. It's not complicated and it's not fancy, but when you need one, you need one and you need it right now.
We think the durable problem in enterprise AI is accuracy. Not access.
AI is really exposing… organizational debt… Don't confuse the tool with the process that's broken.
My goal was to kind of kill the resume. I do not believe it is a useful signal any longer for who a candidate is.
Models will change, tools will change, people in this room will probably change next year. This formula won't.
In an autonomous supply chain, the pallet has to talk. So we made it intelligent.
The run of show, in one scroll
9:00 · Keith Scandone opened by thanking presenting sponsor Guru, Lovable, and the first-ever advisory committee. He also joked that “whoever organized this thing didn't give me a lot of time.”
Art, AI + new ways of seeing
Steven Brady · Barnes Foundation
The Barnes rebuilt Barnes Focus, its point-your-phone art-recognition app, with “agentic engineering.”
- It's 5x faster, and operating costs fell 97% after dropping an outside vendor.
- A review tool in progress has AI pick the best match and a human confirm it, taking each review from about 1 minute to 5 seconds.
- The thesis: adopt AI “for the purpose of humans.”
AI for all: the operational shift
Mike Gadsby · O3XO
The keynote that framed the day. Individuals already have big gains, so the next step is going “from prompt to process,” then “from process to product.”
- He walked through O3's AI Operating Model and the Dialed case study: $127 in credits replaced ~$35K/yr, and PMs ship updates themselves.
- AI policy should be a living system, not a document nobody reads.
- He closed on his first-grade report card: “a talented and determined student with a vivid imagination.”
The PB&J principle
Mike Pitone · Wawa
Wawa's CX 2030 blueprint surfaced 550+ friction points, and four themes kept repeating: “get my order right, make it easy for me, keep me informed, and recognize me.”
- Next is an AI “friction finder” with agents for intake, connections, journeys and storytelling, moving the team from “isolated repairs” to “fixing on a mission.”
- The viral PB&J video made the point: give AI loose context and it makes a mess.
- The top friction point is order accuracy. The new metric is effort score: “Effort score is smoke… You should probably fix that before there's fire.”
The AI path forward
Marjón Dean · ServiceNow, moderated by Katie Teuber · Comcast NBCU LIFT Labs
Sort work into three buckets: automate it, keep a human in the loop, or keep it human-only. The last bucket is the company's “secret sauce.”
- She calls herself “the janitor” and compared this moment to Microsoft cannibalizing itself for the cloud: “if you don't disrupt yourself, somebody else will disrupt you.”
- Her do-over: ServiceNow “underestimated” the human impact. “The employees suffer the most,” so “you got to slow down to speed up.”
The SaaS reckoning
Rick Nucci · Guru
Software is moving from tools you log into toward software that does the work for you, and the problem that lasts is accuracy.
- Five lessons: being early is a bet; don't build around a model's current weaknesses (a new one ships about every 11 days); a good brand can trap you; adoption runs 5/20/75; pricing is moving from seats to work done.
- Next wave: personal agents that work for you and send email as themselves, not as you. He expects it to be “wonderful and incredibly chaotic.”
Fireside with Ethan Mollick
Ethan Mollick · Wharton, moderated by Rick Nucci
The peak of the day. The “co-intelligence” era is over and the “coexistence” era has started: we now live with machines that are sometimes better than us.
- Organizations need leadership, the crowd and a lab working together. People still bring four things AI doesn't: deep expertise, wide knowledge, agency and taste.
- He warned about the McNamara fallacy and about a future where people only validate AI output.
- “Every one of your companies is in major cyber trouble.” His new book, Coexistence, is out Oct 20.
The new front door
Christina Clark · HubSpot
42% of HubSpot's prospects used AI search to evaluate it, and that was the #1 indicator of purchase intent.
- Results: up to 1,850% more AI-referred qualified leads, 433% more citations, and 3x conversion.
- To “become the answer”: make every section of a page a complete answer, back claims with stats (they built an internal “ProofGPT”), and keep your tone consistent across channels.
- Models kept inventing hubspot.com URLs, so HubSpot now routes those 404s to the right page.
The story of becoming an AI company
Danny Fisher · West Shore Home
West Shore grew from an $18M business to a ~$1B run rate and “became an AI company by being a process company first.”
- His three “nuggets”: get clean data through daily operations, build on what never changes, and “AI doesn't need a use case, it needs a process.”
- Small in-house image models read bathroom photos (drain location included) better than frontier models. A bathroom goes from sale to install in about 7 days.
Voice AI and recruiting
Ophir Samson · Greenhouse, moderated by Marcia Zaruba O'Connor · The O'Connor Group
The most clip-able demo of the day. Applications per job are up 239% because of GenAI (Marcia's stat), and voice screens scored against a rubric cut through the noise.
- Live demo, for a fake “Head of Conference Catering” role at O3: the AI refused to ask about parental leave and shrugged off a prompt-injection attempt.
- Zappos screened 10x more candidates, and 97% rated the experience excellent. About 20% of candidates interview after 10 PM. Ophir is also a professional magician.
AI from the president's seat
Brett Norton · Buffalo Construction, moderated by Mike Gadsby
A CEO who builds his own “toothpick” fixes.
- Project closeout dropped from 75–120 days to 25–35, and closeouts have been cash-flow positive since July.
- Estimating 1,500-page bid specs went from 10 days to 2.
- An AI Council keeps experiments from becoming silos. His rule: if you've done a task 2–4 times, automate it.
Beyond the vibes
Lazar Jovanovic · Lovable
Building is becoming free, so the new risk is “AI sprawl.”
- His four steps: align (interview coworkers as if they were customers), build together, put guardrails around one “mothership” app, and name an accountable champion. “I think that in O3, that's James.”
- Lovable's internal sales tool, Catalyst, spread to every department and saved about $1M in software.
Three AI-first startups, eight minutes each
O3's way of giving back to Philly's early-stage community. A panel of judges picked one winner.
BleedIO
Leonardo (Lee) Beup · Co-founder & CMO
Edge-based wireless mesh networks that run without Wi-Fi, 5G or GPS, pitched as “the connectivity layer for physical AI.” They have paid pilots with Chevron and Lufthansa, interest from the Army and Pentagon, and are raising now.
“We get an email on Tuesday from the Army, can you be in Huntsville, Alabama on Thursday?… We have no idea why.”
Adaptive Pallet Solutions
Adam Posner · Co-founder & CEO
Pulse puts sensors and AI on recycled smart pallets, creating a “pallet passport” for every handoff. It flagged ~$3M of at-risk cargo on a single lane.
“The pallet is truly the only independent witness in your supply chain.”
MyMonitor
Charles Mbata · Founder
Disease-specific AI that pre-screens dermatology trial candidates from phone photos. Today, finding 1,000 patients means screening about 7,000.
“When the patient shows up and they don't qualify… they have to cut them a check before they leave.”
Built from the 16 Granola session recordings. Speaker names were checked against the 1682 agenda. Quotes are lightly trimmed, not reworded.