The day, session by session.

Jannis opened the day and set the frame for what followed.

The Morningside AI founder walked through five common AI-adoption mistakes drawn from recent client work, and his fix: an "AI workspace" where non-technical staff build their own tools with a coding agent. Attendees were handed a starter workspace zip to try themselves.

Key ideas

• Five mistakes: scattered data AI cannot reach; no centralised AI strategy; treating AI as a developer-only job; rigid automations on unstable processes; work invisible outside meetings.
• Skills vs automation rule: flexible skills for unstable processes; rigid automation only where nothing will change within about six months. Worked example: an accounting firm's working-papers process cut from about an hour to about 15 minutes (their figure).
• The "Ledger": every piece of work, human or agent, continuously logged so managers see progress without meetings.
• Delivery model: the "AI Makeover", an intensive week of setup ("seven months of progress in seven days"), after which the client's team self-manages.
• Small teams are winning right now: enterprises are slowed by security and permissions; open, flexible systems favour smaller companies.

Slides (5)

A layered AI-operating-system architecture and how it scales from 5 to 1,000 people: identity at the centre, then substrate, rules, hooks (the only deterministic layer), skills and connections. The commercial heart of the talk was "rot": a system starts becoming obsolete the minute it is built, and someone has to be its plumber.

Key ideas

• Layer model: identity (changes about yearly, includes a librarian index of file paths), substrate (durable files), rules (change often), hooks (when X then Y, 100% of the time), skills and workflows, connections.
• "Rot": he keeps a maintenance file cataloguing every file and process with its expected obsolescence window per layer. "Plumbing is the most profitable thing in AI."
• Hooks as compliance: he spent a month hardening hooks for one client so wrong-client data could never route to a wrong-client surface.
• Rollout by company size: 5 people = a NOW vs BACKLOG list; 15 = AI stewards per team, department by department starting revenue-adjacent; 50+ = departmental playbooks and information gating; 1,000 = hardened hooks and strict governance.
• Agent-hiring philosophy: an agent is a job; start with one, split only when scope gets weird, because errors cascade.
• Six-month purge: delete old skills and instruction detail every six months or so; old scaffolding handicaps newer models.

Slides (9)

The ClickUp co-founder's growth playbook: audience beats product, scrappy acquisition, and the strategic bet that context is the moat. The talk ranged from the "Jira Sucks" ad to answer-engine optimisation and employee-generated content.

Key ideas

• Audience beats product: their pre-ClickUp app was better than Snapchat's clone and still died. "No matter how good your product, it means nothing if no one sees it."
• Scrappy plays: scraped competitors' 2-star reviews and contacted the unhappy customers; when a competitor shut down, they built a migration tool in three days and onboarded around 20,000 users (his figure).
• "Swing up": attack bigger competitors by name; incumbents will not punch down.
• AEO replaces SEO: get cited by the answer engines people actually ask, by showing up on the platforms those engines read, often via many small creators rather than a few big ones.
• Employee-generated content: 30+ employees posting daily; people follow people, not logos; raw phone-shot beats polished.
• Rule of three: every post must solve something, make them laugh, or make them feel something. Otherwise do not post.
• "Context is our focus": models are interchangeable; the company that owns the context of work wins.

Slides (4)

Why chat-first AI assistants break past about 30 users, and the three-part architecture that replaces them for mid-market companies: a document brain, text-to-SQL business intelligence over an enterprise ontology, and deterministic scheduled workflows that cost nothing in inference.

Key ideas

• EU enterprise buyers often reject hosted-LLM stacks on data residency before the demo; his clients "bring their own LLM" and pay for the system, not tokens.
• Quality must come from the system (ontology, SQL, deterministic routines), not the frontier model - model-agnostic by design.
• A chat-first brain degrades past about 10,000 documents and 40+ concurrent users (his figures); deterministic workflows make daily tasks free at runtime.
• Departments as workspaces, with the C-level federating across all of them - a clean permission model for multi-user setups.
• An honest ceiling: chat-first operating systems are "great for SMBs, but when it's more than thirty people, it's not enough" - past that you need the BI and ontology layer.

Slides (7)

Two operators who productised agency delivery into installable skill bundles ("plugins"): process capture beats engineering, a skill is inputs plus rules plus tests, and the retainer is a slice of the measured saving.

Key ideas

• A plugin is a bundle of skills mapped to a departmental workflow; one 2-hour boardroom session with the domain experts can automate a large share of a 10-20 hour process (their claim).
• A skill is inputs, rules, tests - validated with 20-50 isolated trials before it is trusted, because a misquote can cost hundreds of thousands.
• Worked example: construction estimation cut from about two weeks touching eight people to same-day - faster quotes mean more won work, not just saved hours.
• The ROI sale: audit first to baseline hours, touches and cycle time, then "give me $1 now and get $2 back in two weeks".
• Traffic / System / Skill: every service business is bottlenecked by exactly one of lead generation, conversion or delivery at a time - diagnose the binding constraint and work only on that.

Slides (5)

Why service businesses exit at roughly 2.7-3x annual profit, and how to break the ceiling: change the asset class, not the multiple. Package accumulated data and intelligence into a proprietary technology asset that is transferable and defensible.

Key ideas

• Grounding exercise: average monthly profit x 36 - that is the ceiling while buyers are valuing cash flow tied to human delivery.
• The escape: package your intelligence into a technology asset that is transferable (documented, a new owner can run it) and defensible (hard to recreate).
• "The multiple increases when buying what you build becomes easier than recreating what you know."
• Service firms beat pure software plays to product-market fit through warm client introductions.
• He shared a working tool with attendees that reads your own workspace and identifies licensable IP and data products.

Slides (5)

A closing morning session on authentic AI-driven content at scale: a content factory trained on his own scripts, a sponsor play built on showing brands their finished sponsorship before they pay, and a daily one-document sweep of everything that matters.

Key ideas

• Content factory: an agent trained on all his past scripts drafts in his cadence, and a director agent assembles a near-complete edit; he cited 5.6M views in 30 days without appearing on camera (his figure).
• Sponsor play: hundreds of personalised AI videos, his face plus each target brand's logo, showing sponsors what their sponsorship would look like. "Show them the end result before they pay."
• His daily "sweep": one document consolidating all inboxes, tasks and money, read before the day starts; three must-do items, everything else is noise.
• Core thesis: with everyone on the same models, authenticity is the only durable differentiator - AI distributes your original thoughts, it never generates them.

Slides (1)

A walkthrough of Hostinger's internal AI stack: Dex, a Slack-native company agent with context, memory and per-employee dashboards, and Lumos, plain-English text-to-SQL - plus, refreshingly, the two things that failed.

Key ideas

• Dex use cases: a daily sweep of 70-80 competitors posted as per-product Slack summaries; screen-record a bug and Dex files the ticket and drafts the fix for human approval; most influencer research automated.
• Around 500 staff now have a personal AI agent (their figure); their 2025 goal of 50 automation-capable employees was beaten at around 100.
• What failed: unreviewed AI output shipped to customers caused complaints (now human-reviewed), and automated KPI-setting was reverted because it removed judgement.
• Education-first rollout: workshops, learning hubs, and an AI champion per team.

The team behind Glaido, a fast, private, EU-hosted dictation tool and conference sponsor, on building the product in the open: component-level design systems, feature-creep discipline, waitlist pre-validation and compliance as a moat.

Key ideas

• Design systems past brand guidelines to component-level definitions (button states, hover) so AI tools build pixel-perfect UI - "do it once even for internal products".
• Waitlist as the demand test: "if I can't get 300 people to say they want this, how am I going to sell it?"
• Feature-creep discipline: they experimented with about 20 features and shipped almost none - "we first really need to do one thing right".
• Compliance as moat: understand compliance deeply enough to automate it, so you can sell into enterprises without adding headcount.
• The product processes data without storing it, with an on-device layer in development and one-click dictionary import - switching is cheap.

Slides (1)

Jack Roberts on the systems agencies actually use to deliver with AI.

Two co-founders of a done-for-you Instagram acquisition agency for high-ticket offers: a full-funnel machine with exactly two human touchpoints - recording the content and taking the sales call. They shared a skills pack with attendees.

Key ideas

• Content factory: finds outlier competitor posts (5x average views), scripts in the founder's voice, and A/B tests three hooks per video via Instagram trial reels before promoting the winner.
• Ads only amplify organic content that has already proven itself - never cold creative.
• AI DM setter: speed-to-lead under 20 minutes, every engagement triggers a conversation; they say almost no leads clock that it is AI (their figure).
• Where the AI stops, the CRM hands a human a follow-up task inside Instagram's 24-hour window - deliberate AI-human fusion.
• The close: "You were here - the system you just experienced is what I'm selling."

Slides (10)

How he reached 100k YouTube subscribers in around 100 days (his figure) and turned attention into clients without outbound: a three-level system that starts with pure consistency and keeps AI strictly in the organising seat.

Key ideas

• Level 1 is consistency only: one video a week for ten weeks, quality explicitly not judged ("climb the cringy mountain"), enforced by forfeit-based accountability - miss a week, run 10km.
• Personal brand over company brand: "opportunities change but trust moves with you."
• Long-form first; one long video repurposes into about ten assets - "one unit of work, ten of distribution".
• Anti-slop filter: only make content you can teach without opening Google; ten minutes daily writing ten possible titles.
• AI organises, never writes: ramble first, let AI structure it into Hook / Setup / Points / CTA - scripts written by AI sound like nobody.

Slides (1)

A live coaching hot-seat: Lauren working through Dave's creator-to-client funnel on stage. The mid-funnel machine: a monthly live "training" (never say webinar) with one evergreen opt-in page and a rotating backend, a thank-you page that books sales calls before the event, and AI-mined sales transcripts for real buying language.

Key ideas

• The thank-you page is the sales page: ICP headline, short pitch video, embedded calendar. Their benchmark: 5-7% of opt-ins book a call from the thank-you page, and the live event itself converts about 5%.
• Opt-in benchmarks: 20-25% cold, 35-40% warm organic (their benchmarks).
• Mine closed-won sales-call transcripts with AI for the language that actually made buyers buy - it becomes titles, hooks and page copy.
• Never say "webinar" - call it a training; the title should name the audience's deepest current fear.
• LinkedIn's Services marketplace: list your services, post on the topic, and inbound arrives from outside your network.
• Repurposing rule: write for the prospect's awareness level, not the client's - hook, literal bullets, one golden nugget.

How AI consultants stay relevant as the tools change: diagnose the real constraint, agree one objective KPI before building, and price on value - roughly 10% of projected annualised year-one value, so the client sees a 10x return path. He shared his pricing masterclass document with attendees.

Key ideas

• Clients buy belief, not value: diagnose the real constraint, not the request. His med spa example: she asked for lead generation; the leak was no-shows and zero follow-up.
• Two discovery questions, each followed by deliberate silence: "If you had 10x the business tomorrow, what would break first?" and "How do you get 10x the water into your pipe tomorrow?"
• Agree one objective KPI (baseline plus target) before building - "feeling less busy" is unprovable; "5 appointments a week to 10" is defensible.
• Pricing: about 10% of annualised year-one value. The worked maths from the slide: $40/hour x 10 hours x 52 weeks = $20,800 of value, so price around $2,080. Walk the maths confidently, then silence.
• Never discount - slice scope instead: define a v1 that fits the budget at full rate and phase the rest. Discounting on first pushback teaches clients to push.
• Retainer path: return at months 1, 2 and 3 with the agreed number moved - objectivity converts a project into a retainer.

Slides (7)

The day closed with an announcement, then moved to dinner at sea from Port Pine, Tivat.