Session 2: How to Plan, Scope and Sell your AI Projects

When a stranger evaluates you, they don't trust your website — anyone can type anything there. They trust LinkedIn, where your background, past companies and connections are verifiable. So the storefront work is profile mechanics: a clear professional photo (a selfie run through Gemini or ChatGPT is now legitimate), a banner that states your one-line offer with your domain name placed where the profile picture won't cover it on phones — check the mobile render — and the bio line that does all the selling: 'I help [niche] [automate/build/grow] [problem] with AI.' One line; a stranger should know in seconds what you can do for them.

The under-used surface is the experience section: every entry takes links and media, which is where case studies belong — 'conducted this workshop for a 9-figure D2C brand', linked. The trainer audited his own profile live and admitted the gap: his case studies live buried on his website where nobody digs. Signals of having done the thing before are the biggest credibility currency, and they belong one click from your name.

When people ask ChatGPT or Perplexity for 'best AI agencies for marketing automation', the models don't consult a ranking — they source links whose content answers the query, weighted by credibility and backlinks. The trainer's discovery: ChatGPT cites a LOT of LinkedIn posts. So the content strategy writes itself — post consistently about your niche in your niche's terms ('AI for real estate agents', story after story), and when a few posts get traction, you start appearing inside AI answers rather than under them. He learned this when a client reached out because his agency surfaced in an AI overview.

Reddit is the second citation source — the models constantly surface its threads for sentiment ('how do people feel about X'). Harder to work deliberately, but it means the communities where you're already mining pain (next concept) are also where citations grow. SEO took a decade of tactics; AEO is young enough that consistent niche posting is still the whole playbook.

The pyramid builds from the base: first the business function you resonate with — sales, marketing, accounting, finance, HR — the kind of problem you have a knack for. That alone is too broad. Add the industry — healthcare, coaches, real estate, e-commerce — and now 'sales' becomes 'sales for real estate agents': a niche with edges. The peak is the exact problem, which you don't invent — you find it where the niche complains (next concept), stated in their own words.

The payoff is operational, not aesthetic: function+industry immediately tells you where to research ('where do real estate agents hang out?' is a question Perplexity answers), and the completed pyramid produces the storefront's one-line offer. Session 1 taught WHY narrow beats broad (3-5x pricing); this is the HOW of constructing the narrowness.

The biggest question — 'what should I solve?' — has a mechanical answer: go where your niche already complains. Reddit subreddits, triaged by title, upvotes and comment count. skool.com and Facebook communities (pick ones with 1,000+ members; he joined a wholesaling-real-estate community live and surfaced 'I spent 4 hours on the dialer and felt my soul leaving my body' within minutes). YouTube channels serving the niche — where the creator's manual how-to videos are effectively automation specs, and the comments show which parts hurt. Even WhatsApp groups and hyperlocal apps for hyperlocal offers.

Two multipliers make it scale. AI does the reading: a browser extension summarizing a 295-comment thread into 'the painful problems being discussed' turns an afternoon into minutes. And the harvest isn't just the problem — it's the language. When many people resonate with the same soul-crushing complaint, you have the validation signal AND the exact words for your offer line. If your offer doesn't sound like their comments, it won't land.

The services-first logic is evidence-based product discovery. Serving clients puts you inside real workflows, where one solution eventually clicks — for the trainer it was UGC-ad automation for D2C brands, spotted via a YouTube comment section, sold as a service to three clients, then productized as Glued. The clients who bought the service became the product's first users and its proof; one creative agency he eventually stopped charging now onboards her own clients onto it — evangelism worth more than her subscription. Selling the same solution five times is the signal to productize into recurring revenue.

The strategic asymmetry: services pivot at the speed of the technology — n8n workflows yesterday, Claude skills today, same business — while products carry re-engineering overhead every time the ground shifts, and standalone AI products are 'getting crushed left, right and center' by the labs' own releases. Even the investors agree now: Y Combinator's requests-for-startups explicitly seeks AI-native services companies — don't just build AI for lawyers, provide the lawyers too, and charge 10x. Their startup directory doubles as free market research on what's fundable.

The discovery call's engine is curiosity about the business, not the feature request. Clients arrive with a narrowed symptom — 'we need this generation automated' — because they don't know what's buildable. Ask how the business operates: what's the manual workflow, who touches it, what does each tool produce, what are the high-level goals. Often the real opportunity is one level up — the founder fixating on one automation actually needs the whole content operating system, and proposing that ('forget this branch — what if I automate the entire thing?') is how scope and price both grow legitimately.

The part everyone forgets: metrics. Agree on the KPIs that define success before building — revenue moved for sales systems, leads/reach for marketing — because an undefined finish line guarantees a dissatisfied client. Do the metric research beforehand, and shape your booking intake (Calendly questions) so calls start half-scoped. Session 1's constraint diagnosis runs inside this same conversation.

His design discipline: after discovery, reverse-engineer the requirements into a flowchart — when does the system trigger, what does each step do, what data moves where. The live example: a client asked for 'lead qualification'; the diagram that sold the project showed a LinkedIn lead-form trigger, Apify scraping the lead's profile, company page and website, an LLM extracting data points into their GoHighLevel CRM, a scoring model ranking leads 1-10 (work the 8-10s; dip lower only when volume is thin), and an optional voice agent bolted on top — a bigger, clearer system than the request, visible at a glance.

The diagram works three jobs at once: the client understands exactly what they're buying; you gain a build plan that prevents mid-development chaos; and if you're business-side, it's a precise spec to hand a technical partner. Hence the writing rule: diagram first, then bullet points, and no paragraphs — nobody reads them. Tools don't matter much (FigJam, Canva, Excalidraw, Miro — or ask Claude for an HTML flowchart and screenshot it); the discipline does. The proposal then assembles itself: kit template plus meeting notes into Claude.

Founders arrive with transformation-sized asks. Committing to them cold is how engagements die: trust hasn't been earned, scope is foggy, and the first stumble sinks it. The move is decomposition: find the atomic milestone — the smallest self-contained deliverable you're confident lands in two weeks — and propose it as a paid pilot. Quote a month; deliver in two weeks. Underpromise, overdeliver, on purpose.

The pilot is mutual qualification dressed as a project: the client risks little, you learn whether they're workable, and the early win installs the belief that 'this person actually delivers' — which is what converts the 3-month contract into the 6-month partnership and later the retainer. He opens with it explicitly: 'I want to work on a pilot first.' Confidence, not hedging.

AI has made humans negligent: outputs get piped along unread until 'Here is the email for you' appears in a client's inbox. The antidote is layered: an AI QA stage inside the system, a human QA pass, and the founder's own approval — three gates before 'done'. Testing is artifact-driven: ask the client what they'll actually throw at the system, collect 10 real examples, let AI pattern-generate 20 more, and run 10-15 through before delivery.

Error handling is the professional signature. If a system breaks and the client tells you, you've failed twice — once technically, once reputationally. Everything you deploy should alert YOU on failure, and long-running jobs need explicit monitoring: his n8n workflow silently timed out on 500-page uploads while everyone just waited. Security now belongs in the same bucket — compromised mainstream libraries (the Axios incident, the GitHub hack) mean an automated daily audit comparing your stack against vulnerability news is a buildable, sellable habit. When customer counts climb from 10 toward 50, this whole function becomes a hire.

Pricing confusion is a credibility leak: if you don't know your price, the client concludes you don't know your work. The anchor discipline: pricing = value = business impact — and the cleanest source of impact is the client's own books. His masterclass case: a firm doing contract due-diligence on Fiserv's billion-dollar payment-processing deals paid a Vietnamese legal team $9-15k/month for manual document review. Once that number is revealed, automating the workflow prices at 5-10x margins with total legitimacy — the value conversation is already over. When no manual system exists, research the counterfactual: what would the 2-4 staff needed cost at market salaries?

Your negotiating range comes from three multiplied variables — AI skills (the craft), domain experience (years they can't fake), business impact (the revealed number) — each one a lever in the conversation. Frame impact positively: not 'halve your accounting staff' but 'take per-employee revenue from 5k to 15k' — raise the ceiling, don't lower the floor. And keep deals win-win: 2-3x the manual cost closes; 10x insults; and starting cheap isn't fatal — his first $700 project became a $10k/month retainer because the value showed up.

Fixed price fits defined deliverables — 'build this pipeline, hand it over' — and carries a natural upsell: training their team, billed hourly, which he sells constantly. Subscription/retainer is the destination: maintenance, support, consulting hours, the recurring revenue every first win should be steered toward (the trend-chasing content client who needs a new automation weekly is the archetype). One caution for productized AI: tokens crush subscription margins — heavy users cost you real money per interaction, so price from unit economics (cost per interaction plus margin) or sell credits.

Performance-based is the high-risk, high-trust play: 'you're at 50k MRR; work with us 3 months, we take you to 150k and keep 10-25% of the delta.' It requires the client to open their books — do it only with people you trust — and the professional structure is the hybrid: a fixed monthly floor that covers your costs, plus a slightly reduced percentage (20-22% instead of 25) on the upside. Session 1's guarantee discipline applies verbatim: his recruiting-era scar of delivering appointments no one closed is why the floor exists.

His GSB mentor's report: the books are the same everywhere; the differentiators are network and one operational truth — winners simply run more experiments. The taught exercise: before building anything, stand up a one-page landing site for the imagined offer, run Google Ads, capture emails. Students who raced five pages found two with real demand and built one. That's validation as a numbers game, not a debate — and AI's deepest business gift is parallelism: five landing pages, five offers, five niches tested simultaneously, then total concentration on whichever clicks.

The Q&A wrapped operating wisdom around it. Inbound beats outbound: when clients find you (content, AEO), you hold the pricing power; when you reach out, they do — so brand building is the structural answer to discount pressure. Niche selection includes buying habits: his worst mistake was building developer tools, because developers build rather than buy — solve rich problems for people who pay. And the 1-minute demo rule disciplines all outreach: if the Loom can't convince in 60 seconds, the 5-minute version won't save it.