How customers choose AI projects

👋 What’s up peeps!

This is Prompt Punk — the best AI sales newsletter 😉

TL;DR - inside today’s newsletter

Let’s rip…

I wish you could win AI projects simply with paper ROI. But in my experience, that’s not how they’re won. They’re won because of a company’s internal politics, timing, AND ROI math. In practice, buyers greenlight AI for three reasons.

Once the “why” is settled, the “what” is surprisingly consistent: fast, simple, provable ROI beats grand strategy every time.

The moment you rewire a sales process, you touch finance, product, engineering, senior and junior sellers, and marketing. Each handoff adds approvals, approvals add delays, delay kills momentum. Earn trust with solo wins, then stitch big projects together one at a time.

A few domains justify complete cutover rather than caution:

Other functions will come as AI develops.

Bottom line: Customers choose AI projects that are obvious to start, quick to prove, and cheap to defend.

Sequence your wins, publish the gains, earn the right to rewire the messy middle.

Do that, and you won’t need a miracle. You’ll have a track record.

Airbnb wired an AI customer-service agent into its app. They evaluated OpenAI, but production leans hard on Alibaba’s Qwen for speed + cost.

Result: Fewer tickets hitting humans (15% reduction) and half of U.S. users routed through the bot during rollout. Qwen does much of the heavy lifting, OpenAI is in the mix but not the core production engine right now.

Customers notice: Faster resolutions inside the app, fewer handoffs to live reps as the agent learns from tens of thousands of prior conversations.

How they win
They didn’t just “turn on AI.” They built a multi-model fleet (OpenAI + Google + Qwen + OSS), picked Qwen for cost/perf on core workflows, and trained/guardrailed the bot on real support data then shipped it to real traffic.

Why it matters
Proof that the ROI isn’t from the brand of model it’s from shipping the workflow. The playbook: test widely, productionize the cheapest fast model that meets quality, and let humans handle the edge.

Virgin Atlantic plugged AI into paid seat selection pricing and saw >10% seat-revenue uplift. No new routes, just smarter monetization. Using their AI engine, prices adapt to demand in real time, turning each row into micro-inventory. Maybe that explains why prices keep changing 🙄.

How Virgin wins
• Contextual pricing by route
• Continuous experiments to maximize conversion
• Bundles that sell: seats → priority boarding, bags, Wi-Fi

Why it matters
Add-ons are airlines’ cleanest margin. A 10% lift can be the difference between profitability and loss. Start with seat maps, then roll to bags, upgrades, lounges, and Wi-Fi.

This kind of thinking can apply to other industries such as hotels (room/view/late checkout), rail & buses (seat selection), sports/concerts (dynamic seating, VIP access), car rentals (vehicle class, insurance), e-commerce (shipping speed, gift wrap), SaaS (usage add-ons, priority support), gaming (skins, battle passes), and telco (data boosters). The potential is endless.

What it does: Copy.ai’s “AI Sales OS” lets you stitch workflows that take a call from transcript → summary → CRM updates → follow-ups → next-step tasks. It can auto-join meetings to capture notes, coach deals off transcripts, and push structured data back into your stack. Think codified playbooks—not one-off prompts.

Example: After a discovery call, the recorder bot posts a summary, creates action items, updates fields in Salesforce/HubSpot, drafts a recap email, and schedules the next meeting.

Why it’s valuable: You standardize what “good” looks like across reps (and quarters), compress admin time, and keep deals moving while the team sleeps. Workflows make wins repeatable.

Website: copy.ai

What it does: A real-time email coach for Gmail and Outlook that scores your draft, suggests subject lines, trims fluff, and personalizes with context, right in the compose window.

Example: Draft a cold opener in Gmail, Lavender grades tone and length, surfaces a personalization angle, tweaks for mobile reading.

Why it’s valuable: Fewer rewrites, more replies.

Website: lavender.ai

I installed the new ChatGPT Atlas browser and my life got easier. Not “reinvent your life” easier, more like a steady 20–30% lift.

It has speed you actually notice, a clean UI that stays out of the way, and a memory that keeps your research thread alive across days.

Search turns into “context first, links second” instead of a click-maze. Voice works. Agents… not yet.

If you spend your day researching, writing, or bouncing between docs and links, this is the quiet edge you’ve been missing.

What Atlas nails

1) Fast enough to feel different
Pages pop. Scrolling is smooth. The UI friction is low.

2) Minimal on purpose
Toolbars are sparse, dropdowns are fewer, cognitive load is low.

3) Search that starts with AI context
Type and you get reliable top links plus an AI answer. Need a specific URL? It’s right there. Want summary/context? It’s there too. Result: fewer tabs, faster decisions.

I only wish AI responses were faster.

4) Memory that compounds
Research a holiday or a multi-day project and pick it up later. Atlas still remembers the thread. That continuity is the overlooked superpower.

5) Voice that’s actually useful
Talking your way through tasks is surprisingly effective. You won’t use it always, but when you do, it clicks.

Where it stumbles

Agents are slow and trust-draining (for now)
I tried a basic “email my wife hello” flow with email access granted. It lagged and fizzled. The promise is there, the execution isn’t. Treat agents as beta-adjacent.

Bottom line

Atlas is a 20–30% upgrade over Chrome for real work. It won’t blow your mind, it will grease your day.

Final scorecard

DeepSeek released an open-source OCR model that shrinks long text into image-based “vision tokens,” cutting context size by ~7–20×. Instead of chewing through costly text tokens, it renders pages as high-res images and lets a vision-language decoder recover the words turning sprawling docs something models can afford to read end-to-end.

Why it matters
If you pay per token, this is a margin machine. Cheaper long-context means you can put “all your docs/files/code in one prompt” for legal, finance, and engineering this is a game changer.

Meta is trimming about 600 roles across its AI research, product, and infrastructure groups while continuing to hire into a new advanced lab focused on next-generation models.

The research group isn’t disappearing, but the center of gravity is shifting from open-ended exploration to productized, ship-fast AI. Inside the company it’s framed as moving to smaller, more accountable teams to speed decisions.

Why it matters
Meta has been the anchor for Western open-source models. If its center of gravity moves from open research to product packaging, expect less western open source and Chinese domination. Our last hope is Mistral.

It ain’t perfect folks 😂

Thanks for reading! 👋

If you liked this:

📨 Forward to a biz techie friend
💬 Hit reply and tell me what you want more (or less) of
📊Vote in the poll below 👇

— John
Prompt Punk