The day after an AI implementation team left the offices of U.S. merch company Imprint Genius, staff used Slack to build a live client proposal on their own. Minutes later, the client replied saying they literally gasped out loud. Then came the catch: it looked great, but parts of it felt a bit too "AI." A day later, the team realized two days of casual testing in that same Slack channel had burned $100 in generation credits.
Imprint Genius is a grounded business with zero venture-capital hype. Founder Isaac Hetzroni, known online as The Sourcing Guy, has spent ten years in corporate merch and global sourcing. Brands come to him for hoodies, mugs, and gifts, and his team finds factories, runs production, and ships the goods. Annual revenue crossed $10 million long ago. Embedding an AI pipeline into that machine over seven days exposed every pitfall businesses hit right after the shiny demos end.
The Million-Dollar Bottleneck: Proposals in 20 Minutes for $4.31
When consultants walk in trying to automate everything at once, projects stall on day one. At Imprint Genius, they worked backward from revenue on paper and hit one clear bottleneck: building sales proposals.
Account managers built 15 to 20 proposals a month, spending about ten hours on each one—picking items, mocking up client logos, calculating prices, and formatting decks. Basically, one full-time employee spent all month assembling proposals. Nine out of ten never closed, though every win brought a first order around $10,000.
That kind of bottleneck is ideal because AI improves it on all three fronts: speed, cost, and quality. Clients used to wait a week for a deck; now they get a link the same day. Custom 3D renders cost $500 to $1,000 each, while an AI-generated product photo with a logo costs about eight cents. Instead of a static PDF, the buyer opens a private web page that's easy to tweak on the fly.
The pipeline follows a strict sequence:
- A request arrives from the website or Slack.
- Web scraper Apify visits the client's site and pulls their logo, brand colors, and fonts.
- The system picks matching catalog items and posts the list in Slack—nothing moves forward until a human manager clicks approve.
- The generator creates branded photos and six-second videos, while a second AI checks the logo in every frame against the original so it never warps by a single pixel.
- The output compiles into a private web page hosted on Railway, a cloud platform for running web apps without server headaches.
CodeRabbit watches the codebase as an automated AI reviewer, catching bugs and security gaps after the outside engineers leave. Crucially, the system never sends anything to a client on its own—a human always reviews the final page.
In a live test, a proposal with three products, three photos, and three videos cost $4.31 in tokens and took about 20 minutes of compute time. Running the whole stack costs roughly $500 a month plus a few dollars per proposal, compared to paying a full-time specialist for 150–200 hours of manual layout work.
The Data Ceiling and the Final 20% of Taste
On paper, the pipeline looked flawless until it met Imprint Genius's real database. The engineers made a classic mistake: they started building automation and didn't inspect the raw data until mid-week. Ten years of operations had turned the product catalog into a patchwork quilt.
To build a proposal, the system needs three things for every item: a clean photo, a price, and specs. Filtering the database revealed a brutal reality.
Imprint Genius catalog readiness for the AI pipeline
Only a tiny fraction of the catalog made it across the finish line. Human managers had worked around this mess for years: they kept prices in their heads, confirmed details in email threads, and remembered which factory stocked what. An AI can't read an employee's mind—if it isn't in the table, it doesn't exist.
It got worse. Out of 1,600 product photos, the algorithm threw out more than a thousand because they already had past clients' logos printed on them. You can't slap a new brand on top of someone else's. Worse still, the scraper once misfired and grabbed a third-party Costco badge off a client's website instead of their logo—which is why they added the second AI verification step. The entire catalog and factory list moved into Supabase, a Postgres cloud database that works like a smart spreadsheet readable by humans, apps, and AI agents alike.
The second surprise hit when the team started using it. The first demo page built purely by AI worked fine, but looked plain. The next day, Isaac's designer Devin set the visual style and picked the products himself, letting AI generate the layouts, specs, and videos. Quality jumped a level—that was the page that made the client gasp.
Even then, the buyer spotted machine fingerprints in templated copy and flashy video transitions. Devin compared it to school PowerPoint decks: the best slideshows stick to one clean transition, while the worst cram in every 3D spin available. AI works the same way. It handles the first 80% of the heavy lifting, while the last 20% is human taste—and that's what closes the deal. To get drafts closer to that bar on the first try, the team writes the top specialist's rules and best examples straight into the model's instructions.
Teaching Them to Fish and Moving AI Into Team Chat
Custom dashboards often die quietly a couple of months after handoff. For a previous luxury travel client, the same implementation team built a bespoke portal, and a few months later only two screens were still in use. For another client, Daniel, they built a giant system with 30+ screens and 13 agents—once the contractors left, his team had to maintain it alone and couldn't adapt it when needs changed.
At Imprint Genius, they skipped the bloated UI. Data lives in Supabase, and all proposal logic (31 tools covering site analysis, product matching, and image rendering) sits inside an MCP server hosted on Railway. Put simply, MCP (Model Context Protocol) is a standard adapter that lets any compatible AI agent call internal company tools without custom glue code.
The team got two workflows. First is solo work with coding agents like Claude Code or Codex. Employees learned to build small tools themselves instead of waiting on contractors. After Daniel's team learned Claude Code, revenue per employee doubled, and Isaac is already spinning up his own micro-apps on Railway to replace paid SaaS tools for a couple of dollars a month.
Second is the team workflow. AI hidden in a founder's private tab won't move a company forward. So they added an AI teammate named Viktor to their Slack workspace. It has a normal account: staff tag @Viktor in shared channels just like a human coworker, and it joins the thread with full access to the database and tools.
Viktor — an AI teammate for collaboration in Slack and Microsoft Teams (Source: viktor.com)
Dropping a bot into chat isn't enough—you have to tie it to a habit people already have. At Daniel's company, a standalone dashboard sent over 250 notifications in two weeks, and staff acted on just 15. Meanwhile, a content tool stuck immediately because it replaced daily grunt work. Imprint Genius already lived in Slack and built proposals every day, so the bot fit right in. Inbound lead triage works the same way: an agent filters spam from real buyers and sponsors, researches legitimate companies, and posts a score straight into chat.
The Token Bill and Knowing When to Step Back
AI has a sneaky trait: every request costs pennies until someone starts hitting the button with no brakes. Imprint Genius's proposal pipeline stayed at $4.31 because every step was strictly scripted. Once staff started chatting freely with Viktor in Slack and asking for extra variations on the fly, $100 vanished in two days—mostly eaten by video generation.
Rate limits caused headaches during setup, too. Their product-photo API ran out of credits twice, the second time live in front of the client's team, forcing an emergency switch to another provider that same day.
If your API and media generation costs start climbing, ask the AI system itself to audit its own usage logs. An agent will quickly spot the most expensive steps—usually extra video runs or heavy models used where cheap ones work—and suggest optimizations with clear quality trade-offs.
There is always a gap between a working v1 and a mature tool built for daily order volume. Outside engineers can build the frame in seven days, but the client's own team—people who know merch, factory quirks, and client taste inside out—can polish it faster and better. Once staff have a clean database, a clear MCP toolkit, and coding-agent skills, the best thing an integrator can do is step back and only return when the system hits a real ceiling.


