ROI-first AI for operators
Marc Pelberg
I help operators put AI where it pays for itself, cutting cost or growing revenue, and skip the projects that won't.
95%
Enterprise AI projects never show a real ROI.
I'm building the other 5%. The difference isn't the model. It's about putting AI only where it reduces costs or increases revenue, and discarding everything that doesn't.
Built across operating companies
The ROI test
If it doesn't reduce costs or increase revenue, it won't get built.
ROI is the gate
Every stage is screened against one question: does it cut cost or grow revenue? Work that passes gets built. Everything else is skipped.
-
Core business
The work that drives revenue and margin. That's where AI pays first, by making it sharper.
-
Secondary work
The tasks quietly draining cost: cleanup, matching, reconciliation, reporting. Prime targets to automate.
-
Automation boundary
Dashboard, quiet background agent, or recommend-only, matched to the risk and the payoff.
-
Trust before action
Review, audit trails, and thresholds. Autonomy grows only as the return is proven.
Why this perspective is different
Operator first, builder second, AI only where it earns trust.
I have built and run real operations: international e-commerce, retail automation, and AI-driven product data, not just advised on them. That is why my AI work is not decoration. It closes the gap between a process, a decision, and a result you can trust.
- RealArb
- Marketplace and arbitrage operating experience across international e-commerce, where small decisions compound into margin.
- SkuTrue
- AI-driven retail operations and product-listing automation for catalog work that should not stay manual forever.
- Risk judgment
- Private pilot, EMT, and state/regional paragliding distance record holder across New Jersey, Pennsylvania, New York, Michoacán, and Puebla. I am comfortable deciding where automation is safe and where a human has to stay in the loop.
- Consulting & ownership
- Owner-level operating context from Midas Holiday Lighting, Eli & Associates, Inc., and Artex Knitting Mills, plus advisory work with e-commerce and retail brands.
From the build log
Real systems, straight from the repos.
Not slideware. These run today — for my own companies first, which is the harshest client there is.
-
An agent fleet with me in the loop
A fleet of AI agents runs research, maintenance, and deploys across my companies, commanded from a single Telegram thread. Every agent registers a heartbeat on a live dashboard, and anything risky, credentialed, or expensive stops and waits for me.
It's where the show got its name.
-
Automated market scraping
RealArb's engine turns retail sites into structured product, price, and availability signals. Inventory rebuilds itself on a schedule, and opportunities arrive ranked, so a human reviews decisions instead of spreadsheets.
Margin found while nobody was watching.
-
10-marketplace catalog sync
SkuTrue syncs catalog workflows across 10 marketplaces, including Amazon, eBay, Walmart, and Mexico's Sellopublico.mx, normalizing listings, attributes, and channel rules without hand-copying every SKU.
Thousands of SKUs, zero hand-copying.
-
Two subscriptions, canceled by code
When a paid tool is just a middleman over a free API, I build the middle out: a self-hosted ads-reporting server that talks to eight ad platforms directly, and a drop-in Amazon-data scraper that replaced a metered API. Both read-only by default — no live campaign changes without a human.
The cheapest subscription is the one you stop needing.
-
Real-time product data across 450+ retailers: look up any UPC and get prices, availability, images, and margin estimates. Shipped as a public API and an MCP server, so AI agents can plug straight into it.
The internal tool that became a product.
-
A trading bot that knows its limits
An automated BTC perpetuals bot on Hyperliquid: strategies are backtested before they touch money, position size and drawdown have hard ceilings, and a dashboard shows every decision it makes. Autonomy is earned here the same way it is everywhere else: by proving it.
Trust before action, with real money on the line.
How engagements work
How a project actually runs.
01
Audit the work
I map where your team's time and money actually go, and separate revenue work from the repetitive tasks no system owns yet.
02
Build the first system
We start with one high-leverage workflow, with a human in the loop, and ship something real instead of a slide deck.
03
Raise autonomy carefully
The system takes on more only once accuracy, the review trail, and the business outcome prove that it should.
Stay in the loop
The show, on YouTube
Marc in the Loop
One real business problem per episode. I build the AI system that solves it on camera — the wins, the dead ends, and what it actually cost. If it doesn't pay for itself, I say so out loud.
First episodes are in production. marcintheloop.com is where they land first.
Reach out
Have a workflow that is eating your team's time?
Send me the process, the pain, and what a good outcome looks like. I will help you decide whether it should stay human, become software, or run quietly with the right checks in place.