How to Actually Make Money With AI in 2026: An Engineer Ranks 13 Side Hustles

Last updated: September 2026

TL;DR: Almost every AI side hustle ranking you’ve watched was made by someone whose actual business is selling you the ranking. I build the systems instead, so here are 13 ideas graded on five criteria I state up front: upside ceiling, moat, technical leverage, time to first dollar, and fulfillment burden. The S tier is distribution, an owned audience, because it makes the other twelve easier, cheaper, and faster. The D tier is anything Claude Code can finish in an afternoon, because the margin on that work is already gone.

Why most AI side hustle advice is wrong

Search “make money with AI” and you get a wall of thumbnails from people whose revenue comes from the video, not from the business in the video. That’s an incentive problem, not a character flaw. A course seller needs every idea to look reachable in 30 days, so the ranking optimizes for how good it feels to start, and says nothing about whether it still pays in month nine.

What you get here: five criteria stated before the ranking, applied to all 13 ideas the same way, with the specific technical reason each one sits where it sits. Where I’m guessing, I say so. Where there’s a real number, it links to its source. There’s no course at the end of this.

I’m a founder and engineering leader with 13+ years across distributed systems, multi-cloud infrastructure, and production AI. I’m founding engineer at EnergyConnect, I co-founded PopSocial and scaled its engineering team from 1 to 15+, and I build AI agents and automation systems in Go and Elixir, for my own companies and fractionally for founders.

How I ranked these

1. Upside ceiling. Can this plausibly scale past a few thousand a month, or is it structurally capped at freelance income?
2. Moat. Could Claude Code or a free template do it in an afternoon? If yes, the margin is already gone. I wrote about the durable version of this in the schlep is the moat agents can’t cross.
3. Technical leverage. Does knowing how to build give you an edge, or does the tool flatten everyone to the same output?
4. Time to first dollar. How long from starting to money in the account, assuming you work on it seriously.
5. Fulfillment burden. Once it works, does the business own you? Revenue that needs you personally, every day, forever, is a job with extra steps.

Quick reference: all 13, ranked

Idea	Tier	Time to first dollar	Ceiling	Moat	Best for
AI clipping	D	Days	Low	None	Nobody, as a business
AI website generation	D	Days	Low	None	Practice reps only
Faceless channels	C	1 to 6 months	Medium	Weak	Editors with taste
AI voice agents	C	2 to 6 months	High	Real	Patient builders
AI UGC ads	C	Weeks	Medium	Weak	Performance marketers
AI newsletters	B	2 to 6 months	Medium	Audience	Writers with a niche
Small business consulting	B	Weeks	Medium	Trust	Local operators
AI live shopping	B	Weeks	High	Operations	On-camera sellers
Trading augmentation	A	Months	High	Your capital	Existing investors
High-ticket consulting	A	Weeks	High	Judgment	Senior operators
AI digital products	A	1 to 6 months	High	Distribution	Builders who ship
AI automation agency	A	2 to 8 weeks	High	Engineering	Technical founders
Media and distribution	S	3 to 18 months	Highest	Compounding	Everyone, eventually

D tier: AI clipping

Clipping is taking long podcasts or streams, cutting them into vertical shorts, and getting paid per view by a creator’s affiliate program or a brand running a clipping campaign.

It’s D tier because the “AI does it all” claim is false. The models are good at finding speech boundaries and bad at finding the moment that makes someone stop scrolling. Taste and manual curation still decide what works, so this is piecework at a rate set by the cheapest person willing to do it.

Right for: nobody, as a business. It’s a decent way to learn what makes a hook land, and you can get paid a little while learning.

The detail the thumbnails skip: the platform side is tightening. YouTube renamed its “repetitious content” monetization policy to “inauthentic content” in July 2025, and eligible content must “not be mass-produced, generic, repetitive, or manipulative”. If your workflow is one template applied 40 times a day, that policy was written about you.

D tier: AI website generation

You sell websites to small businesses and generate them with an AI builder instead of hand-coding or buying a theme.

This is the clearest case of the moat criterion firing. I can stand up a clean, fast marketing site for a local business in an afternoon with Claude Code, and so can a smart 19-year-old on a free plan. When the supply of a service goes to infinity and the quality floor rises at the same time, the price goes to the floor.

Right for: getting reps. Build ten cheap sites for local businesses, learn how to talk to owners, then sell them something that isn’t a website.

The detail: the website was never the expensive part. The expensive part is the content, the photography, the Google Business Profile, the review pipeline, and the person who answers the phone when a lead comes in. That’s all schlep, and schlep is where the money still is. For the build side, my Claude Code resource bible covers the setup I use.

C tier: faceless channels

A channel with no on-camera presence: script, synthetic voice, stock or generated footage, published on a schedule to YouTube, TikTok, or Instagram.

Better than clipping because you own the asset, and the ceiling is real if a channel lands. Still C for the same reason as clipping: the automation gets you a publishable video, not a good one. Every faceless channel that works has a human making dozens of taste calls per video about hook, pacing, and topic. That’s the exact labor the pitch says you’ve removed.

Right for: people who enjoy editing and know a niche cold. Finance, history, and gear reviews still work. Generic “top 10 facts” is a dead lane.

The detail: the failure mode is the algorithm’s cold start, not policy. You spend real money on tokens, voice, and stock footage for months before a channel tells you whether it works, and none of that spend is recoverable. Budget it like R&D, not inventory.

C tier: AI voice agents

You build a phone agent that answers calls for a business, books appointments, answers questions, and hands off to a human when it should.

This one has a real moat, which is why it isn’t D. Telephony is hard: latency budgets, barge-in and interruption handling, turn detection, call transfer, recording consent, and the integration into whatever booking system or CRM the business already runs. It’s C anyway because the ramp is slow, the buyer is a small business that moves slowly, and the space is filling up with people selling the same white-label demo.

Right for: patient builders who can sit through a six-month sales cycle and like plumbing.

The detail: the latency math is unforgiving, and most demos hide it. ElevenLabs lists its Flash v2.5 model at about 75ms for synthesis, which sounds like plenty of headroom until you add speech recognition, the LLM turn, network round trips, and telephony jitter. Every one of those stacks, and my working rule is that the whole path has to land under roughly a second before a caller starts feeling the machine. Getting there is an engineering problem, which is good news if you’re an engineer. A call is a long-lived stateful process, so I’d reach for the supervised, concurrent architecture in building AI agents on the BEAM. Treating a call as request/response is how you get agents that drop context mid-sentence.

C tier: AI UGC ads

You generate creator-style video ads with tools like HeyGen or a synthetic voice, and either sell them to brands or run them yourself on performance.

C because it’s a real, paid service with an obvious buyer, and because the differentiator is the ad concept, not the generation. A brand doesn’t need 200 variations. It needs three that convert, and the model doesn’t know which three. Whoever brings the media buying judgment captures most of the value.

Right for: performance marketers who can already read an ad account. If you can’t tell a hook problem from creative fatigue, you’re selling assets, not outcomes.

The detail: the uncanny-valley discount is real. Avatars that read as synthetic get scrolled past, and the fix is a scripting and framing change, not a better model. That’s a craft skill, and it doesn’t transfer from the tool.

B tier: AI newsletters

A niche newsletter where AI does research, drafting, and summarization, published on something like Beehiiv, monetized by sponsorship, affiliate, or a paid tier.

B because you own the list. Email is the one distribution channel nobody can algorithm away from you, which makes every subscriber a durable asset instead of a rented impression. The cap is that the content itself has almost no moat. A competent summary of the week competes with a hundred identical ones, and a model can produce all of them.

Right for: writers with a real niche and a point of view, ideally with operator credibility in it.

The detail: the money depends on the niche, not the list size. A 5,000-subscriber list in a B2B niche with real buying power can out-earn a 50,000-subscriber general-interest list, because sponsors buy the reader, not the count. Build for the narrowest audience you can stand.

B tier: AI small business consulting

You go to local businesses, find the five hours a week they waste, and fix it with AI tooling. Invoicing, intake, scheduling, review responses, quote follow-up.

B because the demand is enormous and almost completely unserved, and because trust is the moat. The same owner who ignores a cold email will hand you their whole operation if you fixed one thing for their neighbor. The cap is geography and your calendar. This is per-hour work until you productize it.

Right for: people comfortable walking into a business and talking to the owner. Technical depth helps less than you’d think here. Willingness helps more.

The detail: the sale is almost never the AI. It’s the report you hand them in week one showing where the hours go. Most owners have never seen their own process written down, and that document makes them believe you before you’ve built anything.

B tier: AI live shopping

You sell products in live video streams, on Whatnot, TikTok Shop, or similar, using AI for clip repurposing, listing descriptions, and post-stream follow-up.

The category isn’t speculative. Whatnot raised $545 million at a $20 billion valuation in August 2026 on $8 billion of 2025 live-sales GMV, which is what a real market looks like. It’s B because the business is operations: sourcing, inventory, shipping, returns, and your face on camera for hours. AI helps at the edges and can’t help at the core.

Right for: people who like being on camera and are good at sourcing. If either is false, skip it.

The detail: the constraint is inventory cash and fulfillment throughput. The AI tooling shaves an hour off your listing and clipping work. It does nothing about the boxes in your garage, and the boxes cap your growth.

A tier: AI trading and investing augmentation

Using AI to screen securities, summarize filings and earnings calls, and build or evaluate backtests. Augmentation of your own investing process, not a robot that trades for you.

A as augmentation, and useful there. Reading 40 filings in an afternoon instead of four is a real edge for a serious investor. It isn’t a side hustle in the sense the rest of this list means, though. Your returns are capped by your capital, the outcome is volatile, and it gets emotional in a way that damages judgment right when judgment matters.

Right for: people who already invest seriously and have capital at work. If you’re here because you need income this quarter, this is the wrong list item.

The detail that stops most people from turning this into a business: the moment you charge anyone for advice about securities, you’re probably an investment adviser. The SEC’s framing is that an adviser is anyone who, for compensation, is engaged in the business of providing advice to others about securities, and unless an exemption applies you register with the SEC or a state regulator. “I built an AI stock screener and sell a subscription” walks straight into that definition. Talk to a securities lawyer before you take a dollar.

A tier: AI high-ticket consulting

Selling AI strategy and implementation to companies at real prices. In my experience that means low five figures per project or a monthly retainer.

A because the ceiling is high, the delivery is finite, and the leverage is your judgment, not your hours. Here’s the part people get wrong: businesses aren’t buying credentials. They’re buying a high-agency person who will brute-force their specific problem with these tools and won’t disappear when it gets ugly. Nobody has ten years of experience in something that’s three years old, and buyers know it.

Right for: senior operators. If you’ve shipped systems under real constraints, you already have the only qualification that matters.

The detail: pricing is where technical people torch this. Scoping AI work is different from scoping normal software work because the uncertainty sits in the model’s behavior, not the implementation, and hourly billing hands all of that risk to you. I wrote up how I estimate it in estimating client work when agents write the code. The same consultant can sit at the bottom or the top of the rate band based entirely on whether the engagement is scoped to an outcome.

A tier: AI digital products

Tools, templates, and small software products sold on your own store or a marketplace.

A, with one path that works and several that don’t. The path that works: build the tool you use to fulfill your service work, then sell it. You know the workflow is real because you run it daily, you know what breaks, and you already have the customer profile. The internal tool becomes the SaaS. The path that doesn’t work starts from “what product could I sell,” which is how you end up with 200 Notion templates and no buyers.

Right for: builders who already do service work. The service work is the market research.

The detail: your first version should be embarrassingly small. The agent loop that powers most useful AI products isn’t a large system. I’ve published a working one in 50 lines of Elixir. The real engineering is everything after the loop works: retries, idempotency, and durable state so a crashed job doesn’t double-charge a customer or lose a half-finished run.

A tier: AI automation agency

You build and operate automation systems for businesses. Lead intake and routing, document processing, reporting, agentic workflows that touch their real systems.

This is the highest-confidence A on the list, and it’s where technical skill matters most. Three reasons. It’s verifiable: three named workflows, delivered, working, done, which is a much easier sale than “AI strategy.” It scales, because the same workflow shapes recur across clients in a vertical. And the technical gap is enormous, because most of the competition is duct-taped no-code that works in the demo and falls over in week three.

The systems that survive have real data pipelines, real memory, a harness around the model, retries, and idempotency. That’s a backend engineering job, not a no-code feature list. n8n is fine for a demo. The ones that survive get rebuilt as real services with a durable runtime underneath, so a failed step resumes instead of silently dropping a client’s order.

Right for: technical founders. This is where being an engineer is worth the most money.

Two details a marketer won’t know. First, licensing: n8n is fair-code, not open source, and its Sustainable Use License limits you to using the software “only for your own internal business purposes or for non-commercial or personal use”. Hosting your clients’ workflows and credentials on an instance you operate reads to me as outside internal business use, and that’s a paid license conversation to have before you’ve got 30 clients on it. Second, your COGS isn’t zero. Each of my team’s $200-per-month AI coding seats consumes $1,800 to $3,500 per developer per month in API-priced tokens under heavy agentic use, and client-facing automations meter at API prices with no subscription arbitrage to hide behind. Price the tokens into the retainer or your best client becomes your least profitable one. If you’d rather have this built than build it, that’s what I do.

S tier: media and distribution

Building an owned audience. A channel, a list, a body of public work that means people know who you are before you ask them for anything.

S, and it’s not close. This is the only item on the list that makes every other item better. The automation agency with an audience doesn’t cold email. The digital product with an audience has a launch. The consultant with an audience gets inbound. Distribution is the one asset that compounds while you sleep and that nobody can deprecate.

Right for: everyone, eventually. People skip it because the payback period is brutal, my rough guess is three to eighteen months before it does anything measurable, depending on niche and consistency.

The detail: publish the specifics of work you actually do. The generic AI-tips account is saturated and worthless. The account that shows one real architecture decision, with the tradeoff and the thing that broke, builds trust that converts, because nobody can fake having done it. That’s also the only moat on this list that doesn’t erode when the next model ships.

Which one should you pick?

Pick based on which of the two scarce inputs you already have: technical ability or an audience.

Non-technical, no audience. Start with AI small business consulting. Walk into five local businesses, offer to document their process for free, and fix the worst hour of it. You’ll get paid within weeks, learn what businesses buy, and build the local trust that makes everything after it easier. Start publishing what you find at the same time.

Technical, no audience. AI automation agency, then productize it. Your engineering skill is worth the most per hour here, the sale is concrete, and the second or third client reveals the workflow you should turn into a product. Build your own tooling deliberately as you go. My Claude Code plugin stack is the setup I use to keep fulfillment work fast.

Technical, with an audience. Digital products and high-ticket consulting. The consulting funds the product, the audience de-risks the launch, and you already have the distribution the other two profiles are still building. Don’t start an agency in this position. You’d be giving away leverage you already paid for.

The one thing every idea depends on

All 13 of these are downstream of distribution, which is the honest reason the rankings above matter less than they look.

Two people run the identical automation agency. One cold emails 200 businesses a month and closes three. The other publishes twice a week about the systems they build and closes three from inbound while doing a quarter of the work. Same service, same skill, radically different business. The difference is whether anybody knows you exist.

So the practical sequencing isn’t “pick the best idea.” It’s: pick the idea that gets you paid soonest given what you already have, and start building distribution the same week, because in twelve months the distribution is worth more than the idea was.

Questions people search for

Can you make money with AI with no coding?

Yes, and small business consulting is the clearest path. The bottleneck for most local businesses isn’t technical difficulty, it’s that nobody has ever mapped their process, and the AI tooling needed is mostly configuration. The ceiling is lower than the technical paths because you’ll hit a wall the first time a client needs a real integration. Plan to partner with a builder at that point.

What’s the fastest AI side hustle?

Service work sold to local businesses, typically within a few weeks of starting. You’re selling a fix for a problem the buyer already knows they have, to someone who can decide on the spot, so there’s no audience to build and no algorithm to satisfy. Content paths like faceless channels and newsletters are slower by months because the audience has to exist before the money does. Speed to first dollar and ceiling trade against each other, so fastest is rarely best.

Is selling AI websites still worth it?

Not as a business, only as practice. The tools compressed the work to an afternoon and raised the quality floor at the same time, so the supply of competent website builders is effectively unlimited and the price reflects that. What’s still worth money is everything around the website: content, local search presence, review pipelines, and lead follow-up, because those need ongoing human work that doesn’t compress. Sell the outcome the website was supposed to produce.

Is the AI automation agency space saturated?

The marketing is saturated. The delivery isn’t. Plenty of people sell automation, and far fewer can build a system that survives six months of contact with a client’s real data, because that takes real error handling and durable state, neither of which comes from a no-code canvas. If you can build the durable version, the crowd that can’t is doing you a favor by making the category familiar to buyers.

How much do AI consultants charge?

In my own engagements and the quotes I see, independent AI consultants land in a wide band from low hundreds per hour up to several hundred, and it depends almost entirely on whether you’re selling advice or delivery. Project pricing in the low five figures is common for a bounded implementation, in my experience. Hourly billing underprices this work because it hands all of the model-behavior uncertainty to you, so scope to an outcome where you can.

Do you need an audience to make money with AI?

No, but everything is harder without one. Service businesses sold locally or through cold outreach work fine with zero audience, which is why they’re the right starting point for most people. An audience changes the economics: inbound leads instead of outbound, higher prices, shorter sales cycles, and a launch channel for anything you productize later. Start earning without one and build one in parallel.

What AI skills are worth learning?

The durable ones are system design around models, not prompt tricks. Specifically: how to build an agent loop and know when it’s stuck, how to make a workflow idempotent so a retry can’t double-charge anyone, how to manage state across long-running jobs, and how to evaluate output quality so you know when a change made things worse. Those transfer across every model release. Skills tied to one tool’s interface have a shelf life measured in months.

I build AI agents and automation systems for businesses, mostly in Go and Elixir, with the durability and error handling that keeps them running after the demo. If you want a system like that built, here’s how I work.