Skip to content
Day 69: new resources are added every day. See what's new

New ideas every day

AI and tech business ideas from today's news

280 business ideas, each grounded in a trending AI, tech, robotics or hardware story on X, with honest ratings, pros and cons and a verdict. 20 new ideas are added every morning. Last updated 16 November 2026.

Built with love by Shushant Lakhyani. Building bestproductivityresources.com

Browse business ideas

Search, filter and sort by rating, cost, difficulty, business model and date.

Category

Showing 20 of 280 matching ideas (280 in total)

AI securityAdded 16 November 2026

Connection trimming and token scoping audits for agents wired to many MCP servers

Review which tool servers an AI agent is connected to, remove the ones it does not need, and shrink what each token can do.

6.8/ 10

A VentureBeat guest article attributes to Palo Alto Networks' Unit 42 a 78.3 percent attack success rate in testing when five MCP servers are connected to one agent, because injected instructions carry into calls to servers that were never compromised. It also points to the OWASP MCP Top 10, which ranks token mismanagement and secret exposure highest. Most teams keep adding servers because it is easy. Sell a fixed audit that lists every connection, cuts the extras, issues short-lived least-privilege tokens and writes a one page rule for adding new ones.

Target customer
Companies running internal or customer-facing AI agents with several tool connections
Business model
Consulting
Startup cost
Under $1k
Difficulty
Medium
Time to launch
Under 1 month
Overall rating
6.8 out of 10
Market demand
7/10
Competition (10 = open field)
6/10
Profitability
7/10
Ease of execution
6/10
Trend momentum
8/10

Pros

  • Security buyers already understand access reviews
  • The audit method works across agent platforms
  • The research gives you a concrete risk to explain

Cons

  • The 78.3 percent figure comes from one test setup and may not match a client's
  • Needs real security and identity skills
  • Agent platforms may soon bundle these checks

Verdict

Strong for security or identity consultants. Start with a flat-fee audit of one agent, quote the research with its limits, and leave clients with a short rule they can follow without you.

LLMsAdded 16 November 2026

Private hosted trials of giant open-weight agent models for teams wary of API lock-in

Run a time-boxed private trial of a freshly released open-weight agent model on a client's own tasks, so they can compare it with their paid API before committing.

6.6/ 10

Meituan released the weights of LongCat-2.5, a 1.6 trillion parameter mixture-of-experts agent model, under an MIT licence that allows commercial use, with a 4-bit version and an API that speaks the OpenAI and Anthropic formats. The sample self-hosting guide needs 194 accelerators, which almost no company will buy just to test a model. Sell a fixed-fee trial: you rent the capacity, run the client's real coding or agent tasks, and hand back a plain report on quality, cost and what it would take to host it permanently.

Target customer
Software and operations teams that want leverage against their current AI provider
Business model
Consulting
Startup cost
$1k to $10k
Difficulty
Medium
Time to launch
Under 1 month
Overall rating
6.6 out of 10
Market demand
7/10
Competition (10 = open field)
6/10
Profitability
6/10
Ease of execution
6/10
Trend momentum
8/10

Pros

  • The MIT licence removes the usual legal question for a trial
  • Buyers already ask whether an open model could replace a paid one
  • A short fixed-fee project is easy to explain and price

Cons

  • Renting enough capacity for a giant model can eat the margin
  • Benchmark figures came from the maker, so clients will want their own proof
  • New models arrive weekly, so each trial template goes stale fast

Verdict

Good for a small team with MLOps experience. Cap each trial at one or two weeks, price compute at cost plus a fee, and be blunt when the answer is that the paid API is still the better buy.

Business & marketsAdded 16 November 2026

Product data cleanup so shopping agents read a maker's catalogue correctly

Fix a brand's product titles, specs and identifiers across every listing so AI shopping tools and giant product databases describe the items right.

6.4/ 10

Amazon is reportedly building a World Product Graph, a database covering virtually every item sold online, including items it does not sell, with makers such as Coca-Cola and Bayer reportedly supplying data. Amazon told Reuters the document is a draft brainstorming paper and it is unclear whether or when the project will launch. If any of this ships, rival shopping agents could also pull from that data, so messy listings would spread everywhere. Offer a catalogue audit that standardises names, specs, images and product codes and keeps one clean master sheet.

Target customer
Small and mid-sized brands that sell the same products across many sites
Business model
Agency / services
Startup cost
Under $1k
Difficulty
Easy
Time to launch
Under 1 month
Overall rating
6.4 out of 10
Market demand
7/10
Competition (10 = open field)
5/10
Profitability
6/10
Ease of execution
7/10
Trend momentum
7/10

Pros

  • Clean product data helps every sales channel, whatever Amazon does
  • Easy to demo with a before and after of a real listing
  • Recurring work as catalogues change

Cons

  • The project is only reported and may never launch
  • Many feed and catalogue tools already claim to do this
  • Brands with huge catalogues will want software, not a manual service

Verdict

A sensible service to start small. Sell the cleanup as useful in its own right, treat the Amazon report as a reason to move now rather than a promise, and automate the checks as you grow.

ChipsAdded 16 November 2026

Camera redesign cost studies for product makers eyeing single-chip AI image sensors

Show device makers what moving from a three-part camera design to one programmable AI sensor would change in cost, power and software before the chips ship.

6.2/ 10

Bharat Semi raised a $24 million seed round for programmable image sensors that combine capture, processing and AI on one chip, and says this could cut system costs by at least 70 percent against a separate sensor, image processor and system-on-chip. First-generation parts are still in prototyping, with commercial production targeted for 2027. Product makers in smart home, industrial and drone markets need to decide now whether to wait, so sell a study that models their bill of materials, power budget and software rework under both paths.

Target customer
Hardware product teams in smart home, industrial automation and drones
Business model
Consulting
Startup cost
Under $1k
Difficulty
Medium
Time to launch
1 to 3 months
Overall rating
6.2 out of 10
Market demand
6/10
Competition (10 = open field)
7/10
Profitability
6/10
Ease of execution
5/10
Trend momentum
7/10

Pros

  • Product teams plan chips years ahead, so early clarity has real value
  • Needs little more than spreadsheets, datasheets and engineering experience
  • Works for any sensor family, not just one startup

Cons

  • The 70 percent saving is the company's own claim and the chips are not in production
  • Buyers with the budget may already have in-house hardware analysts
  • A vendor delay would leave your study without a decision to support

Verdict

Best for an ex hardware engineer. Present three scenarios (stay, switch, wait), mark every vendor claim as unverified, and refresh the study when real samples appear.

Business & marketsAdded 16 November 2026

Junior role redesign for teams where AI now does the starter tasks

Rebuild entry-level jobs around review, checking and client work so companies still grow future seniors while AI handles the easy tasks.

6.2/ 10

An analysis by the Reserve Bank of Australia found young workers aged 15 to 24 in the most AI-exposed jobs have weaker employment than older workers, and that job postings have fallen most in the highest-exposure group. The bank says it cannot prove AI is the cause. Any company that stops hiring juniors will have no mid-level staff in a few years. Offer a short engagement that maps which starter tasks AI now does, then designs apprentice-style roles and a training path that make juniors useful from week one.

Target customer
Companies in software, support and admin-heavy fields that have frozen junior hiring
Business model
Consulting
Startup cost
Under $1k
Difficulty
Medium
Time to launch
Under 1 month
Overall rating
6.2 out of 10
Market demand
6/10
Competition (10 = open field)
6/10
Profitability
6/10
Ease of execution
6/10
Trend momentum
7/10

Pros

  • The pipeline problem is real and easy for managers to grasp
  • Needs mostly workshops and templates, not software
  • Can be repeated across industries

Cons

  • The research shows a link, not proof that AI caused it
  • Buyers may prefer to keep cutting costs
  • Results show up over years, which makes success hard to prove

Verdict

Worth trying for an experienced HR or team lead. Pitch it as protecting the talent pipeline, show the study's caveats up front, and offer a pilot with one team.

Business & marketsAdded 16 November 2026

Reseller and cloud partner due diligence for AI chip buyers and funders

Check the intermediaries behind a big AI chip order so buyers and lenders know who really receives the hardware.

5.8/ 10

The Wall Street Journal reportedly reviewed documents showing Nvidia cloud partner Megaspeed brokered about $12 billion in high-performance chip orders for Chinese companies between March 2024 and June 2025. Megaspeed calls its business legitimate and lawful and has hired outside counsel for an internal investigation. Whatever the outcome, lenders, cloud buyers and investors now have a reason to ask who stands behind a chip broker. Sell a plain-language review of ownership, customers and export-control exposure for any intermediary in an AI hardware deal.

Target customer
Banks, data center investors and companies buying GPUs through resellers
Business model
Consulting
Startup cost
$1k to $10k
Difficulty
Hard
Time to launch
3 to 6 months
Overall rating
5.8 out of 10
Market demand
6/10
Competition (10 = open field)
5/10
Profitability
7/10
Ease of execution
4/10
Trend momentum
7/10

Pros

  • Large deal sizes support high fees
  • Compliance work tends to be repeat business
  • The topic is getting more scrutiny, not less

Cons

  • Needs legal and trade-control expertise that is hard to fake
  • Big law and compliance firms already sell similar work
  • The allegations are reported and denied, so avoid claiming wrongdoing

Verdict

Only for people with trade compliance experience. Stay factual, work from public records, and add a lawyer to the team before writing a single client report.

HardwareAdded 16 November 2026

Wi-Fi buying and placement checks for small offices facing Wi-Fi 8 hype

Tell small offices whether to buy now, wait, or just move the router, before they pay for draft-spec hardware.

5.8/ 10

Engadget reports Wi-Fi 8 (802.11bn) focuses on reliability, with final approval expected in 2028, while ASUS and TP-Link have already announced routers based on the draft spec. Both the router and the device need Wi-Fi 8 to see the benefits, and the article says Wi-Fi 7 remains a reasonable choice today and that moving a router to a central spot will likely help more than any new standard. Offer a flat-fee site check that maps dead spots, advises on kit and placement, and sets up a mesh only if the building truly needs one.

Target customer
Small offices, clinics and shops with patchy wireless
Business model
Agency / services
Startup cost
Under $1k
Difficulty
Easy
Time to launch
Under 1 month
Overall rating
5.8 out of 10
Market demand
6/10
Competition (10 = open field)
4/10
Profitability
5/10
Ease of execution
8/10
Trend momentum
6/10

Pros

  • Almost any office has some Wi-Fi pain
  • Honest advice to wait builds trust
  • Low tool cost: a laptop and a survey app

Cons

  • Many local IT shops already do this
  • Low ticket size per job
  • Hardware changes could date your advice

Verdict

A simple local-services play with a little hype to ride. Charge a flat fee, be willing to say do nothing, and keep a short checklist you can update when the Wi-Fi 8 spec firms up.

AI infrastructureAdded 16 November 2026

Mixed-chip training cost audits for self-driving and robot AI teams

Help teams that train driving or robot models find which jobs can run on mid-tier or alternative chips so one supply problem does not stop them.

5.8/ 10

At Pony.ai, the CFO said demand for compute far exceeds supply, and the company is training models more efficiently, sometimes on midtier compute, and turning to domestic chips. He said companies have prepared for tighter Nvidia export controls for a couple of years and that alternatives exist, though they are not yet the most cost-effective. Many teams without Pony.ai's resources have no plan B. Sell an audit that sorts training jobs into those that need top chips and those that can move, with a price and timing estimate for each.

Target customer
Autonomous vehicle and robotics companies with growing training bills
Business model
Consulting
Startup cost
$1k to $10k
Difficulty
Hard
Time to launch
3 to 6 months
Overall rating
5.8 out of 10
Market demand
6/10
Competition (10 = open field)
6/10
Profitability
6/10
Ease of execution
4/10
Trend momentum
7/10

Pros

  • Compute shortages are an urgent budget issue
  • Savings are measurable, so fees are easy to justify
  • Applies beyond driving to any heavy training workload

Cons

  • Needs deep machine learning and systems skills
  • Clients may not share workloads with outsiders
  • Alternatives change fast and the article says they are not yet the most cost-effective

Verdict

Suited to a former ML infrastructure engineer. Start with one client's biggest job, measure real cost and time on a second chip type, and report numbers instead of promises.

Generative mediaAdded 16 November 2026

Stroke-by-stroke replay videos and prints for AI-painted artwork

Turn the full stroke log of a code-driven painting into a replay video and a limited print that carries its own record of how it was made.

5.6/ 10

Anthropic's Claude Fable 5.2 reportedly paints rather than generates images: it writes a Python program of roughly 1,000 to 2,000 lines that drives a simulated paint engine, with no external image model and no traced source images. One example scene used about 700,000 strokes, and the model has no undo, so every stroke is committed. That leaves a complete, replayable record of each artwork. Package the log as a timelapse video plus a print with a scannable link to the replay, for artists, galleries and brands that want a story and proof of process.

The news behind it

Anthropic's Claude Fable 5.2 doesn't generate images, it paints them

The Decoder · 15 November 2026

Target customer
Digital artists, small galleries and brands commissioning AI artwork
Business model
Marketplace
Startup cost
Under $1k
Difficulty
Medium
Time to launch
1 to 3 months
Overall rating
5.6 out of 10
Market demand
5/10
Competition (10 = open field)
6/10
Profitability
5/10
Ease of execution
6/10
Trend momentum
6/10

Pros

  • The stroke log is a built-in proof of process that image generators cannot offer
  • Replay videos are cheap to render and good for social posts
  • Print on demand keeps stock and cash risk near zero

Cons

  • Buyers may not pay extra for process proof
  • Earlier painting attempts by a sibling model drew criticism for imitating existing art
  • The capability is new and access details were not in the article

Verdict

A small, creative experiment, not a safe income plan. Test with a handful of pieces, be clear about what the model did, and stay away from copying living artists' styles.

Open sourceAdded 16 November 2026

Setup and hosting for open-source e-book lending boxes for small libraries and schools

Install and host an open-source lending system so small libraries can own the e-books they lend instead of renting licences.

5.0/ 10

The magazine Flaming Hydra is running a $30,000 Kickstarter that includes Lenny for Libraries, an open-source library-in-a-box e-book lending system that lets libraries buy and own e-books rather than license them. The campaign also plans to turn paywalled newsletters into EPUB files for lending. If the code ships, small libraries and schools will not have the staff to run it. Offer setup, hosting, catalogue loading and support for a monthly fee, and help them source e-books they can actually own.

Target customer
Small libraries, schools and community archives
Business model
SaaS subscription
Startup cost
Under $1k
Difficulty
Medium
Time to launch
3 to 6 months
Overall rating
5.0 out of 10
Market demand
4/10
Competition (10 = open field)
8/10
Profitability
3/10
Ease of execution
5/10
Trend momentum
5/10

Pros

  • Open source means no licence fee for the core system
  • Libraries value ownership and long-term access
  • Hosting and support are steady recurring work

Cons

  • It is only a crowdfunding campaign and the software may slip or change
  • Library budgets are small and slow to move
  • Content supply depends on publishers willing to sell ownable copies

Verdict

A niche mission-driven business, not a fast earner. Wait until the code is released and tested, run one pilot with a friendly library, and price for low volume.

LLMsAdded 15 November 2026

Cost-per-solved-task model routing for maths-heavy work

Help teams pick the cheapest model that still solves their hard reasoning tasks instead of buying the top benchmark score.

6.8/ 10

On the MathArena benchmark, Google's Gemini 4 Carbon took first place at 85.4% with an expected cost of $15.70, while OpenAI's GPT-6.1 Sol scored 80.6% at about $1.07. The article puts the cost gap at roughly 15 times for a few points of accuracy. Offer a service that tests a client's own hard problems across models, then sets rules for when to use the cheap model and when to pay for the expensive one.

Target customer
Teams running AI on maths, finance, science or engineering problems
Business model
Consulting
Startup cost
Under $1k
Difficulty
Medium
Time to launch
Under 1 month
Overall rating
6.8 out of 10
Market demand
7/10
Competition (10 = open field)
6/10
Profitability
6/10
Ease of execution
7/10
Trend momentum
8/10

Pros

  • The price gap is large and easy to explain
  • Savings show up on the next invoice
  • Can be packaged as a fixed-fee audit

Cons

  • Model rankings change every few weeks
  • Clients must share sample tasks
  • Routing tools may do this for free

Verdict

A solid starter service. Charge for the test and the rules, and keep your results tied to the client's own tasks, not public leaderboards.

Business & marketsAdded 15 November 2026

Compute commitment stress tests for AI companies buying capacity years ahead

Model what happens to cash if revenue grows slower than the compute a company has already agreed to buy.

6.8/ 10

At a JPMorgan event, Anthropic CEO Dario Amodei said it must buy compute one to two years ahead, before knowing if revenue will cover it. Bloomberg reported nearly $901 billion in commitments, with about $71 billion due by September 2027. Smaller AI companies face the same trap at a smaller scale, so offer scenario models that show when growth, margins and contract terms break a plan, and what to renegotiate.

Target customer
Finance leads and founders at AI startups with large cloud or GPU contracts
Business model
Consulting
Startup cost
Under $1k
Difficulty
Medium
Time to launch
Under 1 month
Overall rating
6.8 out of 10
Market demand
7/10
Competition (10 = open field)
6/10
Profitability
7/10
Ease of execution
6/10
Trend momentum
8/10

Pros

  • Compute is a top cost line for AI firms
  • A sharp, fixed-fee deliverable
  • Spreadsheet models are cheap to build

Cons

  • Needs finance and cloud contract skills
  • Clients must share sensitive numbers
  • Auditors and fractional CFOs compete

Verdict

Strong for a finance person who understands cloud contracts. Sell a two-week model plus a renegotiation checklist.

Business & marketsAdded 15 November 2026

Exit kits for customers of AI apps that might shut down

Sell a simple pre-mortem and data-export kit so teams can leave an AI tool without losing work.

6.6/ 10

TechCrunch's running AI graveyard lists tools that closed or were pulled, including Relay in August 2026 and the family assistant Ollie on November 6, plus Character.ai's bankruptcy filing in early October. It also cites S&P Global Market Intelligence's estimate that about 42% of AI initiatives are abandoned by their corporate parents. Teams adopting new AI tools rarely plan their exit, so offer a checklist, export tests and a backup tool plan before they depend on a vendor.

Target customer
Small and mid-sized teams adopting several new AI tools
Business model
Consulting
Startup cost
Under $1k
Difficulty
Easy
Time to launch
Under 1 month
Overall rating
6.6 out of 10
Market demand
7/10
Competition (10 = open field)
6/10
Profitability
5/10
Ease of execution
8/10
Trend momentum
7/10

Pros

  • Real shutdowns make the risk easy to explain
  • Needs little more than process know-how
  • Can be sold as a recurring review

Cons

  • Buyers only feel the pain after a shutdown
  • Big vendor-risk firms may cover it
  • Hard to prove value when nothing fails

Verdict

Easy to start and easy to sell after any new shutdown headline. Keep it practical: export tests, backups and a named fallback for each tool.

Consumer AIAdded 15 November 2026

Ownership cost calculators for people eyeing a self-driving car with a monthly fee

Build a simple tool and guide that shows what a driverless personal car really costs once the monthly operating fee and service-area limits are counted.

6.4/ 10

Tesla opened pre-orders for its Cybercab for personal use at $29,990, but owners must also pay $299 a month to operate it, and the car runs unsupervised only in the same geofenced areas as Tesla's robotaxi service. Musk called the fee too high and said it should be closer to about $115 a month, so the real price is still moving. Buyers will want to compare owning against ride-hailing and renting, so sell a calculator and a short decision guide that updates as the terms change.

Target customer
Early adopters and families weighing a driverless personal car
Business model
Content + affiliate and paid reports
Startup cost
Under $1k
Difficulty
Easy
Time to launch
Under 1 month
Overall rating
6.4 out of 10
Market demand
6/10
Competition (10 = open field)
6/10
Profitability
5/10
Ease of execution
8/10
Trend momentum
7/10

Pros

  • Pricing is public and still shifting, so people need help
  • Cheap to build and easy to update
  • Works for any future driverless car offer

Cons

  • Terms may change fast and make the guide stale
  • Buyers can use free comparison tools
  • Only a small group is eligible to pre-order today

Verdict

A quick, low-cost test of demand. Build it as a living page, date every number, and widen it to other driverless cars as they launch.

ChipsAdded 15 November 2026

Buyer briefs for AI teams weighing wafer-scale inference hardware

Sell independent briefs that compare specialist inference chips with GPU capacity for teams deciding where to run their models.

6.4/ 10

Cerebras reported that Q3 hardware sales fell 28.8% to $68.9 million while its revenue backlog reached $26.1 billion, most of it tied to an OpenAI deal. It also says its next CS-5 system enters production in early summer 2027 and that it has added another hyperscaler. Teams outside those big deals cannot tell when specialist chips beat GPUs on cost, supply and software fit, so offer a plain-English comparison with a short test plan.

The news behind it

Cerebras Readies CS-5 Systems For 2027 Launch As It Bags Another Hyperscaler

The Next Platform · 14 November 2026

Target customer
AI startups and enterprise platform teams planning inference capacity
Business model
Consulting
Startup cost
$1k to $10k
Difficulty
Medium
Time to launch
1 to 3 months
Overall rating
6.4 out of 10
Market demand
6/10
Competition (10 = open field)
7/10
Profitability
6/10
Ease of execution
6/10
Trend momentum
7/10

Pros

  • Chip choice is a real cost decision
  • Few neutral voices explain the trade-offs
  • Reports can be resold to several teams

Cons

  • Needs hands-on hardware knowledge
  • Vendors may not share test access
  • Roadmaps change, so briefs age quickly

Verdict

Good for an engineer who already knows inference stacks. Start with one workload type and publish a clear method so buyers trust the numbers.

LLMsAdded 15 November 2026

Prompt habit refresh sessions for teams moving to Claude Opus 5.5

Run short workshops that strip out old prompting habits and set up effort levels, one-message briefs and file inputs for the new model.

6.4/ 10

Anthropic says Opus 5.5 is now the default on Pro, Max, Ultra and Team plans and that old habits, like writing think hard, are holding it back. It suggests starting at Medium effort, which matched Opus 5 on High in its tests, and describing the goal, format and when to ask questions in one message. Many teams still prompt the old way, so sell a half-day session plus a team prompt guide and a before and after test.

The news behind it

Anthropic says outdated prompting habits are slowing down Claude Opus 5.5

The Decoder · 14 November 2026

Target customer
Teams on paid Claude plans who rely on AI for daily work
Business model
Agency / services
Startup cost
Under $1k
Difficulty
Easy
Time to launch
Under 1 month
Overall rating
6.4 out of 10
Market demand
7/10
Competition (10 = open field)
5/10
Profitability
5/10
Ease of execution
8/10
Trend momentum
7/10

Pros

  • Anthropic itself says habits need to change
  • Easy to run remotely and repeat
  • Tied to real time and cost savings

Cons

  • Anthropic and others publish this advice for free
  • Advice goes stale with each release
  • Hard to prove the before and after gain

Verdict

Works as a quick entry offer. Bring measured before and after examples from the client's own tasks, and update the guide with every model release.

AI researchAdded 15 November 2026

Chart-reading briefs for investors studying AI company roadshow slides

Turn dense compute and efficiency charts into plain notes about what they do and do not prove.

6.2/ 10

A former researcher, Jacob Coxon, says an Anthropic roadshow slide shows efficiency gains accelerating and that Mythos could be retrained with a fraction of the original compute. The slide's compressed axis makes a 10x and a 100x increase look alike, and Anthropic says its data points are internal estimates for illustration only. Retail and smaller investors have little help reading such charts, so sell short briefs that flag axis choices, missing units and claims that need outside evidence.

Target customer
Independent investors, analysts and newsletter writers covering AI
Business model
Subscription newsletter
Startup cost
Under $1k
Difficulty
Medium
Time to launch
Under 1 month
Overall rating
6.2 out of 10
Market demand
6/10
Competition (10 = open field)
6/10
Profitability
5/10
Ease of execution
7/10
Trend momentum
7/10

Pros

  • Big IPO coverage drives attention
  • Skill transfers to every future filing
  • Low cost to start

Cons

  • Cannot give investment advice
  • Claims rest on public slides only
  • Audience may prefer free commentary

Verdict

Fine as a niche newsletter. Stay on method, not predictions, and always separate what a chart shows from what a person says it means.

Business & marketsAdded 15 November 2026

Succession readiness reviews for founder-led tech companies

Test whether a company can keep customers and partners calm if its leader leaves suddenly.

6.2/ 10

Nvidia CEO Jensen Huang resigned on November 12, and a petition on standwithjensen.com had passed 115,000 signatures by Saturday, with thousands of current employees and shareholders among them. CFO Colette Kress is interim CEO while a search runs, and the stock fell sharply during the week. Few founder-led firms have a tested plan for a sudden exit, so offer a review of decision rights, board processes, customer messaging and a first-72-hours checklist.

Target customer
Boards and executive teams of founder-led tech companies
Business model
Consulting
Startup cost
Under $1k
Difficulty
Medium
Time to launch
1 to 3 months
Overall rating
6.2 out of 10
Market demand
6/10
Competition (10 = open field)
6/10
Profitability
6/10
Ease of execution
6/10
Trend momentum
7/10

Pros

  • A very public example makes the risk real
  • Reviews can be priced as a project
  • Reusable templates keep delivery light

Cons

  • Boards may avoid the topic until a crisis
  • Senior sales cycles are slow
  • Governance firms already offer similar work

Verdict

Suits an ex-executive or governance adviser. Keep it about process, never about the people involved.

RoboticsAdded 15 November 2026

Trial benchmarks for robot makers testing general AI models as controllers

Run side-by-side tests of general-purpose models against specialist robot models on a client's own tasks.

5.8/ 10

OpenAI moved its robotics researchers into its main research and applied teams, reportedly because general models like GPT-6 Astra may control robots directly. In simulated DROID tasks an Astra agent hit 74.8% against 51% for the best specialist models, but on a desk-object test with dual-arm robots it fully completed only 7 of 100 tasks. Small robot makers cannot tell which result applies to them, so offer a repeatable trial on their hardware and tasks.

Target customer
Robot makers and warehouse or lab automation teams
Business model
Consulting
Startup cost
$10k to $50k
Difficulty
Hard
Time to launch
3 to 6 months
Overall rating
5.8 out of 10
Market demand
6/10
Competition (10 = open field)
7/10
Profitability
5/10
Ease of execution
4/10
Trend momentum
7/10

Pros

  • The two test results differ sharply
  • Few neutral test providers exist
  • Reusable test rigs lower later costs

Cons

  • Needs robots, space and safety rules
  • Clients may want vendor-run tests instead
  • Model access terms can limit what you publish

Verdict

Best for someone with a robotics lab already. Start with one arm and five common tasks, and publish the method openly.

RoboticsAdded 15 November 2026

Adaptation service for robot makers using open-weight robot control models

Fine-tune and test an open robot model on a client's arm or gripper so they skip the data and GPU work.

5.8/ 10

Runway released its Praxis-1 robot control model with open weights under Apache 2.0, which allows commercial use. Adapting it to a Franka arm used about 277 hours of DROID recordings, the standard version needs an 80 GB GPU, and Runway admits it can confuse similar-looking objects. Many small robot makers lack the data pipeline and GPUs, so offer a fixed-scope adaptation, test and report service.

The news behind it

Runway's open-source Praxis-1 turns video AI into a robot controller

The Decoder · 14 November 2026

Target customer
Small robot makers and research labs without ML teams
Business model
Agency / services
Startup cost
$10k to $50k
Difficulty
Hard
Time to launch
3 to 6 months
Overall rating
5.8 out of 10
Market demand
6/10
Competition (10 = open field)
6/10
Profitability
6/10
Ease of execution
4/10
Trend momentum
7/10

Pros

  • Open licence means no vendor deal is needed
  • Clear scope: data, tune, test, report
  • Real hardware results are valuable to buyers

Cons

  • Needs robots and GPU rental budget
  • Results are not independently verified
  • Runway or others may offer official partners

Verdict

A strong niche for ML engineers with robot access. Price per robot type and be upfront about the limits the model has on long tasks.

These ideas are written for research and inspiration. Ratings are our own judgement, not data, and nothing here is financial, legal or investment advice or a promise of income. Each idea links to the public news item that inspired it; check the original and do your own research before spending money. Product and company names belong to their owners, and we are not affiliated with them.

18 free downloads

Free databases, guides and hacks

Everything I've collected, free on Gumroad. Pay what you want, or nothing at all.

See all 18