Build Your Own AI Tool? First Ask If It Needs the Internet
On February 3, Reuters reported a selloff in software and data stocks after Anthropic released plug-ins for Claude Cowork. LegalZoom fell 19.7%. RELX fell 14%, its biggest drop since 1988. CNBC reported a partial rebound on February 24, and Wedbush called the selloff "overblown." Wall Street was pricing a question you can ask of your own card statement: if AI can build this tool for me, why am I still renting it?
The pages ranking for that question all start with price. Smart Marketer, Sengo and James Devonport rank candidates by monthly cost and payment risk, and Devonport's decision test includes a cost threshold of about $50 to $100 a month. All three then default to hosted stacks, meaning web apps running on rented servers. My first question is where the tool will live. A job that only touches your own files belongs in a desktop tool on your own computer. A job that needs outside logins, customers or payments stays rented, or a professional builds it.
I have made this exact call for my own operation. I replaced Acrobat Pro and Descript/CapCut with tools I built in Claude Code. I use them every day on my own Windows machine, and I built them to run locally on purpose. I judge a new tool by what it exposes before I look at what it costs.
Price picks the candidate and says little about what the build exposes
Owners are already building. Retool's 2026 Build vs. Buy Report, published February 17, 2026, surveyed 817 builders in late 2025. 35% said their teams had replaced at least one SaaS tool (software you rent by the month) with a custom build. 78% plan to build more in 2026. Retool sells an app-building platform and counted its own customers in the sample. Read those figures as the enthusiast end of the market. The direction still matters for any owner holding a stack of subscriptions.
The best-documented small operator case comes from Smart Marketer, published September 7, 2026. A member, an ecommerce operator who is not a developer, replaced about ten subscriptions with tools Claude built for him. He estimates the savings at about $60,000 a year. That figure is his own estimate, reported as a case study. His stack was GitHub, Vercel and Neon, which covers code storage, web hosting and an online database.
Ethan Mollick described the shift in a post on X on January 9, 2026:
"There is a new category of software which is 'useful tool for me' which was previously either outside of people's ability or not worth the time, but which is now doable. It isn't just AI coding ability, it is the fact that it does it well without a lot of errors. Unexplored space"
Ethan Mollick, post on X, January 9, 2026
He has the category right. A "useful tool for me" serves one person, on one machine, doing one job. The ranking guides then build that tool as a public web app anyway. Choosing hosted signs you up for a login system, a public web address and an online database. The three pages ranking for this question do not frame the choice around where the tool lives. Those parts are where homemade apps leak.
Hosted means a login, a public address and a database, and that is where homemade apps leak
The login decides who gets in. The public address lets anyone on the internet find the front door. The database holds whatever your customers type into the app. Get one rule wrong and strangers can walk straight in.
On March 21, 2025, researcher Matt Palmer reported what became CVE-2025-48757, a public vulnerability record. He found 303 endpoints across 170 projects built with Lovable, an AI app builder, that lacked row-level security. Row-level security is the database rule that lets each user see only their own records. Without it, attackers needed no credentials at all. They could pull usernames, emails, phone numbers, payment status and API keys.
An API key is the password one piece of software uses to talk to another. A leaked key can let someone run up charges on your accounts. The CVE was published May 29, 2025, and Superblocks walked through the case in a March 9, 2026 analysis. For a small business, that list of exposed fields reads like a breach notice, and customer trust goes with it.
The code underneath carries its own problems. Veracode's 2025 GenAI Code Security Report, from October 2025, found that AI-generated code introduced security flaws from the OWASP Top 10, the industry's standard list of common weaknesses, in 45% of 80 coding tasks. I unpacked what that means for a small team in why AI coding multiplies the skill you already have. Addy Osmani described the pattern back in December 2024, in "The 70% problem":
"They can get 70% of the way there surprisingly quickly, but that final 30% becomes an exercise in diminishing returns."
Addy Osmani, "The 70% problem," December 4, 2024
On a hosted app, that final 30% includes the login rules, the database permissions and the security patches. A non-developer is least equipped to check those parts, and a stranger probes them first. Osmani also wrote:
"This cycle is particularly painful for non-engineers because they lack the mental models to understand what's actually going wrong."
Addy Osmani, "The 70% problem," December 4, 2024
When an app you can't diagnose breaks, you are left with code nobody on your team can read. On a hosted app, you are also left with a public front door that only that code controls.
A desktop tool on your own computer has no public door to leave open
My PDF desk runs on my own Windows machine. It merges and splits files, stamps page numbers, compresses documents and runs OCR in English and Spanish. OCR, short for optical character recognition, turns a scanned page into searchable text. The PDF desk has no login, no server and no web address. My video captioner follows the same design. It transcribes on my own graphics card, exports SRT and VTT caption files, burns captions into the video and reframes clips for vertical feeds, all locally.
The rest of my daily setup uses the same pattern. My content command center runs at a local address on my machine, and only my computer can reach it. My live meeting transcriber and my interview translator also transcribe on my own graphics card. They call Claude only for their Claude features. Some of these tools reach out to the internet, but no stranger can type an address and reach a login page for any of them.
Mollick flagged the assistant itself in "Claude Code and What Comes Next," published on One Useful Thing on January 7, 2026:
"It can read all your files and create new ones (PowerPoint and Word are just code, in the end, and Claude knows how to write code), access the web using your browser, write and execute programs for you, and more. Of course, AIs are not flawless and giving an AI access to your browser and computer creates all sorts of new risks and dangers."
Ethan Mollick, One Useful Thing, January 7, 2026
On a local build, the exposure comes from the AI assistant you build with, because it can read whatever folders you open to it. Point it at one project folder. Keep client records out of that folder while you build. With that boundary in place, the jobs below are good candidates for a tool on your own machine.
Ask who touches the data before you ask what the build saves
Retool's report also asked which software categories builders see as at risk of replacement. Look at who uses each category. Several of them are built around other people typing into them.
| Software category | Retool: at risk of replacement | Who touches the data |
|---|---|---|
| Workflow automations | 35% | Your team, plus the outside services and logins it connects |
| Internal admin tools | 33% | Your own staff |
| BI (reporting) tools | 29% | Your own staff, reading your own numbers |
| CRMs and form builders | 25% | Customers and strangers who fill in your forms |
| Project management | 23% | Your team, sometimes clients |
| Customer support | 21% | Customers |
Percentages: Retool 2026 Build vs. Buy Report, Feb 17, 2026, 817 builders (Retool sells app-building software). Right column: my read of who touches the data in a typical small business.
Picture a two-dentist practice paying for a PDF editor, a transcription app and a form builder. The PDF editor only works on files already on the front-desk computer. That makes it a candidate for a local build. The patient intake form collects health details from strangers on the internet. Keep renting it from a vendor that handles health-data compliance, or pay a professional to build it properly.
The transcription app depends on what goes into it. If the audio contains patient details, check the app's privacy page or ask the vendor where recordings are stored and processed before you replace anything. For that practice, the first money question is what a homemade tool would make it liable for. Savings come second. Run every subscription on your statement through the same questions, in order.
Smart Marketer and Sengo agree with question three. Smart Marketer says to leave payments, billing, payroll and tax alone. Sengo says to keep buying identity, payments, accounting and regulated data. Each still defaults to a hosted build.
Your Monday move depends on who does the work. If you build it yourself, put question one to every subscription on your card statement this week. If you delegate, give your developer the questions and ask for written answers before any build starts. If an agency pitches you a custom app, ask where it will run and who can reach it from the internet. A vague answer to that question is your answer.
Local, hosted or rented: the exposure gap is wider than the price gap
Exposure never appears on an invoice, and it stays invisible until something goes wrong. The table compares the options on the points that decide what you are liable for.
| Local desktop tool | Hosted homemade web app | Keep the subscription | |
|---|---|---|---|
| Who can reach the data | You, on your own machine | Anyone who finds the address, if one rule is wrong | The vendor's systems, under their contract |
| Login needed | None | You build and maintain one | The vendor runs it |
| What breaks it | Your computer or a system update | Expired keys, hosting changes, a missed security rule | A vendor outage or a price change |
| Who patches security | You, with no public entry point | You, with the whole internet able to probe it | The vendor |
| Good examples | PDF merging, video captions, file cleanup | A small team's internal dashboard, after a professional review | Payments, payroll, tax, patient intake, customer support |
Comparison drawn from building and running my own local tools. Not survey data.
The middle column carries the most risk because it pairs public exposure with your own patching. A local tool keeps the patching job and drops the exposure. A subscription hands the patching and the exposure to the vendor for the monthly fee. You choose which of those trades fits each job, one subscription at a time.
Run the new tool beside the paid one before you cancel anything
The Smart Marketer case gets this part right. The operator ran each replacement beside the paid tool for about two weeks. His attribution numbers matched the paid tool's within about 98%. Copy the method, whatever you build.
Retool's report asked how teams treat AI-written code. 44% test it thoroughly. 32% review it briefly, 8% put it to work immediately, 8% treat it as inspiration and 7% rewrite large portions. On a hosted app, that code faces the public internet from the first day.
The same report found that 60% had built something outside IT oversight in the past year. Most small businesses have no IT department to bypass. The problem is harder to spot there. If your staff are building tools on their own, start with a written AI policy for a small team before the first homemade app touches customer data. Pierre Yves Calloc'h of Pernod Ricard gave the enterprise view in the same Retool report:
"There's no way you can go live with a vibe-coded solution."
Pierre Yves Calloc'h, Pernod Ricard, quoted in Retool's 2026 Build vs. Buy Report, February 17, 2026
Small business owners are learning AI the hard way, by trying it. Heartland Forward and Echelon Insights published "AI on Main Street" on October 1, 2026, based on 691 owners surveyed September 3 to 11. 43% of small businesses use AI. 70% of those users report a net-positive effect on their finances. 83% of AI-using owners taught themselves by trying it, and only 12% took a class. Learning by trying is fine for a PDF tool on your own desk. It gives you a thin base for a login system that holds customer data.
Track the right things during the trial. Count how often you fall back to the paid tool, and log the hours you spend fixing the new one. Note whether the output matches on the same inputs. The number of subscriptions you cancelled is a vanity metric. It looks good in a screenshot and tells you very little about whether the work still gets done safely.
Frequently Asked Questions
Can I replace my paid software with something I build using AI?
For some jobs, yes. Jobs that only touch files on your own computer, like merging PDFs or captioning videos, are good candidates for a desktop tool on your own machine. Jobs that need outside logins, customer records or payments should stay with a vendor or be built by a professional. Run the new tool beside the paid one for about two weeks before you cancel.
Is an app built with AI safe to use with customer data?
Only if someone who understands security has checked the login rules and the database permissions. In CVE-2025-48757, at least 170 Lovable-built apps exposed emails, phone numbers, payment status and API keys because a database rule was missing. Attackers needed no login to get the data. If customer, patient or payment data is involved, rent from a vendor that carries the compliance, or hire a professional to build and review it.
Which subscriptions should a small business keep paying for?
Keep anything that handles payments, payroll, tax, health records or customer logins. Those tools need security patching, compliance work and someone on call when they break. Tools that only process files already on your computer are the safest to replace with your own build. Ask who touches the data before you compare prices.
The price on a subscription is the easy part of this decision because it is printed on the invoice. The harder part is whether your homemade version will need a login, a public address and a database, and who will guard them. If you are weighing which tools to cut, or how AI fits into the way customers find your business, book a call with me and we will go through it together. Some of the subscriptions on your card statement are the cheapest security you own.
