AI coding tools such as Codex and Claude Code have made tightly scoped custom software for small business faster to build. A focused internal tool can sometimes take 10 active work hours. A larger first release may take 40 or 80. At Website Genii’s standard rate of $200 per hour, those examples create useful planning points from $2,000 to $16,000 before extra security, data, or support.
Generating code is only one part of the job. With a capable coding agent, the longest stretch of elapsed time may be waiting for a carefully planned run to finish. Human time still goes into understanding the workflow, revising the plan, making client decisions, reviewing the output, testing failures, launching safely, and maintaining the tool. AI can do a remarkable amount of busy work. People remain responsible for the result.
What Can a Small Business Build Now?
Custom software is any tool shaped around the way one business works. It can be a small feature on a website or the main system a team uses every day.
Here is a simple range:
- Simple: a public calculator, searchable directory, research organizer, quote helper, or dashboard that displays non-sensitive data.
- Focused: an invoice sorter, lead review screen, job-photo organizer, inventory checker, or approval workflow used by a few employees.
- Connected: a reporting portal that collects information from several services through an API. An API, or application programming interface, is a standard way for software systems to share data.
- Complex: a multi-user customer portal or CRM with permissions, history, email, files, reporting, and several integrations. A CRM, or customer relationship management system, stores contacts, sales activity, and customer history.
A sales competition dashboard is a good example. A basic version could pull approved sales totals, show team rankings on a TV, and let employees view the same board from home. Add private commissions, manager-only views, live CRM updates, alerts, and rewards, and the same idea becomes a much more serious application.
Complexity rises when software gains more users, sensitive data, outside connections, or power to change records. A read-only research tool is not the same security job as a system that can email customers, move money, or edit orders.
Real Small Businesses That Built Useful Tools
The examples below come from business-owner accounts and vendor case studies. They show what is possible, but their time and savings figures are self-reported, not independent guarantees.
A Public Locator Built in an Afternoon
Christian Ortega of The Original Tamale Co. had never written code. He used ChatGPT to build and launch a searchable farmers-market locator in one afternoon, according to OpenAI’s small-business story. It is a strong first project because it is useful without needing customer accounts or private business records.
A Contractor’s Invoice Sorter
Builder Morgan Sterling used ChatGPT to create an email-connected script that reads invoices, matches them to roughly 30 active Beach Cities Builder jobs, and files them in Google Drive. OpenAI’s contractor case study says the first version took about two hours and saves five to six hours weekly. Because it touches email and financial documents, it needs limited permissions, logs, and human review when unsure.
A Local CRM for a Training Academy
Spatial Thoughts had participant data across Google Sheets, WordPress, Stripe exports, and Eventbee reports. Its owner used Claude Code in an afternoon to join that data and build a searchable private CRM, according to the founder’s account. It runs on one computer without hosting, logins, or multiple users. That boundary removes work, although the device, exports, and backups still need protection.
A Fleet Prototype That Stays Away From Real Data
At five-person fleet startup Proaction, non-engineer cofounder Colin Knudsen uses Codex to turn customer conversations into working demos. The Proaction story says they use sample data in separate test environments while engineers build the secure production version. The team can prove a workflow without putting real fleet or customer data into the experiment.
A Client Reporting Portal for an Accounting Firm
Officeheads built a Softr and Airtable portal used by more than 50 small-business clients. Its vendor case study says the first version took under two days. It includes calendars, documents, project status, and financial dashboards updated monthly. Established tools sped up the interface, but client permissions and the accuracy of connected data still matter.
A Field-Sales Platform Without an In-House Developer
Three CPF Floors employees without an in-house developer used Glide to build a field-sales app for customer and product data, pricing, commissions, purchase orders, and approval rules. Glide’s customer story says it launched in under two months and processed 13,000 orders in eight months. That shows usage, not proven new revenue, and makes testing prices, permissions, and changes essential.
Custom Systems and Software Website Genii Has Built
Website Genii’s own work ranges from focused workflow layers to complete business platforms. Current AI coding tools now help us plan, build, test, and release improvements much faster.
- LeadDecide is a multi-client lead-review application. It imports leads from CallTrackingMetrics and CallRail, shows AI-assisted qualification suggestions, records reviewer decisions and job outcomes, and, with opt-in writeback, can sync mapped outcomes to CallTrackingMetrics Reporting Tags. It is a focused layer, not a replacement CRM.
- Sites for Scouts is an all-in-one troop platform with websites, trip management, calendars, member roles, family access, and scout-account bookkeeping. Its first code was written manually, and AI-assisted development has since accelerated new releases.
- AEO Blocks is a WordPress plugin we built to help create AI-assisted FAQs across Gutenberg, Divi, and other website-building approaches. AEO, or answer engine optimization, means arranging useful content so AI-powered search tools can understand and cite it more easily.
- Our private internal operations AI helps Website Genii manage client work, daily tasks, and company processes. It is a practical example of software that does not need to become a public product to return value every day.
- The GSCSW member experience includes automatic renewals, personalized member portals, exclusive content access, event reservations, a directory, and member-only events.
- The Southern Land Exchange project includes a custom back-office workflow for property listings and agent profiles, plus public property search and dedicated agent pages.
This experience is why we believe AI should handle as much repeatable implementation work as it safely can. It is also why every AI-produced change needs human review and testing. A comprehensive custom CRM, a live portal that combines many reporting sources, and a TV sales leaderboard are all realistic project types, but their permissions, data, and failure risks still have to be designed by people.
Should You Buy, Configure, Connect, or Build?
Use the least complex option that solves the real problem:
- Buy an existing product when the workflow is common, such as accounting, scheduling, file storage, or basic contact management.
- Configure it when the product is close and only needs fields, permissions, templates, or reports.
- Connect good existing tools when the pain comes from copying data between them. Workflow automation and system integration often solve this without another full application.
- Build when the workflow is distinctive, repeated, valuable, and poorly served by existing products.
Custom does not mean creating everything from scratch. A sound tool may rent proven services for sign-in, payments, databases, email, and hosting while custom code handles the part that makes the business different.
What Does AI-Assisted Custom Software Cost in 2026?
There is no honest single price for custom software. Clutch’s current software development pricing guide says projects in its verified reviews commonly fall between $10,000 and $49,999. It also reports an overall average of about $132,480 and an average timeline near 13 months. Upwork’s current developer marketplace guide places custom web applications in a broad $3,000 to $15,000 range.
Those online figures combine many scopes, team sizes, and development methods. They are useful market context, but they are not Website Genii’s pricing model. Our recent AI-assisted work can be much leaner when the business problem is clear, the data is accessible, and an experienced person directs and reviews the build.
Website Genii’s Current Planning Math
Website Genii’s standard rate for this work is $200 per active work hour. Active time is the human work that makes the result useful and dependable: client meetings, workflow planning, directing the coding agent, reviewing its output, testing real cases, checking security, fixing failures, launching, and documenting the tool. Passive time while an AI run finishes is different from active human work.
These examples show the arithmetic, not fixed packages or promises:
- 10 active hours: about $2,000. This can fit a feasibility test, sample-data prototype, calculator, research tool, one-view dashboard, or another small internal helper.
- 20 active hours: about $4,000. This can fit a focused production tool or one comprehensive, read-only reporting view that joins several products when they provide clean, well-documented connections. Most active time on a project like this may go into client conversations and decisions rather than typing code.
- 40 active hours: about $8,000. This can fit a more complex first release with sign-in, several data sources, simple user roles, automated updates, testing, and basic administration.
- 80 active hours: about $16,000. This can fit a larger first release with several integrations, permissions, reporting, data setup, and an audit trail, which is a record of who changed what and when.
A useful 10-hour tool is not unheard of. Neither is a fairly complex 40-hour build when its workflow is clear and the outside systems cooperate. An 80-hour budget creates more room for data cleanup, unusual cases, administration, testing, and a safer launch. A regulated platform, difficult migration, unreliable integration, or complete CRM may still need more.
Hosting, paid APIs, email, storage, model usage inside the finished tool, and ongoing maintenance are separate operating costs unless a proposal includes them.
Plan the Software Before Asking AI to Build It
The strongest first step is a written plan, not a giant coding prompt. Define the problem, users, screens, data, integrations, permissions, failure cases, security needs, acceptance tests, and what the first version will deliberately leave out. Then challenge and revise that plan several times before implementation begins.
That preparation gives Codex, Claude Code, or another coding agent a much better target. With Codex using GPT-5.6 Sol’s Ultra mode, a well-scoped run can complete a large amount of routine implementation while the human team focuses on choices and review. In our experience, the longest part of a well-planned build can be waiting for those runs to finish, not writing every line by hand.
The plan does not need to predict everything. It should make unknowns visible early. If one data source or integration is uncertain, test that piece first before estimating the full production tool.
What Does a Full AI-Assisted CRM Example Show?
Phillips Data Solutions says its founder used Claude Code to replace HubSpot Starter with a custom CRM in roughly one week. The founder’s account lists deals, campaigns, lead intake, integrations, alerts, sign-in, and multi-factor authentication. He reports infrastructure under $25 monthly, but that excludes his skilled labor, testing, data work, and maintenance. It shows what clear workflows and experience can compress, not a beginner’s one-week promise.
Can a $20,000 to $200,000 Project Now Cost $3,000 to $10,000?
Sometimes. A clearly planned workflow that once required a larger team and much more hand-written code can cost far less with today’s tools. That does not prove that every old $200,000 system can be recreated for $10,000, or that a smaller first version is identical to the old full scope.
Research also shows why experience and planning matter. Google’s 2025 DORA research describes AI as an amplifier: it strengthens good teams and processes, but it can also magnify weak ones. A narrow early-2025 METR study found that experienced developers working in familiar, mature projects took 19 percent longer with the tools tested then. That study came before today’s models and does not set a limit on current Codex performance. It still shows that adding AI without changing the workflow does not guarantee a faster result.
The better claim is simple: AI can make a well-defined first version far cheaper to test. It cannot make unclear requirements, dirty data, fragile integrations, security work, and long-term ownership disappear.
Are Tokens the Main Cost?
Usually not. A token is a small piece of text processed by an AI model. OpenAI’s current model catalog lists GPT-5.6 prices by the million tokens. As an uncached example, 250,000 input tokens plus 25,000 output tokens costs about $0.08 on the lowest listed tier or $1.50 on the highest. Repeating an AI task at scale changes the math, but skilled planning, review, testing, and maintenance usually cost more than one build run.
Claude Code is included in Anthropic’s $20 monthly Pro plan, while heavier Max plans start at $100, according to Claude’s current pricing. A finished product may also pay for hosting, a database, email, storage, backups, and monitoring.
What Costs Continue After Launch?
Budget for routine operation even when launch is inexpensive. Outside services change, data grows, and employees need different access. A small tool may need only a few maintenance hours in a quiet month. A business-critical system needs monitoring, tested backups, updates, user support, and a named person who can respond to failures or security issues.
Where Do Agents and Loops Fit?
An AI agent is software that can use tools and take steps toward a goal, not just answer one question. A loop means it checks the result, chooses the next step, acts again, and stops when a clear condition is met.
A reporting agent might:
- collect approved data from sales, accounting, advertising, and support systems
- match names and dates into one standard format
- flag missing or unusual records
- refresh a dashboard and draft a summary
- ask a person to approve the summary before it is sent
Other useful loops can sort incoming documents, prepare quote details, route leads, watch inventory exceptions, or assemble a weekly operating report. Use ordinary rule-based automation when the steps are fixed. Use an agent when the work requires reading, classification, or judgment.
Keep a human approval step before an agent sends customer messages, changes financial records, deletes data, approves work, or publishes legal, medical, or financial conclusions.
Does Codex Hit a Size or Token Wall?
A context window is the amount of information a model can consider in one request. It is not the maximum size of an entire software project. OpenAI currently lists a 1.05-million-token context window for its GPT-5.6 models, and Anthropic lists 200,000 tokens for individual Claude plans. These are published model and product limits, not a promise that every Codex or Claude Code task can use the full amount. Neither number is a repository-size limit, and the limits can change.
Codex and Claude Code can work on repositories larger than one context window by searching for the relevant files and opening only what a task needs. There is no credible rule that Codex stops working at a certain number of files or lines of code.
The real bottleneck appears when one change requires scattered knowledge from many parts of the system. Token use and mistakes rise when names are inconsistent, modules are tightly linked, instructions conflict, or there are no reliable tests.
Keep larger projects workable by using clear modules, short architecture notes, one source of truth for business rules, focused tasks, and automated tests. Give the coding agent the goal, boundaries, affected workflow, and acceptance tests. Treat build-time token use separately from the tokens a finished AI feature may consume every day.
Security Is the Main Boundary
Anyone can prompt a screen into existence. The dangerous gap is assuming that a working screen is a secure system.
Use a simple risk ladder:
- Lower risk: public information, sample data, local research tools, calculators, and prototypes that cannot change important records.
- Medium risk: internal documents, employee-only tools, email access, business reporting, and integrations that can create or edit records.
- Higher risk: customer logins, financial or health data, payments, legal files, broad connector access, automated messages, and systems that run the business.
For anything above the lower-risk level, require these basics:
- give every person and connection only the access it needs
- use established sign-in services and test every role
- keep passwords and API keys in a protected secret store, not in the code
- validate all incoming data and handle errors safely
- record important actions in an audit log
- test backups by restoring them, not just by creating them
- scan dependencies for known flaws and keep them updated
- test normal, failure, permission, and abuse cases
- review AI-generated changes before they reach production
- arrange independent security review for sensitive or business-critical systems
The NIST Secure Software Development Framework says security practices must be built into the full development process. OWASP’s Secure Coding with AI guide adds risks specific to coding agents, including untrusted packages, leaked secrets, unsafe instructions, and agents operating with broad permissions.
Codex also has security tools. Official Codex Security documentation describes repository scans, deeper scans, code-change reviews, finding validation, and help fixing confirmed problems. Those tools can make review faster and more thorough. They do not prove that an application is safe, replace a qualified reviewer, or cover every mistake in permissions, cloud settings, business rules, and operations.
If a prototype touches sensitive information, move it to sample data until the security plan, access design, testing, monitoring, and response process are ready.
Should You Build It Yourself or Hire a Partner?
DIY is a reasonable place to start when the tool is local or public, uses non-sensitive data, is easy to reverse, and has a small group of informed users. A good hybrid is to build a sample-data prototype with Codex or Claude Code, prove the workflow, and then have an experienced developer prepare the production version.
Bring in professional help when the tool has customer accounts, financial or regulated data, several integrations, complex permissions, high uptime needs, or power to take important actions. Ask who owns the code, hosting, database, accounts, backups, and documentation. Require acceptance tests and a maintenance plan before launch.
Start With One Valuable Workflow
AI can lower the cash cost of a narrow first version. The best first project solves one repeated problem, uses the least sensitive data possible, and has a clear owner. Prove that version before adding more systems and authority.
If you want help deciding whether to buy, connect, or build, talk with Website Genii about the workflow. Our custom software and technical consulting can help define the smallest dependable version and the security it needs.