Comparing AI tools for working on your business

These tools are sold to developers, but most of the businesses paying for them are not software companies. They are a ten-person business with a spreadsheet somebody fixes by hand every month, a website form whose submissions reach nobody, a bank export that has to be reconciled against invoices, and two systems that do not talk to each other. All four tools here solve exactly that, they just speak developer.

This guide compares them from the business side: what actually gets closed, who has to sit in front of the tool, what it costs at the end of the month, and what happens to the customer data that goes in. Prices and policies were checked in September 2026.

Four tools, two product shapes

Tool	What it actually is	Default model	What it takes to start
Claude Code	An agent in the command line, with VS Code and JetBrains extensions and a separate app	Claude Opus 5	A Pro subscription and 20 minutes of setup
Codex	An agent in the command line, browser, IDE extension and iOS	GPT-5.6 family	An existing ChatGPT subscription, the browser version installs nothing
Cursor	A full editor, a VS Code fork, with the agent inside it	You pick the provider, Claude and Grok included	Downloading software and working inside it
Grok Build	An agent in the command line, open source under Apache 2.0	Grok 4.6	A SuperGrok or X Premium+ subscription

The difference that matters to a business is between software that replaces a working environment and a tool that runs alongside what you already have. Cursor is something you sit inside, so it makes sense when an employee or contractor writes code most of the time. The other three open in a command window, take an instruction in plain language, and work on whatever folder is in front of them. For a business with no permanent developer, that second kind is far easier to trial: you can drop it at the end of the month without anyone having learned a new application.

Which jobs it actually closes

The job in the business	What the tool does	How long it takes
A bank export that has to be reconciled against invoices	A script that reads both files and lists the gaps	Half an hour the first time, one click after that
A monthly report someone assembles by hand from four sources	A script that builds the report from the raw files	One to two hours
A website form emailing an address nobody reads any more	A small fix in the site's code	Minutes, given access to the code
A messy customer list, duplicates and phone numbers in five formats	Cleaning and normalizing the file, with a report of what changed	Half an hour
Connecting two systems that both have an API	A script that moves data between them overnight	A working day, testing included
A small internal tool, say a screen showing stock levels	A simple app running on one server	One to three days

What this does not close: anything requiring a signature, regulatory approval, or a connection to the tax authority's systems. An invoice sent to a customer has to meet structural requirements and carry an allocation number, and that is a job for an invoicing system, not for a script the agent wrote yesterday. The practical boundary is simple: these tools are excellent at the internal work nobody sees, and a bad idea for anything that goes out in the business's name.

Who in the business sits in front of it

The honest answer is that somebody has to be able to read an error. The agent writes the code, but it also occasionally proposes something destructive, like deleting files or changing something in production, and whoever approves needs to understand what they are approving. Three setups work in practice: a technical owner willing to learn, an employee who already administers the systems, or a contractor who comes in one day a month and uses the tool to do three days of work in it.

What does not work is handing the tool to an employee with no technical context and hoping. Not because they will not understand the tool, but because they will not know when the answer is wrong, and a reconciliation script returning a wrong number looks exactly like one that works.

Where the customer data ends up

This is the question worth asking before anything else, and the answer depends less on the tool than on the plan tier.

Vendor	Personal plans	Business plans
Anthropic	Training on conversations is on by default since August 2025 unless you turn it off in privacy settings. Five year retention with training on, thirty days with it off	Team, Enterprise and API: no training on what you send
OpenAI	A regular ChatGPT account, Plus and Pro included, follows that account's training settings	Business and Enterprise: no training on business data by default
Cursor	Privacy mode blocks code from being used for training, by Cursor and by the model providers it routes to	Same privacy mode, enforceable across the whole organization from the admin panel
xAI	A /privacy command inside the tool turns off retention at session level	Zero Data Retention for eligible enterprise customers, with SOC 2 Type 2

Two things are easy to miss. First, if an employee pays out of pocket for a personal plan and runs it on the company's files, the company is on that personal plan and its training terms. That is the most common setup of all, and the only one nobody documents. Second, Cursor is an intermediary: data passes through its servers on the way to the model provider, which is an extra link in the chain that belongs on your vendor list.

Israel's Amendment 13 and the third-party processor

Amendment 13 to the Privacy Protection Law has been in force since 14 August 2025, with expanded enforcement powers for the Privacy Protection Authority and a data protection officer requirement for certain organizations. For a business using one of these tools the implication is direct: a customer file with names, phone numbers and emails that someone pastes in to clean up duplicates is personal data that just left for a third party, whether or not anyone filed it that way.

What works in practice is separating the kinds of jobs. A script running on invented data, or on the file's structure alone, is one thing. Work on the real file, the live database, or an export of customer records is something else, and it goes through the same process every other cloud vendor does: a business plan, a data processing agreement, and a check on where data is stored and for how long. In many cases the cheapest fix is deleting the identifying columns before the file goes in at all.

What it costs a business

Plan	Monthly price	What you get
Claude Pro	$20	Claude Code for one user, moderate use
Claude Max	From $100	5x or 20x Pro, for working with the agent all day
Claude Team	From $20 a seat, premium seat at $100	Central billing, user management, limits that differ by seat type
ChatGPT Go	$8	Codex in small doses
ChatGPT Plus	$20	Codex in the command line, browser, IDE and on iOS
ChatGPT Business	$25 per user monthly, $20 annual, two user minimum	SSO, MFA, no training on business data
Cursor Pro	$20	The agent in the editor, cloud agents, frontier model access
Cursor Teams	From $40 per user	Organization-wide privacy mode, SSO, analytics, automated code review
SuperGrok	$30	Grok Build, which X Premium+ at $40 also unlocks

The number in the table is not the cost, because these plans measure usage in a rolling five-hour window. One intense two-hour session burns a whole window, and the ceiling arrives exactly mid-job. On OpenAI's side, usage has been billed in credits against token consumption rather than message counts since April 2026.

For a business the relevant arithmetic is different. A $20 to $30 subscription is about fifteen minutes of contractor time. If the tool closes one job a month that would otherwise have gone out to someone, it paid for itself twenty times over. Anyone using it a few hours a day lands around $100 to $200 a month, still less than a single day of development. The economic question is not what the subscription costs but how many hours a month someone in the business spends on repetitive work that a script could do.

Control, permissions and what the agent may touch

Claude Code and Codex both run a permission model where the agent asks before changing a file or running a command, with rules you can set in advance for what needs no approval. Claude Code adds hooks, meaning your own code runs before and after every action, which is where a block on sensitive paths belongs. Grok Build ships sandbox profiles, permission rules and OIDC auth, and it is the only one of the four whose source is open, so you can check for yourself what it sends out. Cursor, as an enterprise product, brings SAML and OIDC SSO, SCIM seat management, audit logs, and SOC 2, ISO 27001 and ISO 42001 certifications.

The practical rule for a small business is short: the agent gets a copy, never the original. A separate folder, a duplicated file, a test database. If the agent makes a mistake and deletes something, it deleted a copy. That is worth more than any elaborate permission configuration, and it takes two minutes.

Which tool suits which business

Situation in the business	What makes sense
An owner who wants to try it, without deep technical background	Codex in the browser, since there is nothing to install, or Claude Pro for a tool that works on a local folder
An employee or contractor who writes code most of the time	Cursor, since they are sitting in an editor anyway
A small team wanting one subscription with administration	Claude Team or ChatGPT Business, whichever the business already uses
A business touching real customer files	Business plans only, with a processing agreement and a check on storage location
A requirement that nothing is retained by the vendor	Enterprise ZDR at xAI, or an Enterprise plan with custom retention at Anthropic and OpenAI
Wanting to verify for yourself what the tool sends	Grok Build, the only open source one in the group
A job that has to run itself every night, unattended	A command line tool, since they run headless

What to check before deciding

Take one real job you have already done by hand and let the tool redo it. That tells you what the right answer looks like, and how much time is actually saved. Run it on a copy. Open the account's privacy settings and screenshot them, because they change without notice. Then ask the simple question: if it turned out tomorrow that everything the tool saw was retained by the vendor for a year, what exactly is sitting there. If the answer includes your customer list, the problem is the file you uploaded, not the tool you picked.

Share