Bubble's AI app builder can generate a working first version of a customer portal: pages, a database, workflows, sample data and login flows. It is not ready for customers. By hand you must set privacy rules so each customer sees only their own records, replace the sample data, test roles and files, and move to a paid plan to go live.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowOne honesty note first. This tutorial is a method, not a report of a finished build: nobody has generated this portal in a live Bubble account for it, so it does not claim what the generator "usually" gets wrong. The hand-fix list below comes from what Bubble's own documentation says you must configure yourself, and from the checks Bubble's security tools run. Treat everything the AI produces as a draft to inspect, then work through the list. The running example is a kitchen fitter whose customers log in to see project progress, documents and messages.
What does Bubble's AI app builder produce for a customer portal?
Bubble's AI Agent produces a working first version of an app from your description, with pages, elements, a database, workflows and sample data for your development environment. Bubble's Manual says it can create a fully functional app from simple prompts: you describe the idea, it proposes a structure, you confirm features and add any it missed, then you generate. Beyond the features you ask for, the Agent also builds typical core features such as sign-up, log-in and reset-password flows.
Three details matter for a portal. First, the sample data is placeholder information that mimics real data, there so you can see the design working; you replace it before launch. Second, the development environment is where you build and test, and changes there do not affect users until you deploy to the live environment. Third, Bubble generates a visual app, not source code, so you can open any element, workflow or data type and change it directly, or keep chatting with the Agent. Bubble labels the Agent chat feature as beta on its AI app builder page.
Write the prompt: a filled-in example
Bubble's prompt guide gives a structure: "A [app type] for [target users] to [main purpose]. Users can [key features]. Style should be [visual preferences]." It advises being specific about purpose and users, naming core features first, and giving visual direction in plain words. It tells you to avoid being vague, to avoid dictating exact layout, and to avoid technical specifications such as database or login-provider choices. Prompts that break Bubble's acceptable use policy, or contain injections, are rejected.
For the kitchen fitter, a prompt in that shape could be:
A customer portal for homeowners who have booked a kitchen fitting,
to follow their project and share documents with the fitting company.
Customers can log in, see their project status (survey booked, design
agreed, order placed, fitting, snagging, complete), view and download
their quote and design drawings, upload photos of their kitchen, see
upcoming appointments, and send messages to the fitting team.
Staff can see all projects, update each project's status, upload
documents for a customer and reply to messages.
Style should be clean and calm, navy blue and warm white, with a
friendly tone.
Notice what it leaves out: no mention of databases, button positions or payment providers. It names two kinds of user, which sets up the permissions work later. Do not paste real customer names, addresses or quotes into the prompt. Use made-up placeholders while building.
Review the plan, features and scenarios before generating
Bubble's docs describe the build process: you may be asked a few multiple-choice questions, the AI builds a plan, and you review it. The Features view lists what the app will do as cards, with an Add feature button. The Scenarios view restates the plan as what each type of user can do, with an Add scenario button. For this portal, check for these and add any that are missing:
- "Staff can see every project; each customer can see only their own project."
- "Customers can upload photos, and only that customer and staff can view them."
- "Staff can change a project's status, and the customer sees the new status the next time they log in."
- "A customer who is not logged in sees only a login page."
Then select Generate. You do not need the plan to be perfect, since you can keep adjusting afterwards. Bubble says app generation currently uses no AI credits, so refining the plan is free.
What must you fix by hand in a Bubble AI customer portal?
Bubble's generator gives a structure, not a security review, so you must check privacy rules, sample data, roles, file access, searches, notifications and deployment yourself. This table lists what to confirm. The right-hand column says what a pass looks like.
| Area | Where in Bubble | A pass looks like |
|---|---|---|
| Privacy rules on every data type | Data tab, Privacy | No customer data type shows "Publicly visible"; each shows "Privacy rules applied" |
| One customer cannot see another's records | Test with two accounts | Customer A gets nothing when trying to reach Customer B's project, files or messages |
| Staff versus customer roles | Data types and privacy conditions | Staff rules key off a role field; customer rules key off "this project's customer is the current user" |
| Uploaded files are private | Privacy rules and the uploader element | A customer's photo URL does not open for another customer or a logged-out visitor |
| Sample data removed | Development data, then live | No placeholder customers or messages remain once you go live |
| Password strength and test environment | Security dashboard | No "weak password policy" or "unprotected test environment" issues listed |
| Searches are constrained | Repeating groups and "Do a search for" | A customer's page searches only their own records, so the server returns only what they need |
| Notifications and emails | Workflows | A real test message reaches a real mailbox from your sender, not a placeholder |
| Live deployment | Deploy, paid plan | The live version loads on your domain, and you repeat the two-account test there |
Everything in the first row is where a portal succeeds or fails, so it gets the next section.
Why do privacy rules matter in a Bubble customer portal?
Bubble's own documentation puts it bluntly: privacy rules are an essential part of an app's security, and any private or sensitive database data needs them to be considered secure. A privacy rule is a condition on a data type that tells Bubble's server to send data to the browser, or write to the database, only if the condition holds. Bubble gives the plain-language form of a rule, for example "Only return the data if the thing's Creator is the current user".
To see where you stand, open the Data tab, then Privacy. The left panel lists every data type with its status, either "Privacy rules applied" or "Publicly visible". For each rule you set a condition in the When field, whether people matching it can find the thing in searches and view attached files, and, field by field, whether they can view it, use it as a search constraint, and save to it with auto-bind.
Filled-in rules for the kitchen fitter's portal
These are illustrative; adapt the field names to what your app calls them.
| Data type | Rule name and condition | What it allows |
|---|---|---|
| Project | "Customer's own project": When Current User is This Project's Customer | Find in searches, view all fields, view attached files |
| Project | "Staff": When Current User's Role is Staff | Find, view and edit all fields |
| Project | Default (everyone else) | Nothing: no find, no view, no files |
| Document | "Own documents": When This Document's Customer is Current User | View and download; the customer is stored directly on the document, so the rule needs no chained reference |
| Message | "Participants": When Current User is This Message's Customer, or Current User's Role is Staff | View and create, but not edit someone else's message |
| User (the app's customer accounts) | "Self": When This User is Current User | View own email and name only; not other users' email addresses |
Bubble warns of several traps that apply here:
- Client-side data is not secure. Once data reaches a user's device, even if no page displays it, that person can see it in the traffic. The only protection is not sending it, which is what privacy rules do on the server. Filtering a list on the page is not a substitute.
- "View all fields" means view. It does not stop someone changing a field through a workflow. The rule settings are meant to be taken literally.
- Rules do not refresh a page already open. If a change in the user's situation alters what they can see, the page shows the old result until it reloads.
- Chained references have limits. Privacy rules can evaluate fields directly on a data type, but a condition that reaches through several levels of related things cannot grant search access. If a rule that should work does not, store the customer on the document itself.
- Slugs are public. Slug fields are designed to be public and searchable, so never put anything sensitive in one.
The longer treatment of this topic, with more rule patterns, is in stopping clients seeing each other's data in Bubble.
Check it with two customers, then with the security dashboard
- Create two test customers in the development version, each with a project, a document and a message.
- Log in as the first. Open the browser's page and try to reach the second customer's project by changing the address or any search on the page. You should see nothing.
- Log out and visit the same addresses. A logged-out visitor should see only the login page.
- Log in as a staff user and confirm every project appears.
- Open the security dashboard. Bubble says the Starter plan includes basic checks such as missing privacy rules, search privacy, constrainable fields, sensitive data exposed in workflows, weak password policies and unprotected test environments, while Growth and above add advanced checks. Run the privacy rules checker, which reviews all data types for publicly accessible fields; Bubble notes it needs sample data in the data types to give meaningful results, and can take a few minutes on a large database.
- Fix every item it lists, then repeat steps 2 to 4.
Repeat the two-customer test on the live version after you deploy.
Can Bubble's AI Agent see your customers' data?
Bubble's AI Agent works with the same permissions as the person who starts the chat, and reading the data stored in your database is limited to owners and admins. Someone with editor rights can have it edit the app; someone with view-only rights gets guidance only. Data is readable only for the data types enabled in Settings, AI.
For a portal, that setting deserves a decision. Leave customer data types off unless you need the Agent to inspect real records, and build with invented test customers. When a question can be answered from the app's structure ("why isn't this workflow running?"), the Agent does not need your customers' messages.
How do you edit with Bubble's AI Agent without wasting credits?
Bubble's visual editor uses no credits, while its AI Agent does, so make changes by hand where you can. After the first version you have those two ways to change the portal. Bubble's AI credits page says there is no fixed credit cost per message; the Agent uses more for requests that take more work, and credits are calculated after the request runs. Bubble's tips map well to a portal build:
- Change colours, wording and layout by hand in the editor.
- Ask the Agent for one change at a time, naming the page, element or workflow, for example "On the Project page, show the status as a progress bar".
- Choose the approval mode that suits you. Ask confirms every change before it is applied, Auto lets the Agent decide when to ask, and Approve applies changes without waiting. For a portal that holds customer data, Ask is the safer setting.
- Undo in the editor instead of asking the Agent to reverse a change.
- If a fix does not work after a try or two, describe the symptom in more detail instead of repeating the request.
If you run out of credits mid-request, the Agent stops, keeps the work it finished and tells you what it completed. You can still edit visually. Included credits refresh on the first of each month and do not roll over; credits you buy expire after six months.
How much does a Bubble customer portal cost to run?
A Bubble customer portal needs at least the Starter plan to go live, because the Free plan runs in a development version only. Bubble's pricing page lists these web plans: Free at $0 with 500 AI credits and 50K workload units a month; Starter at $32 a month billed annually with 1,000 credits, 175K workload units, a live website and custom domain, 2 days of server logs and basic version control and security checks; Growth at $119 a month billed annually with 2,000 credits, 250K workload units, 14 days of logs, two app editors, two-factor authentication and advanced security checks; and Team at $349 a month billed annually with 4,000 credits and 500K workload units. Bubble's page also offers monthly billing at different prices, so check the pricing page for the figure that applies to you.
A portal for customers needs at least Starter to go live. Growth becomes worth considering when you want advanced security checks, longer logs to investigate a complaint, or a second person editing.
Estimating workload for a small portal
Workload units measure server work, not AI. Bubble's Manual lists the cost of activities: loading a page's HTML is 0.15, a database search 0.3 plus 0.015 per thing returned, writing or changing a thing in the database 0.5, a server-side workflow action 0.6, and a server-side plugin call 0.2 plus 0.0005 for each millisecond the plugin runs. Work done on the customer's device does not count.
Take a customer's dashboard: one page load (0.15), four searches (1.2), about 29 things returned across them (0.435) and a little for the text itself (0.06). That is roughly 1.85 workload units a visit. Say the fitter has 80 customers who each visit ten times a month: 800 visits, about 1,480 units. Add 300 messages a month, each written to the database (0.5) and triggering a server-side notification action (0.6) and a plugin email call (0.2): about 390 more, before the small per-millisecond plugin charge. The total is under 2,000 units a month, around 1% of Starter's 175K. These are my estimates from Bubble's published figures, and real apps cost more when searches return long lists. Check the app's metrics after launch.
Bubble says you can disable overages so the app never uses more than its plan allows, and emails you at 75% and 100% of your units. How workload bills spike and how to cap them explains what drives costs up, and Bubble's plans and workload units compares tiers.
If you plan the customer list and workload sums in a spreadsheet first, The Complete Excel Course with AI Integration on Coding Liquids covers that spreadsheet work.
A hypothetical: the search that showed everyone's quotes
This is an imagined case, not a real one, built to show the failure the table is designed to catch. Suppose a fitter generates the portal, sees that each test login shows the right project, and launches on Starter. A customer then notices that a "My documents" list on one page shows quotes with other people's names. The page searched all documents, and the document data type was set to Publicly visible. A screen filter hid the others from view, but the data had already reached the browser.
Bubble's docs explain why: client-side filtering works on data already returned, and does not trigger privacy checks. The fix would be a privacy rule on Document that returns only the customer's own, a search constrained to the current user's project, and a repeat of the two-customer test. The Privacy tab check that would have caught it takes two minutes before launch.
How long does a Bubble AI portal take to build?
A simple Bubble customer portal takes about two and a half working days by my estimate, and extra integrations, payments and unusual rules add to it. A realistic plan for a small portal: half a day to write the prompt, review features and generate; a day to refine pages and data with the Agent and visual editor; a day for privacy rules, the two-account test, the security dashboard and live deployment.
Bring in a Bubble developer when you need payments, links to accounting or booking systems, complex staff permissions, or an audit of the privacy rules before launch. The tutorial on hiring a Bubble developer covers cost and a brief. If you would like to talk through whether a portal is the right project at all, that is what my AI implementation consultation covers; see the contact page.
Questions before building a portal with Bubble AI
Does generating an app with Bubble AI use AI credits?
Not for the first generation: Bubble's pricing page says app generation currently does not require credits. Credits are used when you build and iterate with the AI Agent, generate images or generate test data. Each plan includes a monthly allowance that does not roll over: 500 on Free, 1,000 on Starter, 2,000 on Growth and 4,000 on Team. Bubble says amounts can change.
Can customers use the portal while I build it on the free plan?
No. Bubble's free plan lets you build and test a fully functional app with no time limit, but going live needs a paid plan. Until then the app runs in the development version only. Starter is the first plan that includes a live website and custom domain.
Can the Bubble AI Agent read my customers' data?
Only within limits. Bubble's docs say the Agent works with the permissions of the person chatting, and that reading app data is limited to owners and admins and only for the data types enabled in Settings, AI. For a customer portal, enable only the data types the Agent genuinely needs, and use test records while building.
How do I check that customers cannot see each other's records?
Test with two accounts. Create two test customers with different projects, log in as the first, then try to reach the second's project through searches and links. Also open the Data tab's Privacy section and confirm no customer data type shows Publicly visible. Bubble's security dashboard also has a privacy rules checker, which needs sample data to test against.
Is a portal generated by Bubble's AI ready for customers?
No, treat it as a draft. Bubble's AI can generate a working first version with pages, a database, workflows, sample data and sign-up, login and password-reset flows. Before customers use it you must set privacy rules so each customer sees only their own records, replace the sample data, test roles and files, and move to a paid plan to go live.
Further reads
- Bubble vs Lovable for a Business App: Control, Cost and Upkeep — Control, cost and upkeep against an AI code builder
Browse all AI-for-business tutorials
Sources: Bubble AI app builder page, Bubble pricing page, Bubble Manual pages on generating apps with AI, the AI Agent, AI credits, privacy rules, workload activity types and the security dashboard (read 17 November 2026). I did not build this portal in a live Bubble account; the hand-fix list comes from Bubble's documentation.
