Do You Still Need a DXP If AI Can Write the Code for Your Customer Portal?

AI can build a customer portal prototype in a weekend. Learn what a production portal still needs, what you own if you build from scratch, and where a DXP fits.

Get the Guide

Table of Contents

    Key Points

    • Most evaluations aren’t asking which AI coding tool to use, but whether an enterprise platform is even necessary with the advances of AI code generation.
    • AI code generators are good at producing working software quickly, but speed doesn’t mean a finished, enterprise-ready solution.
    • When you build from scratch, you own everything the generator didn’t hand you, including authentication, permissions, integrations, governance, accessibility, security patching, and years of upgrades.
    • Independent research suggests that the code arrives faster but costs more to maintain. Duplication is rising, refactoring is falling, and review capacity is a major constraint.
    • Vibe coding and a platform are complementary rather than competing. AI shortens development time, but a platform is what keeps the result secure, integrated, and maintainable.

    The New Question Customers Are Asking

    Not so long ago, a project for a customer portal started with a platform evaluation. Today, things look different. Someone on the team has built a working prototype over a weekend using a tool like Cursor, Lovable, v0, Bolt, or Claude Code, and the prototype looks credible. It has a login screen, a dashboard, and a page that pulls real data from a test endpoint.

    The other team members see the prototype, then glance at each other with raised eyebrows, everybody thinking the same question: If AI can generate all of this, do we still need to buy a platform?

    This question needs a more careful answer than yes or no. After all, the prototype - and the speed - are real. But what isn’t factored in is all the work and potential cost that comes next.

    This guide sets out what a live customer portal requires beyond the prototype, what your team takes on if you build those parts yourself, and what an enterprise platform provides instead.

    What “Build It with AI” Actually Means

    “AI-built” can mean two things:

    1) Using AI to assist development with an existing platform. Your team still builds on a governed foundation. Nothing about your security model, integration layer, or upgrade path changes.

    2) Building an application from scratch with AI. Here, an AI generator produces the solution’s source code itself, and you become responsible for hosting, securing, integrating, and maintaining it.

    For this guide, we’ll focus on the differences between the two approaches and why building everything with AI isn’t actually the most cost-effective decision for the long haul.

    Speed of Production Doesn’t Mean “Finished”

    AI code generators have become very good, especially at first glance. Recent research suggests their growing aptitude in creating code that runs. This doesn’t mean, however, that that code is “safe” to use as a base for your website or to power a solution for customers.

    Lines of code on a computer screen, the kind of output an AI code generator produces in seconds

    Veracode has been running the same controlled test since 2023, giving models a coding task that can be solved securely or insecurely and measuring which option they choose. In the Spring 2026 update, across more than 150 models evaluated to date, the models produced secure code in only 55% of tasks on average. In other words, they introduced a known vulnerability class in 45% of them. In Java, the security pass rate was even lower, 29%. Syntax correctness over the same period climbed past 95%. So, yes, the code compiles more reliably - but not safely.

    Another interesting finding? The security pass rates have not improved over the testing period.

    Google’s 2025 DORA report on AI-assisted software development, based on roughly 5,000 respondents, found that AI adoption correlates with higher throughput and, at the same time, with greater instability, more change failures, and more rework.

    Developers report the same pattern from the inside. In the Stack Overflow Developer Survey 2025, the most common frustration with AI tooling, cited by 66% of respondents, was “AI solutions that are almost right, but not quite.” Just over 45% said debugging AI-generated code takes more time than debugging code written by a person. More respondents actively distrusted AI accuracy (46%) than trusted it (33%), and scepticism was highest among the most experienced developers.

    Does this mean we don’t let AI write code at all? No - but we shouldn’t assume that AI code is safe. In fact, AI-generated code still needs the most scrutiny.

    Risks When Building From Scratch

    A customer portal is not just a self-service interface. The table below lists what a production customer portal generally needs, what an AI code generator can produce for each item, and what is still your team’s responsibility if there’s no platform underneath.

    A combination padlock resting on a computer keyboard, representing the security a customer portal must provide
    What the Portal NeedsWhat AI Generation Gives YouWhat Your Team Still Owns
    Authentication and single sign-onLogin forms and a plausible integration attempt against a protocol you nameCertificate handling, session management, token refresh, identity federation, and keeping it correct as your identity provider changes
    PermissionsRole checks written where you asked for themA consistent model applied across every page, record, document, and API endpoint, plus proof that permissions work as they should
    Integrations with systems of recordAPI client code drafted from documentation you supplyError handling, retries, rate limiting, credential rotation, and updating every client when an upstream API changes
    Content management for non-developersContent hardcoded, or a basic admin screen if you ask for oneAn editing experience your marketing and support teams can use without filing a ticket
    Process, approvals, and SLA trackingConditional logic for the paths you describedMulti-step routing, parallel approvals, escalation, reassignment, and reporting on whether steps completed on time
    SearchA database query, or a search box over one tableRelevance, indexing, and results filtered by what each user is permitted to see
    AccessibilityMarkup that may or may not meet the standardTesting against WCAG, remediation, and conformance documentation
    Audit and compliance evidenceApplication logging, if requestedA durable record of who accessed and changed what, in a form an auditor accepts
    Multi-language and multi-region supportText extracted into files, at bestTranslation workflow, regional content variation, and locale-aware URLs
    Security patchingNothing. The generator does not monitor your dependencies.Tracking every library you inherited, assessing each advisory, and patching on your own schedule
    UpgradesNothingEvery framework and dependency upgrade, indefinitely, with no vendor doing it for you
    Hosting, scale, backup, and recoveryNothing, unless you use a managed prompt-to-app platformInfrastructure, load testing, failover, backups, and tested recovery

    Take a look at the right-hand column. This is the actual scope of a from-scratch build, and very little of it appears in the weekend prototype.

    From-scratch isn’t impossible, and many organizations have done it. But you need to ask yourself: do I want the burden of building and maintaining everything AI can’t with my limited engineering resources?

    And What About Maintenance?

    The cost to build is fairly visible and can be estimated. But maintenance cost, a much larger and more variable number, isn’t often factored into a from-scratch customer portal project.

    A technician inspecting server racks with a handheld diagnostic tool, the ongoing upkeep a from-scratch build requires

    The long-standing finding in software engineering is that maintenance consumes between 40% and 80% of total software cost over an application’s life, averaging around 60%. That figure comes from Robert Glass’s Facts and Fallacies of Software Engineering (Addison-Wesley, 2002), Fact 41. This data came way before AI code generation, and that’s important: maintenance has always been expensive. And now code generation is happening at exponential rates.

    Early evidence indicates that AI-generated code is harder to maintain than the code it replaces. GitClear’s analysis of 623 million code changes between 2023 and 2026, published as The Maintainability Gap, found that duplicated blocks of five or more lines rose 81% against the 2023 baseline, the highest level in their records. Over the same period, moved code, which is what refactoring looks like in a diff, fell from 21% of changed lines in 2022 to 3.8% in 2026. Developers are now considerably more likely to duplicate a block than to refactor it. GitClear also measured a 35% fall in how often new code calls existing code, which suggests generated code is being added alongside the codebase rather than woven into it.

    This is the behavior product teams describe when asked what goes wrong in practice. Ask a generator for the same banner across ten pages, and you’ll receive ten separate copies. Each one works. But a change to that banner later means finding and fixing all ten, though there’s nothing in the codebase to tell you ten different banners exist.

    Review capacity is the other constraint. Faros AI, analyzing engineering telemetry from around 22,000 developers across 4,000 teams in early 2026, reported that median code review duration rose 5X. 31.3% more pull requests merged with no review at all, which they attribute to reviewers running out of capacity rather than to any policy change. Their report, The Acceleration Whiplash, is vendor-published but based on observed activity rather than self-reported survey answers.

    Code arriving faster than it can be reviewed, duplication rising, and refactoring falling, all mean a rapid accumulation of technical debt—to say nothing of the security risks we mentioned earlier.

    For just a prototype, this isn’t significant. But for a customer portal that you expect to run for a decade, this is a critical cost problem.

    What a Platform and Enterprise Subscription Offers That You Don’t Build

    The strongest argument for using a platform with the potential for AI-assisted development (the first approach from above) is that a platform comes ready for the enterprise world—built, tested, and maintained.

    Here’s exactly what that means on Liferay DXP.

    Identity and authentication. Liferay DXP supports SAML as either identity provider or service provider, OAuth 2.0 as authorization server and resource server, OpenID Connect as a relying party, LDAP directory integration with user and group import and export, and multi-factor authentication including time-based one-time passwords and FIDO2 devices. On your side is just the configuration, not development.

    Permissions down to the individual record. Roles can be scoped to the whole instance, an organization, a site, or an account, and site administrators can create teams to manage permissions locally within one site. For custom data, documentation states that “with custom objects, you can manage permissions for individual database entities to control access to object data.” Documents and folders carry their own permissions, and search results are filtered by what the user is allowed to see.

    Data models that come with APIs. Liferay Objects let teams define business entities visually, with fields, relationships, and validation rules. When an object definition is published, Liferay generates default REST endpoints for it automatically, and the object is exposed in Liferay’s GraphQL schema. Every object is integrated with workflow.

    Process, approvals, and service levels. The workflow engine supports sequential and parallel approval paths using fork, join, and join-XOR nodes, with tasks assigned to users or roles and notifications built in. Workflow Metrics lets you define an SLA as a start event, a stop event, and a duration, then reports items as overdue, on time, or untracked, with breakdowns by step and by assignee.

    Accessibility you can show. Liferay publishes an Accessibility Conformance Report based on VPAT 2.5Rev, assessed per success criterion against WCAG 2.0, 2.1, and 2.2, Revised Section 508, and EN 301 549, for a named quarterly release. Conformance is reported criterion by criterion rather than as a blanket claim, and custom fragments and client extensions your team builds are not automatically covered. Generated code comes with no conformance report at all, and producing one is extra work on your team.

    Security maintained as a vendor responsibility. Liferay has undergone independent verification of its security, privacy, and compliance controls since 2019 and holds ISO/IEC 27001, ISO/IEC 27017, ISO/IEC 27018, SOC 2 Type 2, CSA STAR, and Spain’s ENS among others. Liferay is a CVE vendor, runs a public bug bounty programme, publishes a known vulnerabilities index per release, and conducts annual penetration testing alongside static analysis, dynamic analysis, and dependency scanning. On a from-scratch build, all of that is your job to run.

    A supported upgrade path. Under the Liferay Service Life Policy, the first quarterly release of each year is designated a long term support version and receives three years of premium support followed by two years of limited support. Client extensions interface only with Liferay’s APIs rather than modifying the platform, and Liferay’s documentation notes that “this loose coupling makes client extensions more flexible and easier to apply to environments after an upgrade, since they depend only on APIs that rarely change between versions.”

    A deployment model that matches your constraints. Liferay offers SaaS, PaaS and self-hosted options. On Liferay SaaS, customization happens through client extensions and Liferay applies platform upgrades and patches. That’s a clear difference between the platform and from-scratch approaches: with a vendor’s SaaS offering, keeping the foundation current is their responsibility. With a from-scratch build, it’s always yours.

    Vibe Coding and a Platform Are Complementary, Not Competing

    Remember: vibe coding is a development approach, not a platform. Vibe coding can have an impact on how quickly software gets built, but doesn’t make that software secure, integrated, scalable, or sustainable.

    When used together, the two form a powerful team. AI shortens build time when properly regulated. The platform serves as the foundation. You get both speed and maintainability. That prototype you built gets used as a useful exploration tool before committing engineering to a full project.

    Liferay’s approach to AI-assisted building follows the same logic. Liferay’s AI generates standard low-code building blocks inside the platform rather than freeform code bolted onto it, so the output lives where teams already work. Generated code respects the roles and permissions you’ve already defined.

    Whatever the AI produces stays transparent and editable through the same visual tools, so there is no separate black box to maintain. AI generates reusable elements such as fragments, templates, and content, and can be directed to reuse what already exists before creating anything new, so one element used in many places is updated once. Your existing review and approval workflows apply to AI-generated changes, which keeps people in control of what goes live.

    Read more about Liferay’s low-code capabilities or AI Hub.

    Questions to Ask Before You Decide

    If your team is weighing a from-scratch build against a platform, here are questions that will uncover your potential costs.

    A team with laptops listening to a colleague at a whiteboard while weighing a build-or-buy decision

    1. How long do we expect this portal to run?

    2. Who will maintain this in three years?

    3. What happens the first time a dependency has a critical advisory?

    4. How will we prove our permissions model works?

    5. What do we owe on accessibility?

    6. Where does our regulated data sit, and who can reach it?

    7. What is the upgrade story?

    8. Would building this be the best use of our engineering capacity?

    What It Looks Like When a Platform-Built Customer Portal Works

    Value from a platform shows in time to launch and measurable outcomes. Take these three Liferay customers:

    A woman using a laptop at home, as a customer would when logging in to a self-service portal

    Broadcom went from vendor and product selection to launch in ten months, implementing 845 features, and reported a 66% decrease in clicks for users now that resources sit in one place.

    IndiaFirst Life Insurance reported a 20% reduction in customer service calls and tickets, a 40% increase in customers paying premiums online, and more than 60,000 registered users.

    HDI Assicurazioni implemented their portal within 18 weeks and registered 50,000 new users in the first six months.

    Making the Decision

    If you’re building a prototype just to test an idea, use the tool that gets you there fastest, which is probably just AI. As long as you keep prototypes away from production data until they pass a security and architecture review, this makes sense.

    But if you’re building the system your customers will use to log in and share sensitive data, i.e., a system that counts as a crucial extension of their trust in you, then you need more than a code generator.

    AI reduces the time it takes to build software, but a platform like Liferay DXP along with its built-in AI cuts time to business value and your lower total cost of ownership.

    Book a demo to talk through your customer portal project requirements or keep exploring how Liferay DXP supports solutions for customers.

    Frequently Asked Questions

    Can AI-generated code be used for a production customer portal?

    In theory, yes, if the organization takes on everything that makes an application ready for production, including security review, dependency management, integration maintenance, accessibility testing, audit evidence, hosting, and ongoing upgrades. The generated code is the starting point rather than the finished system. Read more about the security risks of vibe coding and how thorough review can make it safer.

    Is Liferay against vibe coding?

    No. Vibe coding is one development approach and works well for prototyping, proofs of concept, and low-risk internal tools. Vibe coding becomes a problem only when it is treated as a substitute for an enterprise platform.

    Does a platform slow teams down compared with generating code?

    With a platform, there is more design work upfront, because data models, permissions, and workflows are configured deliberately. That investment is what makes later changes predictable.

    What if we have already built something with AI and want to move it onto a platform?

    That’s not an uncommon path, and it’s the four-stage pattern described in our vibe coding and low-code guide: explore with AI, evaluate the prototype, rebuild the validated concept on a governed platform, then measure and scale. The prototype isn’t wasted but starts out as the specification.

    How do we compare the cost of building versus using a platform?

    Compare the costs over the full expected life of the application rather than at launch. Include maintenance, security patching, upgrades, hosting, and the engineering time spent on parts a platform would have provided. Maintenance has historically accounted for the majority of total software cost, so a comparison based on build effort alone will understate the from-scratch option significantly.