- Anubhav Sarker
- Published: 02/09/2026
- Last Updated: 10/13/2026
Veracode, which sells code scanning tools, has tested more than 100 AI models on 80 coding tasks over four years, and in its 2026 report, the average security pass rate was 56%, a figure Veracode says has not moved. The test used short pieces of code in a lab, not working business apps. So the 56% doesn’t tell you how likely your own CRM is to be hacked. It does tell you that AI often writes code with security gaps, which is why you should test anything you build before customer data goes into it.
Vibe coding is a way to build, and SaaS is a way to buy
The AI researcher Andrej Karpathy named vibe coding in February 2025. He described accepting an AI's changes without reading them, calling it "not too bad for throwaway weekend projects." SaaS, short for software as a service, describes delivery: a vendor hosts the software, updates it, and charges a subscription.
The terms overlap. A CRM built in Lovable or Replit runs on managed cloud services, and a SaaS CRM can contain code its engineers wrote with AI help. The practical choice is between building and operating your own CRM with AI assistance and configuring one that someone else maintains.
Review matters as much as authorship. The developer Simon Willison argues that AI-written code is ordinary software development once a person has reviewed it, tested it, and can explain it. Replit's shared responsibility model says Replit secures its platform, encrypts data, and backs up published code and Replit-managed databases, while the builder handles how users log in, what each user may do, the business logic, and regulatory compliance. Lovable's security documentation describes a scan that runs every time you publish, and it warns that its tools "cannot guarantee complete security."
The crucial question
After the first version of your vibe-coded CRM works, who owns the changes, failures, and customer commitments that follow? With a SaaS CRM, the vendor owns the platform, and you own your configuration and your data. With a CRM you built, you own nearly everything above the hosting, which works when someone will carry it.
Six changes that test a CRM after launch
Each of these changes is ordinary, and each one creates work that someone has to own.
A second person needs less access
A login screen is easy to add to a first version, and it can hide the problem. Signing in proves who a person is. Permissions decide which records that person can open, change, delete, or export. An app can check identity and stop there, which leaves a part-time assistant able to read every customer and every price as soon as they sign in.
In March 2025, the researcher Matt Palmer scanned 1,645 projects from Lovable's public gallery and found 170, about 10.3%, with database access rules too weak to protect data such as names, emails, and payment details. Lovable disputes the resulting vulnerability report and says each customer must protect its own app's data. In December 2025, the security firm Tenzai had five coding agents build the same three apps and found 69 vulnerabilities across the 15 results. One app let a seller account view any order. Tenzai sells security testing. Its findings come from sample apps.
A SaaS CRM ships with a permission model that someone still has to configure. In Bigin, roles decide which records a user sees, and profiles decide whether that user can edit, delete, or export them.
Two people edit the same record
In a simple app, the last person to save overwrites the whole record. A phone number that one person corrected can disappear when someone else saves notes from a page they opened earlier. Developers call this "last write wins." The fix is to save only changed fields, warn when a record has changed, or keep a change history, and each fix waits until someone notices the problem.
The same customer arrives twice
A customer who fills in a web form and then phones the office can end up as two records with two follow-up reminders. Someone has to decide whether an email address, a phone number, or a street address identifies a customer, and what happens to the notes when two records merge.
A connection stops without telling anyone
A Gmail connection that shows emails on customer records depends on a sign-in token. Google's OAuth documentation lists several ways that token can quietly fail. It stops working after six months without use or when the user changes a password, and it lasts only seven days while the app's Google project is marked as testing. Emails then stop appearing on records, and a basic app warns nobody unless someone built an alert.
Every integration needs upkeep. Someone renews tokens, maps "Phone" to "Mobile," stops a retried message from creating a duplicate, and updates the code when the other company changes its API. Managed connectors help. Zapier can retry failed steps automatically on its Professional plan and above, though someone still reads the alerts.
A bad change meets the backup
An AI edit can remove data as quickly as it adds a feature. A backup is a copy. A restore is a return to usable customer history, with deals, notes, and attachments still linked to the right people.
During a code freeze in July 2025, Replit's AI agent deleted the production database behind an app that SaaStr founder Jason Lemkin was building, then claimed a rollback was impossible. Lemkin later wrote that the rollback did work. Replit has since separated development and production databases, and its documentation says the agent cannot modify production data.
Supabase is a database service that Lovable apps can connect to, and its Pro plan keeps daily backups for seven days. The project goes offline during a restore, files in Supabase Storage are excluded, and restoring last night's copy loses anything entered since. Recovery to an exact moment costs from $100 a month extra.
Subscriptions have limits here too. Bigin keeps deleted records in a recycle bin for 60 days. Its paid editions include two backups a month, and its backup FAQ describes downloading records, related records, and attachments without describing a restore into the account.
The person who built it leaves
If the AI builder workspace, the database project, the Google project, and the domain sit in one person's name, nobody else can safely change the app after that person leaves. Each account has to move to a business email address, and someone has to write down where the secrets live, how changes get published, and how a restore works.

Figure 1. Each change after launch creates work, and the diagram shows who handles it in each model.
What a year of software costs on each side
Building in Lovable with your own Supabase database costs $25 a month for Lovable Pro, which includes 100 credits, and $15 for each extra 50 credits. Supabase Pro adds $25 a month for the database, logins, and daily backups. That is $600 a year in platform fees before extra AI credits, point-in-time recovery, a security review, or any developer you hire.
Bigin Premier, the edition with duplicate cleanup, costs $15 per user per month, or $12 billed yearly. Express costs $9, or $7 billed yearly, if you can work without duplicate cleanup and stage transition rules. Subscription fees grow with every user, so 10 Premier users cost $1,800 a year on monthly billing. The build's platform fees barely move with headcount, because Supabase Pro includes 100,000 monthly active users.
I think time is the higher cost on both sides. Building means hours on the first version, every later change, permission tests, connection checks, and restore practice. Subscribing means hours on setup and routine administration. Estimate both with your own numbers before you compare fees. I have priced no outages or incidents, because nothing supports a prediction.
What you own when you own the code
Owning the code lets any developer you hire read and change it. An exported copy leaves behind the database rows, the user accounts, the stored API keys, the scheduled jobs, and the hosting and domain setup. Lovable's Cloud documentation says no one-click migration exists from its built-in backend to your own Supabase project.
A SaaS export has gaps too. A CSV file per module can lose the links between a contact, its deals, and its notes. Bigin's backup includes related records and attachments, but its documentation does not mention automation rules or permissions, so I would document those separately.
Customer stories show the work a CRM carries
These Bigin stories describe requirements, and none involves a vibe-coded CRM or compares development methods.
Goverdhan Retail India, a garment wholesaler with about 100 employees, once ran an in-house, on-premise CRM built by hired developers. Its story says director Vinay Gupta's biggest challenge was upgrading that system whenever the business process changed, on top of an annual maintenance fee. It was conventional custom software. The same pressure falls on any CRM a business builds for itself.
The story of V4 Creative, a nine-person design agency in South Africa, shows how much "build a CRM" can hide. An approved quote moves from a Quotation pipeline to Billing, where estimates and invoices go out through Zoho Books, a separate Zoho product with its own plans. Customers pay through a Zoho Books link, staff manually mark each invoice as paid, and connected pipelines then move the deal into Projects.
WorkCon runs labor hire and traffic control from Brisbane, Australia. Its team splits the territory into sub-regions, keeps client email history on deals, and plans sales work in a calendar view. John Anstee, Brisbane Client Engagement Representative, describes a recurring commitment: "When I send off some rates, I've got it set to give me a reminder to ring them in two days." His advice suits any CRM: "Make sure you've got the time to put the data in and set it up right the first time."
CysterCare, a health and wellness company in Chennai, India, turned down two other CRMs because of inadequate support. While evaluating Bigin, the team hit a problem, and support replied with a fix in the middle of the night. That is one company's experience. Bigin's pricing page says chat and phone support depend on region and language.
Where building makes sense, and where a subscription fits
Building makes sense for an unusual workflow, a narrow tool for one or two people, or a business with someone who can maintain code. If a day of downtime would cost you a customer, the app needs monitoring and someone on call. If the data includes payment or health details, it needs a professional review. Each added user brings more permission rules to write and test.
A maintained CRM fits an established sales process, shared customer history, weekly follow-ups, and a team that wants vendor support. Per-user fees grow with headcount, and editions gate features. In Bigin, duplicate cleanup requires the Premier plan or Bigin 360, and Express allows three team pipelines. Custom modules and approval processes are Zoho CRM features that Bigin's editions do not list, and you live with the vendor's decisions.
A hybrid keeps one record of each customer
In a hybrid setup, the CRM holds contacts, deals, and history, while a small custom app does one job, such as a quote calculator for an unusual pricing model. The app reads customer details and writes back the quote amount. Decide which system owns each value, keep both from editing the same field, and name the person who watches the connection. Bigin supports this through Zapier, toppings, and developer APIs that allow paid editions 50,000 credits a day plus 250 per user, up to 100,000.
What you compare | Building with AI | Subscribing to a maintained CRM |
|---|---|---|
Customization | You shape every screen and rule. | You configure what your edition allows. |
Implementation effort | A first version takes days, and review takes longer. | You set up pipelines, fields, users, and imports. |
Ongoing ownership | You own every change and fix. | The vendor maintains the platform, and you maintain your setup. |
Integrations | You or a connector service keeps each one running. | The vendor maintains built-in integrations. |
Access controls | You design and test every permission rule. | You assign roles and profiles and test them. |
Recovery | You choose backups and practice restores. | You learn the vendor's recovery terms and test them. |
Costs | Fees are low, and staff hours carry the cost. | Fees grow per user, and setup takes fewer hours. |
Portability | The code moves, and data and accounts move separately. | Records export, and settings move by hand. |

Figure 2. The CRM owns the customer record, the custom app owns one calculation, and a named person owns the connection.
Run five tests before you commit
Load 20 representative records, with personal details changed, into each option.
- The daily user takes one inquiry from first contact to a finished job.
- The most restricted user tries to open and export a record outside their access, including by editing the record ID in the address bar.
- Someone revokes the email connection, and you time how long it takes anyone to notice.
- You delete five records and a note, restore them, and confirm that links and attachments returned.
- Someone other than the builder adds a pipeline stage and a user from the written notes alone.
Treat the results as evidence for a buying decision. They do not certify security.
My recommendation
Build with AI when the workflow is unusual, the tool is narrow, the data is low risk, and a named person will maintain it. Buy a maintained CRM when customer history is shared, and follow-ups are promises to customers. Consider a hybrid when a maintained CRM covers most of the work and one step needs custom logic.
If Bigin is on your shortlist, start the 15-day trial, which needs no credit card and moves to the Free plan afterward, and run the five tests there. Check feature availability for the edition your tests rely on.
-
Edited by Anubhav Sarker | Images on this article are AI generated. Please verify thoroughly before using