Business
11 min readLast week, in The DIY Delusion, I argued that building your own commerce engine in the AI era is a high-wire act without a net. Writing the first version is the easy part. Patching dependencies, carrying the legal exposure, and maintaining every line for the next decade is where the money goes. You become your own software maintenance vendor, and nobody pays you for it.
That post ruffled a few engineering feathers. The most common reply was some version of this:
"Okay, Brad. If we shouldn't use GenAI to rebuild commodity commerce plumbing, where should we point it?"
It's a fair question, and last time I answered it in a single line: use AI on the work that sets you apart, not on the components every retailer already has. This post is the longer answer.
The money at stake is real. Gartner now expects worldwide AI spending to reach $2.67 trillion in 2026. Yet in McKinsey's 2026 State of AI survey, 80% of respondents say AI has improved their own productivity, while only 37% can tie it to any EBIT impact, a share that hasn't moved in a year. The same survey found that 32% of organizations have passed on buying software because they figured they could build it with agentic coding tools. That's the DIY Delusion showing up in survey data.
So where does AI actually pay off? Here are five places we've seen it work, with our clients and on our own team.
Automated functional testing has been around since flip phones, and it still has a dirty secret: traditional test scripts break constantly. A frontend developer renames a CSS class or nudges the "Place Order" button, and half your regression suite turns red.
AI testing agents work differently. Instead of hunting for brittle technical dependencies, they read a page the way a person does and work out what it's for. When the layout changes, the agent adapts. That makes realistic exploratory testing practical across browse, cart, and checkout, without your QA team hand-scripting every edge case.
Adoption has moved fast. In Applause's 2026 State of Digital Quality in Functional Testing report, more than 92% of respondents use AI somewhere in testing, up from 60% a year earlier. Among those using it in testing, 65.1% use it to create test cases and 36.6% for autonomous test execution and adaptation. The same report found 29% saw defects rise in number or severity, a good reminder that AI-generated tests still need a human deciding what "correct" looks like.
The bigger opportunity in commerce is going past the UI. Give a test agent read-only access to your staging database and tell it to act like a difficult customer: apply a stacked B2B tier discount, split shipping across two addresses, place the order, then query the order tables to confirm every tax line, promotion adjustment, and inventory reservation landed where it should. Now you're testing the whole order, not just the button.
Anyone who has survived a four-hour backlog grooming session knows the pain of turning a three-sentence executive vision into developer tasks. Vague requirements become scope creep, misaligned sprints, and the classic "well, that wasn't in the ticket."
GenAI has become a real partner here for product owners and tech leads. Give it the business objectives and architectural constraints, and it produces a solid first pass at breaking an epic into structured tickets, with proper user stories ("As a merchandiser, I want...") and clear acceptance criteria. Applause found that 49.1% of respondents using AI in development already use it to develop or prioritize user requirements.
The connection between AI tools and project management has matured, too. Atlassian's Rovo MCP (Model Context Protocol) Server, generally available since February, lets AI assistants create and update Jira issues under your existing permissions, with no copy and paste. If your infosec team hasn't approved an MCP connection yet, have the AI produce a CSV and bulk-import it. Either way, your leads stop formatting tickets and start deciding which tickets should exist.
Last week's post said customers choose a commerce platform for experience, merchandising and pricing, not for a homemade cart. That's where your competitive edge lives: pricing logic, bundling rules, dealer workflows and checkout experience. On a closed SaaS platform, your developers spend half their week guessing how an opaque API behaves.
On a source-available framework like Broadleaf, built on standard Java and Spring Boot, AI tools can read the actual platform code. They can trace class hierarchies, find the right extension points to override, and compare two design approaches before anyone commits production code.
The productivity data backs this up. A field study of nearly 5,000 developers at Microsoft, Accenture, and a Fortune 100 company found that developers with GitHub Copilot completed 26% more pull requests per week. JetBrains' 2025 developer survey found that one in five developers using AI saves eight or more hours a week, a full workday. Instead of spending three days digging through framework libraries to learn how an extension point behaves, your engineers have the answer by lunch and spend the afternoon on features that drive revenue.
Enterprise commerce is rarely one storefront selling ten t-shirts. Once you're running dozens of brands, regional tax and currency rules, B2B distributor portals, inherited catalogs, contract pricing, and sandboxing across tenants, sites and profiles, your data model looks less like a highway and more like an interchange designed by M.C. Escher.
An AI copilot that can read your code and configuration makes it far faster to trace how a sandboxed promotion cascades across parent and child catalogs, or how data tracking resolves conflicts across channels. Just as useful, it surfaces downstream dependencies you might otherwise find in UAT, or worse, on Cyber Monday.
There's a caveat. In a January 2026 randomized trial, Anthropic found that developers who used AI while learning a new library scored 17% lower on a follow-up comprehension quiz than those who coded by hand. The ones who did well used AI differently: they asked conceptual questions and requested explanations instead of just accepting the generated code. That's the right model for architecture work. AI should make your senior architects sharper, not stand in for them.
Last week I called code a liability, not an asset. Here's how that plays out day to day. A developer deep in a checkout enhancement trips over a minor edge-case bug, sighs, and says, "I'll add it to the backlog and clean it up later."
"Later" rarely comes, and the small stuff compounds. Stripe's Developer Coefficient survey (2018) found developers spend about a third of their week on technical debt, and CISQ put accumulated U.S. technical debt at roughly $1.52 trillion as of 2022.
Asynchronous coding agents change that math. In McKinsey's 2026 survey, 31% of respondents at large enterprises say they're already scaling software coding agents. When an engineer spots a minor defect mid-sprint, they describe what they were doing, what they saw, and what they expected, then go back to their feature. The agent reproduces the bug, writes the fix, runs the tests, and opens a pull request for review. The debt gets paid the day it's found, before it accrues interest.
Notice what all five plays have in common: none of them asks AI to write a cart, an order state machine, or a payment flow from scratch. They point AI at the work you'd do anyway on top of a solid platform, and they make that work faster.
That's the other half of last week's argument. AI-written code still has to be secured and maintained. As I noted then, Black Duck's 2026 OSSRA report found the average audited commercial codebase now carries 581 known open-source vulnerabilities, more than double the year before. Writing more code faster means more code to own.
The advantage comes from pairing a proven, extensible commerce foundation with targeted AI workflows in testing, planning, custom development, architecture, and debt cleanup.
If you want to see how Broadleaf's composable, source-available architecture lets your team run these five plays into production without piling on maintenance, reach out. And tell me which of these is saving your engineers the most time right now.
Use AI everywhere it makes your team faster. Just don't use it to build every road by hand.
Doesn't this contradict last week's post? If AI makes coding faster, why not build?
No. AI lowers the cost of the first version, but most of a commerce engine's cost comes after launch, in patching, compliance, and upgrades. All five plays here apply AI to work you'd do on any platform, so the speed shows up in your differentiators instead of in plumbing you'd then have to maintain forever.
Do these plays work on any commerce platform?
Testing, ticket planning, and debt cleanup work on almost any stack. Custom development and architecture discovery depend on how much of the platform the AI can actually read. On a closed SaaS platform, it's guessing at an API; on a source-available platform, it's reading the code.
Is it safe to give an AI agent access to our order database?
Keep it to staging, read-only, with masked or synthetic customer data. Treat the agent like any new tester: least-privilege credentials, logged queries, and no path to production. The goal is verifying what landed in the tables, not touching live orders.
Won't AI coding agents create more technical debt than they remove?
They can, if nobody's reviewing. Applause found that 29% of respondents saw defects rise even as AI adoption climbed. Keep agent work small and scoped, run it through the same tests and code review as human work, and you get the speed without the mess.
Will AI replace our architects and senior developers?
No. The Anthropic study above suggests the opposite risk: people who hand everything to AI understand their systems less. Your senior people make the design calls, and AI helps them get to the answer faster.
Where should we start?
Pick one play with a clear before-and-after metric. Testing and ticket decomposition are usually the easiest, because the risk is low and the hours saved are easy to count. Prove it there, then expand.
Does Broadleaf support AI tooling directly?
Yes. Broadleaf offers an MCP server customers can use today, so AI assistants can work with your commerce platform through a governed connection. We also have our own Developer-assist AI tool, which we call Leafy, currently in beta and open to early-access requests from active clients.