The Common E-Commerce Mistake: Treating Product Setup as a Manual Ops Task

A lot of teams still think e-commerce catalog setup is just tedious back office work. That assumption is the problem.

When you are uploading thousands of SKUs, fixing item names, mapping attributes, updating images, and checking pricing by hand, you are not just wasting time. You are also introducing errors that scale with the catalog. In B2B commerce, that pain can stretch into months. The more products, the more brittle the process becomes.

Balaji Ingole’s work at Amla Commerce is a useful example of how to attack this class of problem with AI, but without pretending AI is magic. The goal is not to replace every step with automation. The goal is to redesign the workflow so people spend less time on repetitive setup and more time on exceptions, validation, and decisions.

Why the manual approach breaks down

For manufacturers and B2B sellers, product setup is often the slowest part of launching or maintaining an e-commerce catalog. The work is repetitive, but it is not simple:

• product data arrives in inconsistent formats
• naming conventions vary across sources
• image and metadata updates need coordination
• errors in one field can affect search, merchandising, and downstream operations

Ingole has said that this process can take customers two to three months to complete. That is not just an inconvenience. It delays go-live dates, slows revenue, and ties up staff in work that does not require deep expertise.

This is where many teams make a bad bet: they add more people to the process instead of changing the process itself.

A better model: AI as a guided operator

At Amla, Ingole is building AI-enabled chatbots that help manufacturers set up and manage large product catalogs in e-commerce platforms. The important part is that the AI is positioned as a guide, not an autonomous black box.

That distinction matters.

A practical AI agent for catalog setup should do things like:

• walk users step by step through product creation
• help organize large batches of product data
• reduce repeated manual entry
• surface missing information before it becomes a production issue

According to Ingole, the aim is to reduce setup time from two to three months to about two weeks. That is the kind of target that forces a team to rethink workflow design, not just add a chatbot on top of a legacy process.

If you are building something similar, the main question is not “Can an LLM do this?” It is “Which parts of the process are deterministic, which parts need human review, and where can an AI agent safely reduce friction?”

What to automate first

A useful implementation strategy is to start with the most repetitive, low-risk work. In product onboarding, that usually includes tasks like extracting highlights from source material, organizing incoming product data, and prompting for missing fields.

Ingole’s own project-management workflow shows the same pattern. He built an AI agent that runs every Monday morning, reads emails, extracts highlights, risks, timelines, and upcoming releases, then produces a written status report. He noted that the workflow involves more than a dozen steps, including defining parameters, managing temporary files, and integrating with existing tools.

That detail is important because it pushes back against another common misconception: automation is not a single prompt.

Real workflow automation usually has a lot of plumbing:

• input collection
• field extraction
• temporary state handling
• integration with existing tools
• final review before output

For builders, that means the win is not “the model wrote something.” The win is “the system reliably performed a chain of useful operations without forcing a human to do the boring parts.”

Guardrails matter more than hype

Ingole’s background in health care IT is relevant here too. He managed core systems that state health departments used for Medicaid benefits, provider enrollment, and eligibility verification. He has pointed out that health care data is unlike other data because the compliance burden is high and the tolerance for error is extremely low.

That lesson transfers directly to enterprise e-commerce.

Catalog data may not carry the same life-or-death consequences as health care, but bad data still causes real damage:

• incorrect pricing
• broken customer experiences
• failed integrations
• lost trust with sellers or distributors

So the AI layer needs constraints. The right pattern is usually:

• let the AI propose or organize
• let humans approve exceptions
• keep auditability for changes
• fail safely when inputs are incomplete

This is especially important for B2B catalogs, where a single error can cascade across large inventories and multiple buyers.

Why Ingole’s research habit matters

Ingole also spends time on independent research in data analytics and AI-enabled health care applications. He has published more than 40 peer-reviewed papers and holds six patents in the U.K. and India. That background helps explain his working style: build, test, review, refine.

He has said he believes in “learn by doing,” and that he likes to prototype ideas to see whether they actually work. For developers, that mindset is a reminder that AI products mature fastest when they are treated as systems under test, not as finished answers.

Publishing research plays a similar role. He has said that journal and conference review forces him to defend methodology and tighten the connection between use case and technical architecture. The same idea applies to product engineering. If you cannot explain how an AI workflow handles edge cases, it probably is not ready for production.

A practical takeaway for builders

If your team is dealing with slow product onboarding, do not start by asking how to make AI “smarter.” Start by mapping the workflow.

Identify:

• which steps are repetitive
• which inputs are structured versus messy
• where human judgment is required
• what kind of review or rollback you need

Then use AI to remove the most expensive manual work, not to replace operational control.

That is the real lesson in Ingole’s work: useful AI is usually specific, bounded, and embedded in a workflow that people can trust. In e-commerce, that means fewer months spent cleaning up catalogs and more time spent shipping products, improving customer experience, and handling the cases that actually need human attention.