Not every automation has to be sophisticated to matter.

This one is a form, an AI analysis step, and a spreadsheet output. It runs on about ~1,500 products a month. Each run saves about 10 of manual work. That adds up to ~250 hours per month and an estimated ~$500,000 per year in labor cost that no longer exists.

This is a post about how it works, what made it possible, and what the experience taught me about where AI automation creates real business value.

The problem

At a large retail company, physical product data has to live somewhere structured before it can flow into retail systems. That somewhere is Salsify, a product experience management platform that retail brands use to manage product content across channels.

For every product in the catalog, that means: a product title, a production description, SEO metadata, and extracted text from the product itself. Accurate, consistent, channel-ready content for thousands of SKUs.

Before this workflow existed, that was a manual job. Someone would open a spreadsheet containing product images, look at each image, write the title, write the description, fill in the SEO fields, transcribe the text on the product, and move on to the next row. Then the completed data would get loaded into Salsify.

It was repetitive, time-consuming, and exactly the kind of work that erodes focus. It was also a bottleneck. The speed at which product data could be enriched was limited by the number of people doing the enrichment.

The solution

The workflow is built in Power Automate with a “Run a Prompt” AI step powered by Power Apps/AI Builder. The architecture is deliberately simple:

1. Input: a Power Automate form accepts a spreadsheet or email attachment containing product images.

2. AI analysis: for each image, a “Run a Prompt” step sends the image to an AI model with instructions to generate a product title, production description, SEO metadata, and extract any text visible on the product.

3. Output: the results are written back to an .xlsx file and returned to the user when processing is complete.

That’s the entire flow. Low-code. No custom model, no complex orchestration, no infrastructure to maintain. The AI does what it’s genuinely good at, which is analyzing visual content and generating structured descriptive text, and Power Automate handles the surrounding process automation.

The completed output goes into Salsify, where it replaces what used to be entered manually.

The numbers

The workflow runs approximately 1,500+ times per month.

Each run automates roughly 10 minutes of manual work (the time it previously took a person to analyze an image, write the required fields) and enter the data.

That’s 250+ hours per month of manual work eliminated.

The estimated annual business value is $500,000, a figure that accounts for fully loaded labor costs, downstream process efficiency, and the compounding effect of faster product data availability across the catalog.

It’s worth being precise about what that number represents. It isn’t $500K in headcount reduction; the people who were doing this work do other things now. It’s $500K in capacity: work that previously consumed 250 hours per month of human attention now consumes none. That capacity redirects to higher-judgment work that automation can’t replace.

What made this possible

A few things aligned to make this the right problem for AI automation:

The task is well-defined. Writing a product title, description, SEO fields, and extracting product text are bounded tasks with clear inputs and outputs. There’s no ambiguity about what “done” looks like. Tasks like this are where AI performs most reliably. The model isn’t being asked to exercise judgment, it’s being asked to produce structured output from visual input.

The volume justifies automation. ~1,500 runs per month is enough that even a modest time saving per run accumulates into something significant. Many potential automations don’t get built because the volume doesn’t justify the effort. This one did.

The tolerance for imperfection is real. AI-generated product content gets reviewed before it goes live in Salsify. The workflow doesn’t need to be perfect, it needs to produce a strong first draft that a human can quickly verify and adjust. That’s a much lower bar than full automation, and it’s the bar most enterprise AI workflows should be designed for.

The tooling was already in place. Power Automate and Power Apps (AI Builder) are part of the Microsoft 365 stack the organization already uses. There was no new vendor procurement, no security review for a net-new platform, no integration work. The workflow could be built and deployed within the existing approved toolchain. In a large organization, this matters more than most technical decisions.

It also takes advantage of low-code tools, which some organizations prioritize when building AI/automation solutions.

The organizational challenge was harder than the technical one

Building the workflow took less time than getting it used.

The resistance was the normal friction of changing how people work. Questions about accuracy (“what if the AI gets it wrong?”), questions about process (“where does this fit in the existing workflow?”), and the general inertia of teams that have been doing something a particular way for a long time.

A few things helped:

Framing it as a first draft tool, not a replacement. The workflow generates content that a human reviews. It doesn’t publish anything automatically. That framing reduced anxiety significantly. The AI isn’t making decisions, it’s doing the tedious part so the human can focus on judgment and quality control.

Starting with a team that had the most to gain. The people running ~1,500 of these per month felt the pain most acutely. Getting buy-in from them first created internal advocates who could speak to the value from their own experience, which mattered more than anything I could say.

Measuring and sharing the impact. Putting a number on it (250+ hours, ~$500K) made the value concrete in a way that general claims about “AI productivity” don’t. People respond to specific numbers.

What I’d build next

The current workflow handles the enrichment step. There’s adjacent work worth automating:

Validation: a step that checks the AI output against Salsify’s field requirements before the file is returned, flagging anything that doesn’t meet format or length constraints. Catches issues before they become manual corrections downstream.

Direct Salsify integration: rather than outputting an .xlsx for manual upload, push the enriched data directly to Salsify via API. Removes the last human step in the process entirely for cases where confidence is high.

Confidence scoring: surface cases where the AI’s output is likely to need more review, so human attention concentrates where it’s actually needed rather than being distributed evenly across all output.

The broader lesson

The $500K number gets attention, but the more important takeaway is simpler: the highest-value AI automations are often the most straightforward ones.

A form, an AI step, and a spreadsheet. That’s it. The value comes from identifying a high-volume, well-defined, repetitive task and removing the human from the loop for the parts that don’t require human judgment.

Most organizations have dozens of processes that fit that description. The bottleneck isn’t the technology. It’s knowing where to look and being willing to build something simple.