Decision-Support RAG for an Online Retailer.

An established online retailer with a team of about 15 people had a data problem that did not look like one. Nothing was missing. Everything was somewhere.

Email marketing history sat in one platform. Order history sat in another. Customer service records sat in a third. Operations, forecasting and supplier decisions ran on spreadsheets that individual people maintained. Each system worked. None of them knew about the others.

The practical effect was that the most valuable question in the business, which is some version of "how did this go last time we tried it", had no cheap way to be answered. Somebody would remember roughly. Somebody else would remember differently. A decision worth tens of thousands of dollars would get made on the stronger memory rather than the better record.

McKinsey research puts knowledge workers at roughly 20 percent of the working week spent searching for internal information. For a fifteen-person team that is a real number, and it lands hardest on the people whose time is most expensive, because they are the ones asked to remember.

Delivered in collaboration with Visionary Automate, a systems-integration partner of Zealous Digital Solutions.

The first half of the work was unglamorous and was most of the effort. Every source got pulled into one place with a consistent shape: orders, customers, campaign history, support records and the operations spreadsheets that mattered.

On top of that sits a retrieval layer. Somebody asks a question in plain language, the system finds the relevant records across all the sources, and it answers from the company's own history rather than from general commercial knowledge. That distinction is the whole point. A generic assistant will tell you what retailers usually do. This one tells you what this retailer actually did, in which quarter, with what result.

The second capability runs without being asked. It surfaces patterns from the historical record that nobody queried for, such as a product line whose repeat-purchase rate has been sliding for three quarters or a campaign shape that has consistently underperformed its cost.

Answers cite the underlying records, so a claim can be checked rather than trusted.

The data unification and the retrieval layer above it were specified and built jointly. Delivered in collaboration with Visionary Automate, a systems-integration partner of Zealous Digital Solutions.

These figures are modeled from the retailer's stated team size, hourly rates and time use. They are not audited results. The assumptions:

• Leadership spending about 8 hours a week on decision analysis, valued at $60 per hour • 10 staff spending about 2 hours a week each hunting across systems, valued at $35 per hour • That cross-system hunting time replaced by self-serve queries, not eliminated entirely • One to two avoidable bad commercial bets per year, valued at $20,000 each • Institutional memory capture treated as strategic value and given no dollar figure at all • McKinsey's finding that knowledge workers spend roughly 20 percent of the week on internal information search, used as a sanity check on the hours above

That models out to roughly $25,000 from leadership time, $36,400 from staff time and $20,000 to $40,000 from decision quality, for a modeled annual benefit of $81,000 to $101,000. The model suggests payback inside the first year. Actual results depend on the client's baseline and adoption. These figures are modeled estimates, not measured client results.

The question changed shape. "Does anyone remember what we paid for that placement in 2023" used to start a thread. Now it is a query with a cited answer.

Decisions started carrying evidence. Not better decisions necessarily, and we are careful not to claim that, but decisions where the prior attempt is on the table rather than half-remembered.

The dependency on specific people softened. In a fifteen-person company that has been running for years, two or three people hold most of the operating history. That is a risk nobody writes down. Capturing it in a queryable form is the part of this build with the longest tail, and it is also the part we deliberately assigned no dollar value to, because we cannot measure it honestly.

Junior staff stopped queuing for answers. When the record is self-serve, a question that used to wait for a senior person's afternoon gets resolved in a minute.

The source inventory and data model were handed over as written joint deliverables rather than as tribal knowledge. Delivered in collaboration with Visionary Automate, a systems-integration partner of Zealous Digital Solutions.

This fits an established retailer or direct-to-consumer brand with a team of roughly 8 to 40 people and at least three or four years of trading history spread across separate platforms. Volume matters less than age here, because the value comes from having a past worth querying.

It is aimed at United States owners and operations leads, and it is region agnostic. The strongest signal that it applies is a team that regularly asks each other what happened last time and does not have a fast way to check.

It is a poor fit for a young business with 18 months of history, where there is not yet enough record to retrieve, or for a company whose data lives entirely inside one platform that already reports well.

Four weeks, and the first two are almost entirely data work. That is not a scheduling accident, it is where the value comes from.

• Week 1, discovery and data access. We inventory every source that holds operating history: the commerce platform, the email marketing platform, the support desk, and the operations spreadsheets that individual people maintain. Deliverable: a written source inventory recording, for every source, who owns it, how far back it goes, and whether it can be read programmatically. • Week 2, unification. Sources are pulled into one place with a consistent shape for orders, customers, campaigns, support records and the spreadsheets that matter. Deliverable: the unified data model plus a gap list naming the history that could not be recovered and why. • Week 3, retrieval layer and supervised pilot. Query answers are checked against people who already know the answer, which is the only honest test of a system built on institutional memory. Deliverable: a verification log of questions asked, answers given and corrections made. • Week 4, cutover. Plain-language querying opens to the whole team, pattern surfacing starts running unprompted, and citations appear on every answer.

The gap list from week 2 usually turns out to be as useful as the system. It is the first written record most companies have of what their own history does not contain.

Five prerequisites, and the first is a threshold rather than a checkbox.

• At least three years of trading history. The value here comes from having a past worth querying, so a business with 18 months of records does not yet have the asset this build reads. • Programmatic access to each source. A platform with an API or a full export is usable. A platform that only lets you look at dashboards is not, and it will have to be replaced or exported manually before it can be included. • The operations spreadsheets identified by name and owner. These are the sources most often forgotten in scoping and most often the ones holding the answer somebody actually wants. • A named person who owns the answer set. When the system returns something the team disputes, somebody has to adjudicate whether the data is wrong or the memory is. • A decided position on what the system may read. Customer records, support conversations and marketing history carry privacy obligations, and which fields are in scope is a business decision taken before the first source is connected.

What happens when the system cannot answer?

It says the record does not contain the answer, and it does not fill the gap from general commercial knowledge. That distinction is the entire point of the build. A generic assistant will tell you what retailers usually do, which is worse than silence when you are trying to remember what you actually did. Answers carry citations to the underlying records, so an unsupported claim is visible as one.

Who owns the data and the unified store?

You do. Every source system stays in your accounts, the unified store sits in infrastructure you control, and the data model is documented rather than proprietary. If the retrieval layer is removed, the unified store remains, which is the more valuable half of the asset.

What drives the ongoing running cost?

Three things: the volume of history kept queryable, the number of connected sources, and query traffic. History is the biggest lever and the one people trade off first, because keeping ten years hot costs more than keeping five. Source count is the second, since each connected platform has to be kept mapped as its vendor changes fields underneath you.

How is success measured in the first 90 days?

By behaviour rather than by dollars. How many questions get asked per week, how many are asked by people who are not the two or three who hold the history in their heads, and how many decisions in the period cite a prior attempt. The second measure is the one that matters, because the dependency on specific people is the risk this build exists to reduce.

Four cases where this build will not earn its keep.

• Young businesses with under about two years of trading history. There is not yet enough record to retrieve, and the honest advice is to start recording deliberately and revisit later. • Companies whose data already lives entirely inside one platform that reports well. Unifying one source is not unification. • Teams with no named owner for the answer set, where a disputed result has no route to resolution and the system quietly loses authority. • Businesses that have not decided what the system may read across customer and support records.

Naming these matters more here than in most builds, because a retrieval layer over thin history looks impressive in a demonstration and delivers nothing in month three.

Delivered in collaboration with Visionary Automate, a systems-integration partner of Zealous Digital Solutions.

The retailer is not named and no product line, supplier or customer is identified. The build described here is real and in production.

The value figures are not measured results. They are modeled from the retailer's stated team size, hourly rates and weekly time use, with McKinsey's published research on internal information search used only as a cross-check on the hours assumption. Actual results depend on the client's baseline and adoption. These figures are modeled estimates, not measured client results.

If your team regularly asks each other what happened last time and has no fast way to check, that question is the asset going unused. Start a conversation with your platform list, your trading history and your team size, and we will tell you whether there is enough past there to be worth querying.