Published on the 01/10/2026 | Written by Heather Wright

Retailer swaps report sprawl without more dashboards…

“3,064. That’s the number of Power BI reports we have in our environment Mitre 10.”

The comment from Nastassia Subritzky elicited plenty of laughs at Snowflake’s World Tour Auckland, but the Mitre 10 head of data and insights admits she was ‘pretty alarmed’ when she did the count during the hardware chain’s migration to Fabric.

“Our human users are also getting trained by our agents.”

“This reporting sprawl, plus the seemingly infinite number of spreadsheets makes it difficult because you’ve got all these different reports and spreadsheets and they all tend to have slightly different numbers, so you end up spending so much time trying to reconcile between all of them,” Subritzky says.

The problem arrived at the worst possible time, with the co-op already in the middle of a major ERP transformation, moving from individually customised store data warehouses to a centrally managed SAP platform and data warehouse.

“The move to a centrally managed data warehouse has naturally resulted in many cries for help with how to actually analyse data in this new SAP world,” she says.

Stores were struggling to work out where to go find the analysis they previously had available. The challenge was compounded by Mitre 10’s cooperative structure, where independently owned stores have their own approaches to performance and operations.

And then there was Excel.

“Turns out people get really attached to their Excel spreadsheets, especially during times of change,” she says.

Breaking free from report sprawl

Rather than create report number 3,065, the retailer decided to do something different.

The goal wasn’t to replace reporting, but to reduce dependence on it through AI agents built on curated datasets. Using Snowflake CoWork, the company created agents that draw from the same underlying data as Power BI, ensuring answers remain tied to a common source of truth.

The governance layer was critical.

Many organisations are discovering that generative AI amplifies existing data problems. Mitre 10 started with carefully controlled datasets, guardrails and verified queries to prevent users from generating yet another collection of disconnected answers. The retailer also maintained control over which datasets agents could access.

“With Snowflake Co-work, we have complete control over what datasets the agents have access to, and naturally to help with some reconciliation issues they use the same base tables as our Power BI workbooks.

“Everything naturally ties back together which has been a fantastic win for manual work reduction because I don’t have to send my engineers out to reconcile yet another Excel spreadsheet.”

Trent: Good tech skills, no business sense

The first AI agent, Trent, is the merchandising team’s ‘general assistant’. He was rolled out to the team as part of a two month deadline-driven migration from a legacy SQL reporting platform.

“This deadline doesn’t sound too bad until you consider how many workbooks someone can connect to a very old SQL reporting server over the course of a decade,” Subritzky quips. “We thought given they would have to get used to a new way of working anyway, why not just chuck them in the deep end and give them something really, really different to worry about?”

Trent helps with common tasks, including looking at supplier performance, undertaking basic rent reviews and looking at product performance by store.

But launching the agent was just the start.

“We got basic Trent up and running in absolutely no time at all, but he was very much like an intern. Really good technical skills, but absolutely no business knowledge whatsoever.”

What followed looks remarkably similar to onboarding a new analyst.

The team reviewed logs, refined prompts, built verified queries and taught the agent how the business actually worked. Instructions were added to standardise responses, eliminate exaggeration and ensure consistent analytical outputs.

Trent’s creator and Subritzky reviewed logs created from the merchandising team’s enthusiastic testing of Trent, turning successful requests into verified queries. There was some fun along the way too, with easter eggs added for testers.

The effort paid off.

“Trent had gone from being a really enthusiastic intern to a competent junior analyst who sometimes surprises us with some quite cool solutions.”

Today the agent is embedded in the merch teams daily workflow and has helped staff focus their analysis rather than build ever-larger spreadsheets.

Agents training humans

The store pilot provided more insights.

Initially, staff asked relatively simple questions. Over time, as confidence grew and the results were verified, the complexity of those questions increased. Store operators moved from asking basic sales questions to exploring inventory optimisation, promotional effectiveness and slow-moving stock analysis.

One example involved a store asking which products should be included in a winter clearance programme for slow and obsolete stock. Instead of simply producing raw figures, the agent asked clarifying questions, identified priority categories and recommended potential actions.

“What we were finding here was they could actually genuinely chat away to it because my team had done all the hard work up front to get all of those table joins going, all of the Mitre 10 context, all the odd little language that you end up having in business.”

Just as interesting was what happened to the users.

Agents were instructed to push back on vague or incomplete requests and ask for more information when necessary.

“In the interest of the cost, the compute time, and user experience, we added in explicit instructions about what sort of prompts the agent should go back to the user with if they didn’t get enough information.”

If users keep asking ambiguous questions, the agent keeps prompting them for more information.

Over time, users learned to write better questions.

“So our human users are also getting trained by our agents.”

That improvement in data literacy surfaced elsewhere as well. Because the platform could return the SQL underpinning its answers, analysts began developing a stronger understanding of how data tables connected and how metrics were calculated. The result was not simply more automation, but more capable users.

Meanwhile, executives discovered a different benefit.

With access to mobile capabilities, senior leaders began using the agents directly during meetings. Questions that previously required analyst support could be answered immediately. In stores, managers started interrogating data from the shop floor, investigating anomalies as they encountered them rather than waiting to return to a laptop.

Most importantly, the initiative appears to be delivering against its original objective. During the pilots, stores participating in the programme requested fewer new Power BI reports and modifications. Users were finding answers themselves rather than generating more reporting demand.

As Mitre 10 looks ahead, the challenge is ensuring history doesn’t repeat itself.

The retailer is already receiving requests for additional AI agents and is wary of replacing one form of sprawl with another. “We’re working out how we can make the existing analysts flexible, what sort of combinations we can use so that we can reuse a lot of what we’ve got and only do new agents when it really makes sense or is a really separate use case.”

After all, nobody wants to spend next year cleaning up 3,064 agents.