Technology
How Skyline Chili's purchasing co-op put 20 AI 'direct reports' to work
Sky Co-op founder Tom Hannon explains how AI that monitors the business, rather than just answering questions, is helping catch margin leaks in weeks instead of months.
October 2, 2026 by Cherryh Cansler — Publisher, FastCasual.com
ChatGPT
Grok
Perplexity
ClaudeMorgan has a big job at Sky Co-op, the purchasing cooperative behind Skyline Chiliand other restaurant concepts totaling about 325 locations. She monitors purchasing data, looks for pricing anomalies and flags potential supply chain problems before they quietly eat into restaurant margins.
Morgan may be busy 24/7, but that's ok since she isn't human. She's one of 20 AI-powered digital direct reports, or DDRs, that Sky Co-op Founder Tom Hannon has built using CollectivIQ, a platform created byBuyers Edge Platform. Her story offers a look at what happens when AI stops acting primarily as a chatbot and starts taking on ongoing job responsibilities.
"We're still early, so I'm continuing to develop that framework, but I generally think about it as putting a digital team member to work for items I want continuously monitored and handled," Hannon told FastCasual in an email interview. "An audit might help me periodically look at commodity pricing, food safety incidents, or distributor fill rates. A DDR becomes more useful when I want the system to continuously evaluate that information and tell me what's changed, what's unusual and what may require action."
From periodic audits to continuous monitoring
Sky Co-op's membership includes 136 Skyline Chili restaurants and a broader group of regional concepts, each with different menus, suppliers and cost structures.
Traditionally, the co-op audited contracted items every 30 to 60 days, meaning a pricing error could run for two or three months before anyone caught it.
Hannon said the problem became clear when one food item's contract price jumped by roughly $20 from one month to the next. An operator happened to notice about three weeks later, but the error could easily have continued until the next audit.
"A dashboard gives me information when I go looking for it," Hannon said. "A DDR has an ongoing responsibility to monitor an area of the business and surface what needs my attention."
That distinction is why he built Morgan as a continuously running monitor. She flags price changes above a threshold, currently 5%, so the team can investigate unusual movements as they happen. Because CollectivIQ, owend by Consolidated Concepts, integrates with ArrowStream and InsideTrack, the co-op's existing data platforms, Hannon can check her findings against the underlying records.
The timing matters. Hannon said traffic is down, costs are up and margins are being squeezed, so catching even relatively small cost increases quickly can make a meaningful difference when spread across hundreds of locations.
Flagging, not deciding
For now, DDRs identify, analyze and surface issues but do not make purchasing decisions on their own. A raw percentage change also isn't enough to determine whether something matters, Hannon said. He wants to know whether the item is contracted, how many cases the co-op bought and whether the move is part of a broader pattern.
"A 40% change in one case is very different from a significant price increase across 20 or 30 cases," he said.
Trust, he said, comes from validation. When a DDR flags a concern, he compares its output against ArrowStream and InsideTrack to confirm the numbers line up, pairing the system's pattern-spotting with human judgment.
Hannon treats performance much as he would an analyst's: Is the data accurate? Is it finding what he asked it to find? Is what it surfaces useful? Morgan's job description has grown along the way. She began by answering one question: "What changed more than 5%week over week? Then, she gained the ability to check whether an item was contracted and factored in purchase quantities.
"As I use it, I learn what context it needs and can continue refining what you want it to pay attention to," he said.
A team of specialists
Morgan is one of many. Hannon has deployed DDRs and apps covering pricing, commodities, food safety incidents, distributor fill rates, transportation and freight costs, manufacturing purchasing and market intelligence. Rather than applying a single model across every concept, he configures each DDR around a specific job, concept or area of responsibility.
"Morgan isn't one AI trying to do every job for every concept," Hannon said.
The specialization matters because context changes what the numbers mean. Ice cream purchasing rises in warmer months, he said, and if higher-cost products make up a bigger share of the mix, average case price can climb even when there is no pricing problem. The team can analyze data by concept, distributor, product category and other variables to tell the difference.
Hannon's rule of thumb for what becomes a DDR is still evolving. An audit works for periodic looks at commodity pricing, food safety incidents or distributor fill rates. A DDR makes sense when he wants the system to keep evaluating that information and report what has changed, what is unusual and what may require action.
"Often I'll start with a report or dashboard, understand what information is useful, and then think about how I can make that process more proactive," he said.
Connecting headlines to the P&L
The newest effort is a "headlines" experience designed to connect breaking developments with the co-op's own purchasing history, market conditions and forecasts.
Public information about beef, fertilizer or diesel prices is plentiful, Hannon said. The harder question is what a development means for Sky Co-op specifically. The goal is to mitigate risk by adjusting inventory, freight decisions and even the timing of restaurant promotions.
For example, before launching a specific promotion, he wants to know what the co-op purchased during the same period last year, what those products cost then versus now, how seasonality might affect demand and what current market conditions mean for the economics. An ice cream concept in July behaves very differently than it does in February.
The longer-term aim is to look forward rather than backward.
"If trend continues, what should I expect next month and how can I make it better?" Hannon said.
What it has and hasn't proven
Hannon is candid about the limits. He said it is too early to put a dollar figure on savings, especially across concepts with such different product mixes.
"It's still too early for me to responsibly say, 'The DDRs have saved us X dollars,'" he said.
What he can quantify is why small errors matter. In a hypothetical restaurant with $1 million in sales and a 5% profit margin, or $50,000 to the bottom line, even several hundred dollars in extra annual costs eat a meaningful share of profit. If a mispriced item runs six or eight weeks across multiple locations before an audit catches it, those costs add up. Catching it nearly instantly stops the erosion far sooner.
There are practical hurdles, too. Hannon cited data connections and the amount of historical information the system can process at once. Adoption is the other challenge. He has given team members access but said he is probably the group's most aggressive early adopter.
His answer is to solve real problems rather than push the technology.
"Rather than telling someone they need to use AI, you show them something that previously took weeks to identify," he said, "and now we have the potential to see it much sooner."
How operators can start
Hannon's experience points to a few practical steps for operators who want to try something similar:
- Start with a problem you can name. Hannon's trigger was a specific price jump that slipped through for weeks. Pick a recurring audit or report that is always late.
- Build the report first, then make it proactive. A dashboard shows what information is useful. Once you know that, you can ask the system to watch it continuously.
- Start with a simple rule, then add context. Morgan began with a 5 percent weekly threshold. Contract status and purchase volume came later, cutting down on noise.
- Validate against source data. Check what the AI flags against the systems you already trust before acting on it.
- Keep humans in the decision seat. Let the AI flag and analyze while people decide, at least until you've built confidence.
- Specialize. Give each digital worker a narrow job and the context it needs, rather than one model for everything.
- Win over teammates with results. Show colleagues a problem that once took weeks to find and now surfaces quickly.
Hannon said his team is still learning where the technology works well today.
"That's the exciting part of being an early adopter," he said. "We're learning where the technology works well today and how to make it better by putting it to work against real operating problems."
What's next for fast casual?
Connect with industry leaders at the Fast Casual Executive Summitin Arlington, Texas, Oct. 4-6, and hear from leaders from Shake Shack, Chicken Salad Chick, Freddy's, Original ChopShop, Consolidated Concepts and more. Register here.
About Cherryh Cansler
Related Media
Subscribe
Get the latest news and resources from Fast Casual.
