The most instructive case this year of a large American company taking labor out of its cost structure has almost nothing to do with a model. Group 1 Automotive — one of the largest US auto retailers, running 145 US dealerships across 17 states and 106 in the UK — cut nearly 700 US full-time positions in early April, roughly 5% of the 13,563 US employees it reported at the end of 2025, and told investors on July 30 that it had completed a $50 million annualized expense reduction, exceeding its targets — a program that also included roughly $14 million of contract and vendor spending, not headcount alone.
Group 1 did not credit a machine for any of it. CFO Daniel McHenry told analysts the trigger was demand: coming out of January and February, "we could see some weakness in the market and our SG&A leverage at that point was much lower than we would have expected." These were cuts made into a soft market, and nobody should read them as a machine displacing labor.
The mechanism worth studying is the one running alongside it. Group 1's virtual finance-and-insurance desk took a role that had been physically stapled to a single showroom and made it a pooled, remote service. CEO Daryl Kenningham gave analysts the throughput number in plain language: "In an average day, an F&I manager might do 3 deals. As a virtual agent, they can do 7, 8, 9, 10, and that's kind of the numbers we see." Same people — many of the virtual managers used to work in Group 1's stores. Roughly triple the output.
That is not an AI story. It is a queueing story. And the difference determines whether the same play works in your company.
What Group 1 Actually Built
Group 1 unbundled a licensed, high-skill role from the building it used to sit in. F&I is the desk you visit after you agree on the price of a car: the manager arranges the loan, presents the warranty and protection products, and produces the contract. It is the single most profitable seat in a dealership — Group 1's US stores earned $2,456 of F&I gross profit per retail unit last quarter, against a total US new-vehicle business where margins are thin and shrinking.
Kenningham told analysts on the Q2 call that the virtual desk, introduced two years ago, is now "installed in 66 stores across the U.S., and in those stores, 20% of our F&I volume is virtual" — more than 40% of the store base, up from about a third three months earlier. Management's claimed results are faster transactions and lower compensation cost. On the April 30 call, SVP Peter DeLongchamps put the rest of it plainly: "we've lowered turnover, we've lowered comp, we've actually increased the PRU on what I'd say the bottom performers."
Read that last one carefully, because it is the tell. Pooling did not make the best F&I manager better. It made the worst one better, by routing deals to whoever was free instead of to whoever happened to be standing in that particular building on that particular Tuesday.
The Unlock Was 100 Pieces of Paper, Not a Model
The enabling change was document digitization, and Group 1's CFO said so on the record. Daniel McHenry told analysts the company rolled out a digital deal jacket "across 100% of our dealerships" in the first quarter: "Traditionally, we would have had a scanner and a dealership scanned in 100-ish pieces of paper that would have formed the deal jacket." Kenningham interjected the clarification that makes the whole story legible: "Scanner being a person."
Once the deal jacket is born digital, the F&I manager no longer has to be within walking distance of a printer, a signature and a filing cabinet. That is the entire unlock. The industry substrate arrived at roughly the same time: Cox Automotive reported in January that 86% of auto finance contracts are now eligible for digital submission, with 91% of buyers already completing some or all of the purchase online and 61% of contracts carrying at least one aftermarket product.
No large language model is required for any of this. What is required is that the artifact the expert works on stops being physical.
Why Pooling Beats Automating
An expert who serves one location is idle whenever that location is quiet, and the idleness is invisible because it looks like availability. This is the oldest result in operations research and it has nothing to do with software: pooling independent demand streams raises utilization far more than making any single server faster. A dealership that sells 90 cars a month cannot keep two F&I managers busy and cannot survive with one on a Saturday. So it overstaffs, and the overstaffing is structural.
Move the same people into a national queue and the arrival pattern smooths out. Three deals a day becomes seven to ten without anyone working harder, because the gaps between deals were the product — not the deals.
The pattern generalizes, and the best evidence outside retail comes from hospitals. Researchers at UNC Kenan-Flagler studied nearly 28,000 inpatient encounters and found virtual nursing — experienced nurses working from command centers or from home, handling admissions, verification, education and discharge for multiple units — cut average length of stay by 7.2% and 30-day readmissions by 2.0%. Read the magnitude as directional: it is a working paper, and the design compares two hospitals inside a single East Coast health system. The finding that transfers is the conditional one — the effect depends on the remote person being more experienced than the bedside staff, and where the virtual nurse was less experienced the researchers found no significant improvement at all. Pooling works when it concentrates scarce judgment, not when it dilutes it.
The Part That Can Still Go Wrong
The virtual desk did not save Group 1's quarter, and pretending otherwise would be the mistake. Consolidated F&I gross profit still fell 8.8% year over year to $216.8 million, US F&I profit per unit slipped 0.4%, and the stock dropped about 15% on the print as consumer affordability bit. A staffing-model change fixes cost structure. It does not fix demand.
There is a real quality risk too, and it is well documented on the vendor side of this market: one digital-F&I implementation guide reports that average dealers see product penetration fall 15% to 25% when they move the F&I conversation online without redesigning it — a vendor's framing, and one that happens to sell the remedy, but consistent with Group 1's own claim that the gains showed up among bottom performers rather than across the board. A manager whose skill is reading a room loses that instrument on a video call. Some of them will never get it back.
Steel-manning the other side: the roles most exposed here are also the ones where the physical presence was doing quiet compliance work. Someone watched the customer sign. Someone noticed the confused eighty-year-old. Digitize the artifact and you digitize the audit trail, which is better — but only if you actually instrument it, and the instrumentation is not free.
Finally, this is not a moat. Assurant launched F&I On Demand in January, selling pooled virtual F&I specialists as a service to any dealer with a staffing gap. What Group 1 built with two years of internal work is now purchasable. That is the normal arc: the operating-model advantage lasts until a vendor productizes it, which in this case took about twenty-four months.
Your Competitors Are Already Running This Play
Group 1's closest listed peer described the same play on its own Q2 2026 call, which is the best available evidence that this is structural rather than a one-company anecdote. Lithia Motors CEO Bryan DeBoer told analysts on July 29 that "our sales departments are re-architecting how they operate with combined roles, removing layers, remote functions" and — the line worth stealing for your own board deck — "these results reflect real structural changes, not one-time cuts." Lithia held F&I at $1,811 per unit while personnel cost improved 30 basis points as a share of gross profit.
AutoNation, reporting July 31, grew customer financial services profit per vehicle about 3% to $2,799 and told investors it expects adjusted SG&A to reach 66–67% of gross profit on a run-rate basis by year end. Group 1's US adjusted SG&A improved more than 400 basis points sequentially to 66.4%. But AutoNation credited productivity initiatives, advertising discipline and portfolio actions rather than role-pooling. The cost pressure is industry-wide; the answer to it is not yet.
How to Find These Roles in Your Own Org
The test is not "which jobs can AI do." It is: which experts are pinned to one location by an artifact, and what would it take to make that artifact digital?
Run the inventory against three filters. First, the role requires judgment or a licence, so you cannot simply automate it away — that is what makes pooling the answer instead of replacement. Second, its demand is lumpy at each site but smooth in aggregate. Third, the thing keeping it local is a document, a signature, a physical handoff or an instrument, not the work itself.
Contract review sitting in three regional legal teams. Credit exceptions approved by one person per branch. Specialist claims adjusters. Radiology, already thirty years into this. Tier-2 support engineers assigned to accounts rather than a queue. Clinical pharmacists verifying orders one hospital at a time. In every case the interesting question is the artifact, not the algorithm.
This Week: Pull utilization, not headcount, for your three most expensive licensed roles. Ask what fraction of a shift is spent waiting for work to arrive versus doing it. If nobody can answer, that is your finding — Group 1 could answer it in deals per day, which is why they could act.
This Quarter: Pick one role and name the artifact that pins it to a site. Fund the digitization of that artifact as its own project, with its own owner, and measure it in documents eliminated. Group 1's number was roughly 100 pieces of paper per deal. Yours is a number, and you should know it.
Before Your Next Renewal: Reprice the seats. If a pooled expert handles two to three times the volume, your per-seat software, per-agent contact centre licences and per-user platform fees are now attached to a different denominator. That is leverage with your vendor, and it expires quietly if you do not use it at renewal.
The Bottom Line
The last technology cycle taught executives to ask what a system can do instead of a person. That question produced a long run of layoffs that companies later regretted and a Gartner finding that most firms cutting jobs for AI never captured the return. Group 1 asked a smaller, older question — why does this person have to be in this building? — and got a tripling of throughput out of a document scanner and a routing rule.
The reason 84% of enterprises have not redesigned a single workflow for AI is that redesign is unglamorous. It looks like paperwork, because it is paperwork. Standard Chartered's back-office restructuring and OpenAI's research on how AI crosses job boundaries both point at the same seam: work is being unbundled from roles and roles from places, and the model is rarely the binding constraint.
Before you fund another pilot, go find the person whose job is scanning 100 pieces of paper. They are not in the AI budget. They are the reason it will not pay.
Continue Reading
- They Fired Workers for AI. 55% Now Admit It Was a Mistake.
- 80% Cut Jobs for AI But Got No ROI: Gartner Study
- 84% of Enterprises Haven't Redesigned Workflows for AI—Yet
- Standard Chartered Cuts 7,800 Jobs to Hit 18% ROE With AI
- AI Crossed 44% of Job Boundaries. Is Yours Next?
- Suleyman's 18 Months: Score Your Job's AI Risk (1-25)
- Cisco Personal AI Agents: 90,000 Employees, CFO Playbook
- Cresta Knowledge Agent: Why Augmentation Beats Automation for Contact Centers