On July 17, Travelers Companies (TRV) reported its second-quarter numbers. Core income came in at $10.04 per share against a consensus near $5.41. The stock jumped roughly 10%. In general, TRV stock has done tremendously well around its earnings:
The important part was in the Q&A.
Goldman's Rob Cox asked whether the claims straight-through processing work and the AI initiatives were showing up in Business Insurance margin yet. CEO Alan Schnitzer said yes. He put a number on it. Half a point of improvement in the underlying loss ratio, and he said it was big enough that it was worth drawing attention to. Here's the full quote:
I think the answer is yes. I mean, that was the reference in Greg's script, Dan referred to it. So we are seeing benefits coming from the investments in the innovation. I mean it's in the -- it was a 0.5 point improvement in the underlying loss ratio. And there's a couple of things going one way or the other in there, but that was big enough that it was worth bringing some attention to it. And I wouldn't say this is the first time it's ever appeared. It's just -- it's in the context of some other pieces that are moving one way or another. But we are very clearly generating benefits from the innovation and AI investments that we're making. - TRV 2Q26 Earnings Call
In the same quarter, the expense ratio went up to 29%, as we can see in the handy overview below.
Think about that for a second. The entire market is waiting for AI to shrink headcount and compress SG&A. Travelers made money on the claims line while its overhead got more expensive. Note that SG&A means "selling, general, and administrative expenses.
That's the core of this article.
And as I have kept you a bit in the dark about why all of this matters and how you can benefit, I would say we get right to it!
The Thesis Wall Street Is Running Right Now
Jordi Visser's framework has been making the rounds, and it is worth taking seriously. He calls it "QE for the Mind."
It was what triggered my research for this article.
The argument goes like this. During the zero rate era, companies manufactured EPS growth through financial leverage. They borrowed at 2%, bought back stock yielding 6%, and watched earnings per share climb without having to grow volumes/revenues. In the AI era, companies manufacture EPS growth through operational leverage instead. Basically, enterprise agents strip out administrative friction.
The kicker is the part that got everyone's attention. The biggest winners are not the software companies building the tools. They are the legacy, low-margin, high-headcount operators. When a company earns 1% on a hundred billion dollars of revenue, a small cut to overhead produces an enormous percentage move in earnings.
That's where many see the money. Just think about it: if a company that has tens of billions (or hundreds of billions) in annual sales and super low margins is able to improve margins even by a little bit, you'll see an explosion in EPS.
I ran the screen. It works beautifully on paper.
CVS (CVS), for example, earns roughly 0.4% net margin on $402 billion of revenue against $45 billion of SG&A. If you cut that SG&A by 3%, tax the savings at 23%, add them back, and GAAP EPS goes from $1.36 to $2.17. That is a 60% increase. At a 5% cut, you get 100%.
For the same scenario, Kroger (KR) came out at 64% and 107%.
I almost published those numbers. However, I am glad I did not, and the three reasons why are more useful than the screen ever was. That's where it becomes both educational, helpful, and tricky.
Reason One: The Torque Was A Write-Down
CVS reported GAAP EPS of $1.36 for 2025. Adjusted EPS was about $6.75. The company guided 2026 to $7.00 to $7.20 adjusted. Now, that guidance has been raised to the $7.30 to $7.50 range.
Either way, the gap is a $5.7 billion goodwill impairment on the Health Care Delivery unit taken in the third quarter, plus litigation charges earlier in the year. If you run the identical 3% cost cut against the adjusted number, CVS produces roughly 12% of EPS tailwind instead of the fictional 60% from my example.
Kroger is worse, and the detail is almost too good.
Kroger's fiscal 2025 GAAP EPS was $1.54. Adjusted was $4.85. The difference includes $2.5 billion of impairment charges worth $2.91 per share on the automated fulfillment network.
Read that again. The charge that manufactured Kroger's spectacular AI torque came from a failed warehouse automation project. The company spent billions on robots and wrote it off. That writedown then made Kroger look like the market's purest automation opportunity.
Here is the general rule, and it is worth memorizing if you ask me. When you rank companies by EPS torque using trailing GAAP earnings, you are ranking them by who took the biggest writedown last year. You are not measuring operating leverage this way.
Reason Two: Most Of That SG&A Cannot Be Touched
Kroger's $28 billion cost pool is reported as operating, general and administrative expense. It is dominated by store labor. Walmart's (WMT) $148 billion line is the same story.
Cutting 3% out of Kroger's OG&A means removing something like 12,000 store jobs. That is a robotics problem with a capital budget and a multi-year timeline. Language models (the AI models most companies use) do not stock shelves, run registers, or unload trucks.
My own screener selected these companies precisely because they employ hundreds of thousands of hourly workers. That is the labor AI cannot reach yet. And the numbers reflect it.
Reason Three: The Savings Do Not Stay
This one matters most to me because it connects to how I think about every business I invest in.
I rebuilt the screen to fix the first two problems. I demanded normalized earnings, and I demanded that the cost pool be structured white-collar work. Claims adjudication, prior authorization, chargeback reconciliation, invoice matching, order-to-cash. You know, stuff that AI can touch.
The rebuilt screen returned six names, and all six were health insurers. UnitedHealth (UNH), CVS, Humana (HUM), Centene (CNC), Elevance (ELV), and Cigna (CI).
That looked like a finding for about ten minutes. Then I remembered how these companies get paid.
Medicaid managed care rates are set by states on an actuarial soundness basis and rebased on observed costs, including administrative load. Medicare Advantage bids work the same way. Lower your admin cost and within one or two rate cycles you either hand the savings back through a lower bid or you spend them on richer benefits to defend share.
According to my numbers, government programs are now the majority of revenue at Centene and Humana and a large slice at UnitedHealth and Elevance. Regulation is not pricing power in this case. If you think about it that way, regulation is the mechanism that takes the savings away from you.
What Travelers Actually Showed
Which brings me back to that half point of loss ratio.
Every version of this thesis, mine included, assumed AI attacks the overhead line. Travelers put a number on AI improving the loss line while overhead went the other way.
That is a better outcome for four reasons.
The pool is bigger. Travelers runs an expense ratio around 28.5% against a loss ratio in the mid-50s. The loss line is roughly twice the size of the overhead line.
It compounds. Firing people is a one-time step down in the cost base. You do it once, and the benefit stops growing. A loss ratio improvement is a permanent ratio gain that applies to every dollar of premium you write from here forward.
It is easy to execute. No layoffs, no severance charges, no organizational redesign, no headline risk.
And the same logic scales everywhere. UnitedHealth carries roughly $370 billion of medical costs. Fifty basis points of medical loss ratio improvement is worth more than 10% of net income, which dwarfs anything you get from trimming the admin line. Kroger cited lower shrink and better sourcing as gross margin drivers this year. Shrink and markdown optimization sit in cost of goods, not SG&A.
So the revised view is this. AI shows up in earnings through better decisions on the cost of goods, and the market is watching the wrong line.
I had that wrong for a while. My original framework explicitly excluded cost of goods on the logic that agents automate workflows, not raw materials. That was correct about automation and wrong about economics. Using commodities as an example, AI does not shovel the coal. Instead, it decides how much coal you need.
How To Find These Yourself: The Four P's
I want you to be able to run this without a data terminal. Everything below is free and takes about ten minutes per company.
Proof. Pull the last four earnings call transcripts and search for "AI." If every hit is "excited about," "exploring," or "transformative," stop reading and move on. You are looking for a number attached to a named executive. A ratio, a dollar figure, a headcount figure. Most companies fail here, and that is the entire value of the test. Talking about AI and benefiting from AI are close to uncorrelated right now.
Placement. Ask which line the improvement landed on. Gross margin, loss ratio, shrink, medical costs, and cost of goods are the large pool and the benefit compounds with volume. SG&A and headcount are the small pool, and the benefit is a one-time step down. If a company only ever talks about reducing headcount, it is playing for the smaller prize.
Protection. Ask who sets the price. If a regulator or a government payer rebases your rates on your observed costs, the savings leave within two cycles. If the industry is fragmented and competes on price, you hand the savings to your customer. You keep the money only when you have contracts, switching costs, or an oligopoly. This is the same toll road test I apply to everything, and it does more work here than anywhere else.
Price. Compare the forward multiple to the five-year average. If the stock has already rerated, the market got there before you did.
And then there are four red flags that should stop you:
Torque calculated off GAAP earnings that contain an impairment. A cost pool that is mostly frontline hourly labor. A revenue base priced by a government payer. And a management team that mentions AI on every call and has never once quantified anything. To me, it seems that the market is waking up to the last point as well.
Three Names, And Only One Of Them Is Clean
Travelers. Passes all four. The disclosure is specific, recent, and came from the CEO in response to a direct analyst question rather than a prepared remark. Placement is on the loss ratio, which is the big pool. Protection is strong because Travelers has a genuine underwriting franchise and has said plainly that competing on price is not a strategy it will pursue. Price is reasonable in the low teens on forward earnings.
The company spends roughly $1.5 billion a year on technology, which is the scale advantage doing the work. Schnitzer was careful, and I want to repeat his caution rather than bury it. He said you cannot apply that half point to any particular quarter or even necessarily a full year. This is a direction of travel, not a run rate.
Equifax (EFX). The best disclosure in the entire U.S. market right now, and it improved recently.
In February, the company targeted $75 million of annual cost savings from AI, mostly in operations, call centers, and paper processing. On the July 22 call, Begor doubled it to $150 million of run-rate savings across 2026 to 2028. When an analyst asked what changed in only a few months, the answer was that agent deployment in operations went faster than planned and the coding and software development benefits arrived on top.
The $75 million figure also sits in the 2026 proxy statement, which is a filed document rather than a call comment. That is the highest standard of evidence in this entire exercise.
Protection is excellent. Equifax sits in a three-firm credit bureau oligopoly, and roughly 90% of revenue comes from proprietary data that only Equifax can point AI at.
Price is the problem. Equifax is the most expensive of the reasonably valued names here, and revenue is exposed to a mortgage market that keeps getting pushed out by rates. However, if we look beyond the fact that it trades at its longer-term P/E ratio, we see that analysts expect close to 20% annual EPS growth in both 2027 and 2028. If the company executes, it could be a great buy.
Allstate (ALL) is the one that teaches the framework.
Allstate passes Proof and Placement convincingly. QuickFoto Claim handles roughly half of driveable auto claims and pulled estimating cycle time from five to seven days down to under 24 hours. Generative AI now drafts something like 50,000 claims communications a day across a 23,000-person claims workforce, per Fortune. Arity has more than a trillion miles of driving data feeding pricing and liability decisions. The fourth quarter was enormous, with adjusted EPS of $14.31 against a $9.86 consensus and operating margin nearly doubling year over year.
Now read what management said about where the money went. They attributed the result to expense reduction and claims process improvements that allowed them to price more competitively without sacrificing margins. The SAVE program lowered premiums for 7.8 million customers.
That is the Protection test failing. Allstate is converting its AI advantage into lower prices to win policies. Over time, that may be the smarter strategy, because policy growth compounds and a cost advantage in a price-competitive market eventually shows up as share. But the shareholder is not collecting the savings today. The customer is.
I own no position in any of the three. I am watching all of them.
The Three That Already Worked
Worth knowing what the proven version looks like, because it tells you what you missed.
C.H. Robinson (CHRW) has cut total employment by roughly 31% since 2022 while deploying hundreds of agents that open and sort freight orders, classify freight, and book appointments. Agents have handled more than three million shipping tasks, including over a million price quotes. It trades north of 30x forward earnings.
Waste Management (WM) has been writing "permanently reducing our cost to serve through technology and automation" into its filings since at least 2023. Recycling EBITDA grew 22% last year while commodity prices fell 20%. Roughly 29x forward.
Republic Services (RSG) is around 30x.
The pattern is uncomfortable, and I think it is the most important thing in this article. Where AI cost reduction is proven and disclosed, it is fully priced. Where the stock is cheap, the benefit is unproven. There is no third bucket.
Don't get me wrong. I love these companies, and I would like to own WM/RSG. However, because they are textbook examples of AI adoption, you're paying a premium.
One more observation on why those three were legible at all. C.H. Robinson's productivity gain showed up during a freight recession. Waste Management's recycling gain showed up while commodity prices fell. Flat or falling revenue is what made the improvement visible.
That gives you a screening rule almost nobody is using. Look for AI cost improvement in businesses with flat top lines, because that is the only place you can actually see it. If UnitedHealth's margins recover over the next three years, nobody will be able to separate AI from medical cost trend normalizing, and the thesis will never be provable.
And then there's a layer of risk to my approach.
What Would Break This
Property and casualty pricing is softening after seven years of a hard market. As rates come down, efficiency gains get competed into price, exactly as Allstate is already showing us. That is the largest risk to the Travelers case.
AI capital expenditure offsets AI savings. Travelers spends $1.5 billion a year on technology. Equifax has spent years and enormous sums on its cloud migration. None of this is free, and the net number is what matters.
And disclosure is not verification. A CEO saying AI delivered half a point of loss ratio is a claim, not an audited fact. Schnitzer himself said other things were moving in both directions inside that number.
One housekeeping note for the income-focused readers. None of these three are yield ideas. Travelers and Allstate pay around 2%, and Equifax pays less than 1%. Treat this as a quality and total return sleeve.
Takeaway
The popular AI trade says buy thin-margin, high-headcount operators and wait for the headcount to fall. I tested that, and it did not survive. And, to be honest, I did not expect that to be the result of my research before I started writing.
The torque was accounting damage, the cost pools were frontline labor that agents cannot reach, and in regulated businesses the savings go back to the payer.
The first real evidence points somewhere else. Travelers improved its loss ratio with AI while its expense ratio got worse. Equifax doubled a filed cost savings target because agents deployed faster than planned. Both improvements sit in large pools that compound with volume rather than small pools that step down once.
If you have some time and are looking for investment ideas beyond the ones I presented today, run the four Ps yourself. Proof, meaning a number from a named executive. Placement, meaning which line it hit. Protection, meaning who gets to keep it. Price, meaning whether the market already knows.
Most companies fail on the first test.
And needless to say, I will keep running this test, as AI is changing rapidly.
Stay tuned!