Employer health costs jump 8.2% in 2027, the steepest since 2003. Marsh’s actuary names two brand-new drivers, and neither one is a drug.
The answer, up front
Marsh told 1,800 US employers what 2027 looks like: an 8.2% increase in health benefit cost per employee, the biggest since 2003, and that is after the cuts employers already plan to make. Before those cuts, the number is 11%. GLP-1 drugs account for one point of it. The two new entries on the driver list are AI billing software on the provider side and out-of-network arbitration awards under the No Surprises Act. Both are cases of the counterparty getting better at extracting money, and roughly two-thirds of covered workers sit in self-funded plans, so the employer’s own cash pays the difference.
The hook
Every year a benefits consultant publishes a trend number, every year the business press writes it up as “health costs rise again,” and every year the story lands on the same villain. In 2023 it was hospital prices. In 2025 it was Ozempic. This year the headlines went to GLP-1s again.
Read Marsh’s actual release and something else is sitting in the middle of it. Sunit Patel, the firm’s US chief actuary for health and benefits, sizes GLP-1 utilization at exactly one percentage point of 2027 cost growth. One point, out of eleven.
Then he names what else changed. The rapid adoption of AI-enabled billing software that helps physicians submit claims, which has produced more claims and higher-level claims. And out-of-network payments awarded through the arbitration process Congress built into the No Surprises Act, running larger than anyone modeled.
Neither of those is a treatment. Neither of those is a drug. They are both the sound of the other side of the table getting a better tool.
What Happened
On September 2, Marsh, the benefits consultant formerly known as Mercer, released preliminary results from its 2026 National Survey of Employer-Sponsored Health Plans. Responses came in from more than 1,800 US employers between the June 10 launch and August 10.
The topline: total health benefit cost per employee rises 8.2% in 2027, the highest since 2003, after employers make planned cuts. Without those cuts, 11%. This year’s comparable figure was 6.7%. It will be the fifth straight year of elevated growth after a decade of mild ones.
More than a third of respondents told the New York Times they expect 10% or more even after cutting. Fifty-nine percent plan cost-cutting changes for 2027. About two-thirds of employers with 500 or more workers expect to raise the employee share of premium, which means a lot of paycheck deductions climb by more than 8.2%.
Marsh is not alone. Aon put the increase at 9.5% and the average cost above $19,000 per employee if nothing changes. Business Group on Health has 9.2% median, falling to 8% after benefit changes, and reckons costs could climb 76% between 2018 and 2027, roughly double general inflation.
The chart below is the part nobody framed.
Marsh publishes both numbers every year. Eleven percent is what the plan costs. Eight point two is what employers will actually book. The 2.8-point gap between them is not a forecasting range. It is the benefit reduction, quantified and printed, and nobody names it.
Run it against Aon’s per-employee base and 2.8 points is roughly $486 per employee per year taken out of the plan. That is an order of magnitude, not a coefficient, since the two surveys poll different samples. But the direction is unambiguous, and it gets worse when you remember only 59% of employers are doing the cutting. Concentrate the same average into the cutters and they are stripping closer to 4.7 points, or something like $823 a head.
Which makes the GLP-1 arithmetic almost funny. The drug class that got the headlines contributes one point. Employers are removing 2.8. They are cutting nearly three times as much benefit as the villain adds, because the villain is the one line item they can actually find and touch.
The Backstory: How the Price of an Argument Collapsed
The No Surprises Act took effect on January 1, 2022. It stopped providers from balance billing patients for out-of-network emergency care and for the anesthesiologist you never chose. Popular law, real protection, and it worked.
What it did not do was set a price. Instead it built an arbitration lane. If a plan and an out-of-network provider cannot agree, either side files for independent dispute resolution, both submit a number, and a certified arbitrator picks one. Baseball arbitration for medical claims.
Regulators expected about 22,000 disputes in year one. They got 489,000, fourteen times the forecast. As of January 31, 2026, more than 5.1 million disputes had been filed. In January 2026 alone, parties opened 248,452 of them, which is 11.3 times what the government budgeted for an entire year of the program.
Then, on May 28, 2026, the Departments of Health and Human Services, Labor and Treasury finalized a rule cutting the administrative fee to initiate a dispute from $115 per party to $15. Effective June 11.
That is an 87% price cut on the act of arguing. The Departments said so plainly: the old fee created barriers for providers challenging underpayments, and the lower one makes it economically viable to contest smaller amounts that previously would not justify the cost.
Read that as a business person and it is a market-expansion move. At $115, a $400 underpayment is not worth disputing. At $15, it is. The government just enlarged the addressable market for claim disputes, and Brookings work by Loren Adler’s team has already found that arbitration awards run well above typical in-network rates in the same communities, so a larger pipeline is a larger transfer.
Now stack the timing. Marsh opened its survey on June 10, 2026. The fee cut landed June 11. Employers priced their 2027 renewals in the exact window when the cost of contesting a payment fell by seven-eighths.
The Plan: Two Software Budgets, One Trend Line
Here is the piece that makes this a business model story rather than a health policy story.
Roughly 63% of health care organizations now use AI or automation somewhere in the revenue cycle, according to a July 2026 synthesis of published HFMA, Bain, Deloitte, CAQH and Gartner research. Only about 27% run it at scale. Only about 15% report positive ROI on their own books.
Hold that next to what Marsh’s chief actuary just said. The technology whose returns most providers cannot yet find in their own P&L is showing up as a named driver in somebody else’s biggest cost increase since 2003.
That is a juxtaposition, not proof of causation, and I will not pretend otherwise. But the shape is familiar to anyone who has watched a technology’s value migrate. The ROI is real. It is just landing on the other side of the transaction, where the buyer of the software does not have to book it and the payer of the bill did not buy the software.
And the payers are not insurers. Sixty-seven percent of covered workers are in self-funded plans, per KFF, rising to 80% at large firms. In a self-funded plan the carrier is an administrator collecting a fee, not a risk bearer. Every upcoded office visit, every arbitration award, every extra claim comes out of the employer’s own cash, the same way a captive insurance structure puts the underwriting result back on the parent’s balance sheet. This is an operating expense line, not an insurance premium.
Meanwhile the defensive spending is climbing too. Cedar puts the cost of working a single denial at $57.23 per claim in 2026, up from $43.84 the prior year, a 30.5% jump in twelve months. Payers deny with software. Providers appeal with software. The arms race is real and both sides’ budgets settle into the same trend line, which the employer pays.
Nobody in that loop is treating a patient.
The Business Model Angle
Value capture is moving to the revenue cycle, not the exam room. Clinical AI has spent a decade struggling to prove it improves outcomes at a price a payer will fund. Billing AI proves itself in one quarter, on a metric the CFO already tracks. That is why it is scaling first, and why the health care AI companies worth watching are the unglamorous ones sitting between the chart note and the claim form.
When a regulator cuts the price of a process, it creates a market. The IDR fee went from $115 to $15 and the volume was already 11x the forecast. Any time a rule change lowers the marginal cost of initiating a transaction, somebody builds a business on the new volume. Arbitration filing, appeal automation and underpayment recovery are now a category, and they are growing because a federal fee schedule made them viable, not because a customer got smarter.
The self-funded employer is an unrepresented buyer, which is a market. Two-thirds of covered workers sit in plans where the employer bears the risk and an intermediary handles the money. Marsh’s own survey shows the buyer waking up: 12% of large employers will offer variable copay plans in 2027, guiding members to high-value providers has jumped from fifth priority to top three, and 58% now call it important. Miami-Dade County Public Schools, 45,000 employees, is contracting directly with hospitals for imaging and demanding claims audits.
The intermediary layer is where the defection starts. The National Alliance of Healthcare Purchaser Coalitions found 54% of employers using one of the three largest pharmacy benefit managers, down from 63% a year earlier. That is a nine-point drop, a 14% relative decline in a single year, in a category everyone assumed was structurally locked. When buyers cannot cut the cost of care, they cut the cost of the people standing between them and the care.
Health benefits are becoming a cost structure problem, not an HR problem. At 8.2% compounding, per-employee benefit cost doubles in under nine years. Wage budgets are planned at three or four. Every operator running a US payroll should model what happens to their total compensation line when one component grows at triple the rate of the other, because the answer is that benefits eat raises.
The Risk: Three Ways This Reading Is Wrong
The catch-up argument. Over the last five years KFF’s average family premium rose 24%, against 26% wage growth and 28.6% inflation. Health benefits actually lagged both. Some of 2027 is a delayed repricing of contracts written when trend ran below inflation, and framing it as a runaway is at least partly a base-year trick.
The upcoding might be correct coding. When Marsh says AI billing software produces “higher-level claims,” that is not automatically a scandal. Physicians have chronically under-documented complexity for years, and a tool that captures what actually happened in the room is capturing revenue that was always earned. The honest position is that nobody outside the payers has the claim-level data to separate legitimate capture from drift, and neither the providers nor the vendors have an incentive to publish it.
It is a preliminary survey. Marsh’s final results come later this year. Responses closed August 10 and renewal quotes often land below the projection. Business Group on Health has 8%, Aon 9.5%, Marsh 8.2%, which is a tight cluster but a cluster of forecasts, not a settlement.
There is also a cheerful counterweight worth naming. GLP-1 prices are falling fast in the cash channel, and Novo’s own pill undercuts its injection while Washington fixed the Medicare price at $50 a month. Marsh’s actuary says as much: the market is evolving in ways that could ultimately lower costs. The employers dropping GLP-1 coverage for 2027, 6% having already dropped in 2026 with another 5% considering, may be cutting the one line that was about to get cheaper on its own.
Quick Questions
Is 8.2% really the highest since 2003? According to Marsh, yes, on total health benefit cost per employee, after planned plan changes. The unmanaged figure is 11%.
How much of this is GLP-1s? One percentage point of the 2027 growth, per Marsh’s chief actuary. Roughly one ninth of the raw increase.
What does “AI billing software” actually mean here? Tools that read clinical documentation and suggest or generate the codes on a claim. Marsh’s stated observation is more claims filed and claims filed at higher levels of service.
Why does the employer pay and not the insurer? Because 67% of covered workers, and 80% at large firms, are in self-funded plans where the carrier administers and the employer holds the risk.
Who is moving first? Smaller employers, per the National Alliance, where the pain is most acute. And large public plans like Miami-Dade schools, which is testing direct contracts with hospitals for services like imaging.
The Business Model Analyst Take
The most reliable way to find where a technology’s money actually goes is to stop reading the vendor’s case study and start reading the counterparty’s cost line.
Nobody put out a press release saying revenue cycle AI works. The vendors are still publishing adoption stats with a 15% ROI rate attached. But a benefits actuary at a firm with $27 billion of revenue just put it on a list of reasons 160 million Americans are about to pay more for the same coverage, alongside a federal fee cut that made arguing about claims 87% cheaper. That is not a product review. That is a payer noticing.
And the payer, in two-thirds of these plans, is a company that never bought the software, never signed the arbitration rule, and gets one lever: cut 2.8 points of benefits and hand the bill to its own employees.
The first genuinely profitable AI in American health care does not read scans. It writes invoices. And its customer is not the person paying for it.
