2026 Is the Year PEO Underwriting Stops Running on Spreadsheets

Walk into the sales operations room of almost any PEO and you will find the same three tools: a payroll system of record, an email inbox, and a folder full of spreadsheets. The spreadsheets do the real work. One rates the workers' comp. One builds the admin fee. One models the benefits spread. Somebody copies the outputs into a Word template and that becomes the proposal.

This setup is not stupid. It grew up around a real constraint, which is that the inputs to a PEO deal arrive as PDFs and the only way to get them into a model was to type them in. The spreadsheet was the best tool available for a world where humans were the parsers.

That constraint is gone. We think 2026 is the year the industry notices.

The back office was designed around manual intake

Think about what happens when a broker sends a submission. An email arrives with a loss run from one carrier, maybe a second loss run from the prior carrier, an ACORD 130, a census export, and sometimes an experience rating worksheet. An underwriter or sales coordinator opens each attachment, figures out what it is, and starts transcribing.

Loss runs get keyed by claim: date of injury, paid, reserved, status, claimant, cause. Census gets cleaned up and mapped to class codes. The ACORD gets scanned for payroll by location and prior carrier history. If the loss runs span three policy years and two carriers, someone has to reconcile valuation dates and aggregate by policy year, and that someone is usually doing it in a tab called "LR Summary v3 FINAL."

Every step is a chance for error, and more importantly, every step takes time. In our own PEO, a submission that arrived Monday morning could easily become a proposal Thursday afternoon, and the delay had nothing to do with underwriting judgment. The judgment took twenty minutes. The data entry took days.

Brokers feel this delay as silence. They send a submission, hear nothing, and start working the next PEO on their list. The deal gets slower for everyone, and the PEO that responds first with a credible number usually wins.

What "AI-native" means in practice

The phrase gets abused, so let us be specific about what we mean when we say a platform is built for AI from the ground up, as opposed to a platform with an AI feature bolted onto the side.

Intake is the first place it matters. A document arrives and the system classifies it before a human touches it: loss run, ACORD 130, census, X-Mod worksheet, dec page. Known carrier templates are parsed deterministically, with text extraction and pattern matching, because a Travelers loss run looks the same every time and there is no reason to spend AI inference on it. Unknown formats fall back to a model. The output is structured data, matched to a company record, routed to the module that needs it. In HALO IQ that is what InboxIQ does, and the reason it exists is that brokers send submissions by email and are not going to stop.

Rating is the second place. Once the loss data is structured and aggregated by policy year, the mod projection and the loss pick are computations, not spreadsheet gymnastics. Loss development, Bornhuetter-Ferguson, credibility blending, upward loads for open reserves and payroll growth: these belong in code that runs the same way every time and shows its work.

Narrative is the third. The executive summary is the page a prospect reads, and today it is often the weakest part of the proposal because it is written last, at the end of a long week, by someone who has written forty of them. A model drafting from the structured deal data gets the numbers right and gives the rep something to edit rather than a blank page.

Underwriting judgment is the fourth, and this is where we have been most deliberate. For risk decisions we run more than one frontier model on the same deal context, score the responses, and merge them. One model's confident answer is not enough when the question is whether to take a roofing account with a climbing mod.

None of this replaces the underwriter. It removes the transcription that stood between the underwriter and the decision.

The spreadsheet has hidden costs beyond speed

Speed is the obvious argument, but it is not the one that should worry a PEO owner most.

The first hidden cost is inconsistency. Two underwriters with two copies of the same workbook will drift apart. One updates the LDF tab, the other does not. Pricing that was supposed to protect margin ends up depending on which person caught the deal. Rules that live in a person's head do not survive that person leaving.

The second is auditability. When a carrier pushes back on a loss pick, or when the loss fund comes up short at year end, the question is always "how did we get this number?" A spreadsheet answers that question badly. The inputs have been overwritten, the version history is a filename, and the person who built it has moved on. A system that persists every input and every intermediate calculation at quote time answers it in seconds.

The third is the handoff between sales and the carrier. Assembling a submission packet for a carrier market means pulling the ACORD, the loss runs, the mod, and a narrative back out of the spreadsheet world and into an email thread. Status disappears into that thread. Nobody knows whether the carrier has opened it.

The fourth is the broker relationship. A broker working the same account with three PEOs has no visibility into any of them. They re-key the same data into three different intake forms, or more likely email the same PDFs three times and wait. That friction is not a sales problem to be solved with more follow-up calls; it is a data problem.

Why this year

Three things have lined up.

Frontier models have become good enough at document understanding that AI fallback parsing for an unfamiliar loss run format actually works. Two years ago it was a demo. Now it is a reliable second line behind deterministic parsers.

Model costs have fallen far enough that running two models in parallel on every underwriting question is an operational expense rather than a research budget.

And the industry has quietly reached the limit of what the existing stack can do. Payroll platforms are good at payroll. They were never meant to be underwriting or proposal tools, and the spreadsheets that filled the gap have become the bottleneck. PEO operators we talk to are not asking whether the back office should change; they are asking what the alternative looks like.

What we expect to see

We expect the time from submission to proposal to collapse. Not from days to hours, but from days to something closer to half an hour for a clean submission, because the parsing, rating, and narrative drafting happen as a sequence rather than as three separate manual projects.

We expect underwriting to become more consistent and more defensible, because the methods live in code rather than in personal workbooks, and because every number carries its inputs with it.

We expect brokers and PEOs to work from the same deal record with their own views of it, rather than passing attachments back and forth. Neither side is the other's customer; both want the account closed. The workspace should reflect that.

And we expect the spreadsheet to survive, but in its proper role: as a place to explore a what-if, not as the system of record for a book of business.

Where HALO IQ fits

We built HALO IQ because we ran a PEO on the old stack and got tired of it. The platform covers intake, mod analysis, carrier rate intelligence, pricing rules, proposal building, and CRM in one tenant, with AI in the parsing, the underwriting, and the narrative rather than attached as a chatbot.

Over the rest of this year we will use this blog to work through the details: how to read a WCIRB worksheet, what goes into a manual rate, where loss runs hide problems, what a proposal has to contain to win. Some of those posts are about the product. Most of them are about the craft, because the craft is what the software has to get right.

If you run a PEO or place business with one, we think you will find the back office looks very different by December.