Weekly Editorial, W43
Deel found the best AI model fully solves 49% of real payroll tasks, and Paychex now pays tax penalties whoever caused them. Payroll vendors are selling correctness as a warranty or an included feature, never as a price, and every new payroll agent this week stopped at a person.
Theme 01
Two of the week's moves answered the same question: who pays when payroll is wrong. Neither answer was a published accuracy rate with a price beside it.
On 21 October Deel published EmBench Payroll: 295 real payroll tasks in 8 countries, graded by 29 in-country specialists. "The best model meets 90% of their criteria but fully solves only 49% of the tasks." Its own model, Pangaea P1, passed 61% of criteria on the 110 hardest, at about one fortieth of the top model's cost.
Deel put that saving into price, not margin. "That helps us keep AI included in your Deel plan, with no charge per question." The benchmark is public and open to any model builder. Deel also built it, staffs it and pays the graders, and no independent body has audited it.
Two days earlier Paychex launched Tax Penalty Protection: it pays eligible payroll tax penalties "regardless of cause", including the customer's own errors. Intuit has offered the same since May, capped at $25,000 with published exclusions. Paychex publishes no cap, no terms and no price.
A warranty priced across a whole customer book needs a loss rate only the vendor can see, and it gets cheaper per customer as the book grows. Paychex published an AI correction figure in August and chose not to improve on it in public. It chose to underwrite the outcome.
The other failure mode got no cover at all. On 21 October BambooHR took its platform offline for about six hours in the US, Canada and Ireland to patch PHP2SHELL, a critical flaw in the language much of the web runs on. Its status page said the shutdown interrupts "essential work, including payroll."
Read together: accuracy is now measured in public and still not priced. One vendor grades it and gives AI away, another insures the result, and an outage on a payday falls outside both. A buyer should ask for the cap, the loss rate and the plan for a shared-stack flaw.
Theme 02
HR Tech opened on 20 October with the suites putting agents on the pay run and inside other companies' assistants. The question every release answered was where the agent has to stop.
On 19 October UKG said its AI will "Identify unusual pay results, explain potential issues, and prepare corrections for payroll professionals to review." The release named no product and gave no date or accuracy figure. A review queue only catches what the AI flags; a missed error leaves the queue clean.
A day later isolved moved the line. For three kinds of routine fix (missing occupation codes, characters a file will reject, and new hires missing from the state new-hire file), administrators now choose whether its agent asks first or corrects alone. In August the approval step was mandatory.
All three fixes are filing hygiene at the edge of payroll. None touches a pay calculation. isolved measures the agent by "estimated time saved and errors prevented based on industry benchmarks", not by what it caught or missed on the customer's own runs.
The assistants moved too. UKG put its workforce data inside Claude in beta on 20 October, eight days after saying work would start there, with permissions inherited and the leader making the call. Workday's Sana went in a week earlier. Both large US suites are now in Claude before general availability.
isolved's Claude connector took a smaller step the same day. Employees can update contact details and submit a missed punch from Claude, each change shown and confirmed first, while "Pay, benefits and other records are view-only." Manager and administrator writes are announced with no date. That is the harder permission, because it changes someone else's record.
The pattern is consistent: the agent finds, explains and prepares, and a person approves. The vendors describe permissions in detail and accuracy not at all. Until someone publishes how often the reviewer changes the agent's answer, the review step cannot be shown to be doing work.
Theme 03
Three launches this week sold payroll, or the step before it, as part of a product the buyer already holds. Each one moves where the payroll decision gets made.
On 20 October Mercury, a business bank, launched payroll run from the account "in all 50 states, with no cap on how many and no higher support tier." It checks the live balance before each run. It is free for eligible customers through 2026, then $100 a month plus $50 for each person paid that month.
Mercury built on Central, which it bought in April, where Chase chose on 7 October to rent Gusto. One price for any number of states targets the per-state plan tiers small employers usually buy up to. Mercury had shipped its ledger without payroll a month earlier.
Zoho launched UK payroll the same day at £7.50 a month including five employees, then £1.50 each, with RTI, auto-enrolment, benefits in kind and an AI assistant in the single plan. It is a cheap module that pulls a small business deeper into Zoho's accounting and expenses. It lands five months before UK mandatory payrolling of benefits in kind.
The same move ran the other way. On 19 October Justworks added an applicant tracking system under its own name, with hires flowing into employee records and background checks through Checkr. TriNet rented its recruiting module a week earlier. The PEO now holds the candidate before it holds the employee.
In each case payroll is evaluated as part of a larger purchase: the bank account, the business suite, or the PEO contract that starts at the job posting. Standalone payroll and standalone small-business recruiting both lose the moment a buyer used to choose them on their own.
Next