The two totals everyone checks are the two most likely to match, because a rules difference often moves money between codes without changing the sum. Overtime built on a different regular rate, a shift differential in or out of the base, a deduction sequenced differently — gross and net reconcile, and the FLSA exposure sits underneath them.

A benefits or 401(k) feed that drops records rather than erroring is worse than one that breaks, because nothing alerts and the gap compounds every cycle until an audit or a participant finds it. I have rebuilt one of these from scratch in the middle of a cycle when the vendor interface gave up.

Payroll is signed off on whether people were paid correctly. Finance cares where it landed — cost centre, GL account, project, fringe allocation. That half is usually validated last, by fewer people, with less time, and it is the half that shows up in a variance review three months later.

A validation run on a normal payroll proves the normal payroll. Retro pay, off-cycles, terminations mid-period, arrears, negative net, third-party sick pay, a garnishment that started on the wrong date — those are where configuration differences actually surface, and they are the rows a sample is least likely to contain.

Reciprocity, local taxes, work-in versus live-in, SUI state assignment. Nothing goes wrong while everybody stays put. It goes wrong on the employee who relocated in March, and it stays wrong all year.

Group-term life over the threshold, personal use of a company vehicle, relocation, gift cards somebody in another department handed out. These are not payroll’s data and they arrive late, which is why they become a Q4 emergency and a W-2 correction.

Twenty employees in a spreadsheet is the published advice, and it is reasonable advice for catching a gross mistake. It is not a method for finding a rule that affects one earning group, or one state, or the forty-seven people whose schedule crosses midnight.

What it is, how it taxes, where it posts in the general ledger, whether it belongs in the regular rate for overtime, and how it behaves in a retro. The first three get tested. The last two are found at quarter end.

Registration, tax codes, reciprocity, local taxes, work-in versus live-in, SUI assignment. Standing up a state before the first employee is paid in it costs a fraction of correcting a year of filings afterwards.

Rounding, shift differentials, premiums, callback, holiday and what counts as hours worked. This is where the number is actually decided; the pay code is only where it becomes visible.

Entitlement, carryover, caps, service-based tiers, and the balance that has to be right on the day somebody asks for it rather than at the end of the year.

Effective dating, approvals, and the rules that stop bad data being entered at all — which is always cheaper than the report that catches it two cycles later.

Company structure, pay groups, calendars, the full code set, the interfaces out, and a parallel run that proves the whole thing before anybody depends on it.

Two stints running bi-weekly payroll for a global electrical manufacturer with roughly 4,000 employees — over $14M gross a cycle, plus executive compensation, stock options and non-qualified deferred comp. GlobalView is SAP payroll delivered as a managed service by ADP, so this and the ADP work below are the same estate seen from two sides. Earlier, payroll on an SAP implementation for an automotive quality-services firm.

The deepest of these. Configuration end to end in UKG Pro — business rules, timekeeping, earning and deduction codes, pay rules, accruals — as well as running on it. Time and attendance feeding payroll is where most of the interesting defects live, because the error is created in a timekeeping rule and only becomes visible in a pay code.

401(k) and pension reconciliation through Dayforce audit forensics, including the root-cause work behind a $1.2M systemic error that had been reappearing rather than resolving.

Maintaining the time rules and accrual policies in IFS, then exporting what they produced for import into the payroll system. Accruals are the quiet one: a rule that computes the wrong entitlement is a liability that grows every cycle, and it is usually found by an employee being told they have leave they do not, or do not have leave they do.

Payroll on Lawson in a large hospitality-services environment. Older stacks are not a lesser problem; they are usually a worse-documented one, and the people who know why a rule exists have moved on.

Mapping how payroll actually lands in the general ledger on a web-based Oracle ERP at a global manufacturer: which earning and deduction codes hit which accounts, cost centres and company codes, and where the posting stops agreeing with what payroll paid. Costing is the half that gets validated last, by fewer people, with less time — and it is the half that surfaces in a variance review three months later.

ADP is several products rather than one, and the useful experience is across them: GlobalView for the payroll itself, eTime for timekeeping, SmartCompliance for garnishments, tax filing and record retention, and the payroll and HR data moving between those and everything else. The interface layer is where the interesting failures are, because data leaves one system correct and arrives in another incomplete without anything erroring.

A year as an ERP systems analyst on Dynamics AX at a manufacturer. The work was reporting: taking what the system already knew about production and turning it into daily reports that arrived on the right desks without anybody having to ask. Before that, people were each running their own version of the same query — which is how two departments end up in a meeting with different numbers for the same week.

A canned report or the system’s own report writer, a SQL view against a reporting replica, a scheduled SFTP feed, an ODBC connection, or the vendor’s REST API. Which one is available decides the whole design, and it is the first thing worth finding out. Where an agent needs to read the system itself rather than a copy of it, an MCP server wraps that same API read-only.

SQL and Power Query’s M language do the joins, the type coercion and the de-duplication; DAX once there is a model behind it rather than a sheet, which is Power Pivot and Power BI; Python when the shaping is past what a query can sensibly express. Most of the hidden work is here, because two systems almost never agree on what a date, a cost centre or an employee number looks like.

Python for most of it. VBA or Office Scripts when the work genuinely has to stay in the workbook a team already uses and will keep using. Keystroke and screen automation — AutoHotkey and the like — only for a system that offers no other door, because it is the one approach that breaks the day somebody moves a button.

An automation nobody can check is a second system of record with no owner. So it is scheduled rather than remembered, logged so there is a record of what it ran on and what it produced, and reconciled afterwards by an exception report that names the rule behind a difference rather than only reporting that one exists.