Claude for Payroll Prep: What to Check Before You Approve a Pay Run

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Claude for Payroll Prep: What to Check Before You Approve a Pay Run.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Claude for Payroll Prep: What to Check Before You Approve a Pay Run.

Claude can prepare a pay run by totalling hours from timesheets, comparing them with rotas and the last period, and flagging missing punches, overtime spikes, rate changes, starters and leavers. Before approving, check every flag, headcount, rates, leave, deductions, the cash to cover it and bank-detail changes; your payroll provider calculates tax and submits.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

One setting decides whether "prep only" stays true. Gusto's connector in Claude's directory lists a tool called run_payroll and says eligible customers can run payroll in Claude, and Gusto's own developer documentation tells users never to enable automatic tool execution and to confirm every tool call. The Claude for Small Business plugin's payroll skills stage a run and leave the submit button to you, but an ordinary chat with the connector switched on has no such rule built in. Block that tool before your first payroll session, and the rest of this checklist is about getting the numbers right.

What does Claude actually do when it prepares a pay run?

Claude prepares a pay run by turning timesheets into totals per person, comparing them with the rota and previous periods, and listing every anomaly with the person, date, both numbers and likely cost. You resolve each one.

The plugin's published payroll-prep skill is the clearest description of the job, and its rules are worth copying even in a plain chat:

  • Inputs: timesheets from Gusto, QuickBooks Payroll or an uploaded spreadsheet, including a photographed paper timesheet. Pay rates come from the payroll system, never from the timesheet.
  • Totals: regular hours, overtime, paid leave and unpaid time per person, then for the whole run.
  • Flags, never fixes: a missing clock-out is reported as unknown hours. The skill never fills in a likely finishing time, because a guessed time produces a payslip that looks right and is wrong.
  • Comparison: gross pay against the previous period, with any move over 10% explained.
  • Approval: one decision per flag, then a summary of headcount, hours, gross, cash leaving the account, pay date and any flags left open, before anything is staged.
  • Submission: with Gusto connected, approved hours are written into an unprocessed payroll and you submit it in Gusto; with QuickBooks Payroll or no connected system, you get a checked run sheet to key in yourself.

What Claude does not do, and should not be asked to do, is calculate tax, statutory deductions or employer costs. Those come from your payroll provider, which is built and updated for the rules where you employ people. Claude checks the inputs; the provider does the payroll.

Anthropic's usage policy draws the same line from the other side. It lists setting or changing workers' hours or pay among its high-risk areas, where a qualified person must review any AI recommendation before it is acted on, while carrying out a decision a person has already made, or applying a fixed rule, is excluded. Applying the rates and overtime rule you set is the second kind; asking Claude what someone should be paid is the first.

Which Claude setup fits your payroll: a chat, a Project or the plugin?

A Claude Project with a weekly timesheet upload suits most small teams, while the Small Business plugin's payroll workflow suits owners on Gusto or QuickBooks Payroll who want a cash check first. Its planning command can also chase overdue invoices when cash is short.

SetupPlan (Claude pricing page, 22 November 2026)What you getWho submits
Project plus timesheet uploadFree works (up to five projects); Pro $20 a month or $17 billed annuallyTotals and flags from your own rules; a run sheet you checkYou, keying hours into your provider
Plugin payroll-prep with GustoPro, Max, Ultra, Team or Enterprise; Gusto connection needs a full admin accountHours, leave and rates pulled from Gusto; anomaly flags; approved hours staged in the unprocessed payrollYou, in Gusto
Plugin with QuickBooks PayrollAs aboveRoster, rates and last run read from QuickBooks; a validated run sheetYou, keying it into QuickBooks Payroll
Plugin /plan-payrollAs above, plus a connected ledgerCash forecast for the pay date, reminders for overdue invoices if cash is short, then the payroll-prep stepsYou

Uploading an XLSX timesheet needs Claude's code execution and file creation setting switched on; CSV works without it. Whichever route you take, strip columns Claude does not need before uploading: tax ID numbers, dates of birth, home addresses and bank details. The plugin's skills are written never to repeat those fields in their output, but the safest field is the one never uploaded.

If tidying the timesheet export into one row per person per day is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

The 18-point checklist before you approve a pay run

Every item has a reason and a way to verify it. Claude can surface most of them; deciding is yours. The thresholds in group B come from the plugin's published anomaly rules, and they make sensible defaults for a plain Project too.

A. The period and the people

#CheckWhy it mattersHow to verify
1Pay period start, end and pay dateA period straddling a month end is the easiest way to pay a week twice or skip oneCompare with your provider's pay schedule before uploading anything
2Starters in this periodA new person with no hire record may be a mistake, or worse, a fakeSigned contract and new-starter form on file for every new name
3Leavers and people on leaveSomeone paid last period and missing now needs a reason; a leaver still listed gets paidAsk Claude to list everyone on last run but not this one, and the reverse

B. Hours (Claude flags, you decide)

#CheckDefault flag thresholdHow to verify
4Missing or broken punchesAny clock-in without clock-out, or the reverseAsk the employee and supervisor; never accept an assumed time
5Overtime against the person's own patternAbove 125% of their average over the last four periods, or any overtime for someone who normally has noneJob sheets or the supervisor's sign-off
6Hours above the rotaMore than 10% over the scheduled hoursRota versus timesheet, day by day
7Very long shiftsAny single shift over 12 hours; over 16 hours in a dayUsually a missed clock-out; confirm with the person
8Seven days in a rowAny run of seven consecutive working daysFatigue and premium-pay rules; check your policy
9Duplicates and overlapsTwo entries covering the same timeKeep one; never let them be merged silently
10Zero hours for a regular workerNothing logged for someone who normally worksMissing timesheet, sickness or an unrecorded leaver

C. Rates and pay setup

#CheckWhy it mattersHow to verify
11Rate changes since the last runAn unexplained rise of $2.50 an hour is $100 a week on a 40-hour weekSigned pay-change letter with an effective date
12Classification changesHourly to salaried, or employee to contractor, changes tax and rightsYour payroll provider or an employment adviser, not Claude
13Salaried people with logged hoursCan signal an unpaid partial week, or a missed starter dateContract start date and leave records

D. Leave, extras and deductions

#CheckWhy it mattersHow to verify
14Paid leave against the balanceLeave beyond the balance is either unpaid or an errorUse the balance your payroll provider shows, never an estimate
15Bonuses, commission and expense repaymentsEasy to double-pay when they arrive by email and in a spreadsheetWritten approval for each, matched to one line
16One-off deductionsRecovering an overpayment or a loan usually needs the employee's agreementSigned agreement on file; ask an adviser if unsure

E. Cash and security

#CheckWhy it mattersHow to verify
17Gross pay against last period, and cash on the pay dateA jump with no explanation is the best single warning; a run that bounces is worseEvery move over 10% explained; cash forecast covers net pay plus employer costs from your provider
18Bank-detail changes since the last runFake "please update my bank details" emails are a common payroll fraudPhone the employee on the number you already hold; never act on an email alone

One week's pay run for an electrical contractor with 12 staff

Take an illustrative electrical contractor paying 12 people weekly: eight electricians, three apprentices and an office manager on salary. Timesheets come from a phone app as a CSV every Monday morning. The owner set up a Claude Project with the rate list, the rota pattern and the checklist above as project instructions.

For the week in question Claude reported 512 hours, 46.5 of them overtime, and a gross of $19,860 against $17,940 the week before, a rise of 10.7%. It raised six flags:

  1. Apprentice 2, Thursday: clocked in at 7:02, no clock-out. Hours unknown. The apprentice confirmed he left at 15:30 with the van; 8.5 hours entered by the owner, not by Claude.
  2. Electrician 5: 14.5 overtime hours against a four-week average of 3. Real: an emergency call-out on a commercial job. Approved, and noted against the job for invoicing.
  3. Electrician 7: rate shows $32.00 against $29.50 last week. The raise was agreed, but from the following Monday. Corrected to $29.50 for this run, saving $100 that would otherwise have been overpaid.
  4. Electrician 3: left two weeks ago but still appeared with zero hours. Removed from the run before the provider generated a nil payslip.
  5. A new name with 16 hours: a labourer started on Wednesday. No new-starter form on file; the owner collected it before approving.
  6. Bank details: an email from an address one letter off Electrician 6's asked for pay to go to a new account. The owner rang the electrician, who had sent nothing. The change was never made.

The 10.7% rise was fully explained by the call-out overtime and the new labourer. Cash was the last check: the run needed about $23,000 leaving the account on Friday, including employer costs from the provider's preview, and the bank held $31,000, with $9,800 of supplier payments due and a confirmed $4,200 customer payment arriving before Friday. Tight but covered, as the cash check further down shows line by line. The owner keyed the corrected hours into the payroll provider, previewed the run there, and submitted it. Total time from upload to submission: 35 minutes, of which 20 went on flags 1, 5 and 6.

Why does Claude need four past runs before the first real one?

Claude needs past runs because most of its hours checks compare this week with something: each person's usual overtime, the rota and the previous run. On a first run there is nothing to compare against, and the plugin's anomaly rules say so plainly: most hours checks stay silent until a baseline, a rota or a stated policy exists. Missing punches still show up, and the rule for any single shift over 12 hours exists precisely because it needs no history.

So before you rely on the flags, upload the last four approved runs (or their timesheet exports) to the project files and tell Claude they are the baseline. For the contractor above, four past weeks gave each electrician an average overtime figure, which is what made Electrician 5's 14.5 hours stand out. Without them, the same week would have produced one flag instead of six. Refresh the baseline monthly by dropping the oldest run and adding the newest, so a busy season does not become the new normal without you noticing.

How do you check the cash before approving a pay run?

Check payroll cash by taking the bank balance, subtracting everything due out before the pay date, adding only certain receipts, and comparing the result with the cash the run needs. Take that last figure, including employer costs, from your provider's preview.

Here is the contractor's cash check for the week above, the kind of table Claude can build from a bank export and your list of bills (illustrative figures):

LineAmountNote
Cash in the bank, Tuesday morning$31,000From the bank export
Supplier payments due before Friday-$9,800Wholesaler account and van lease
Receipts certain before Friday+$4,200One customer's payment already confirmed by email; the other $6,100 owed is left out
Cash available on pay date$25,400
Cash the run needs-$23,000Net pay plus employer costs, from the provider's preview
Headroom$2,400Tight: no other spending until Friday

The rule that matters is the third line. Money a customer has promised is not money you have, and the plugin's payroll planning chain makes the same point: chasing an invoice that will not be paid for six weeks does nothing for this Friday's run. If headroom is negative, decide before approval whether to move a supplier payment, chase specific invoices or draw on a facility, rather than discovering it when the bank declines the payment.

Does the checklist change for a salaried office team?

The checklist gets shorter for a salaried team paid monthly, because missing punches and long shifts disappear while part-month starters, salary changes and bonuses become the main risks. An illustrative accountancy practice with nine salaried staff uses this version:

CheckHourly crew (weekly)Salaried office (monthly)
Missing punches, long shiftsEvery runNot applicable
Starters and leavers part-way through the monthOccasionalThe main risk: part-month pay must match the contract dates
Salary changesRareUsually all at once after reviews; check each against the signed letter
Bonuses and overtime claimsOvertime most weeksA few bonuses at year end; each needs written approval
LeaveBalances checked each runUnpaid leave and long sickness are the ones to catch
Bank-detail changesEvery runEvery run

For the practice, Claude's most useful job was comparing this month's run with last month's line by line and explaining every difference: two pay rises from the annual review, one starter on the 18th paid for 13 of 30 days, and a bonus paid twice because it had been added in the provider and in the spreadsheet. That last one, $1,500, was caught before approval.

A prompt for payroll prep, with a sample answer

The prompt below encodes the "flag, never fix" rule and the thresholds. The sample output shows the format worth insisting on: one block per person, both numbers, the cost where it can be worked out honestly.

PROMPT (timesheet CSV attached)
Pay period: Mon 10 Nov to Sun 16 Nov. Pay date: Fri 21 Nov.
Use the rate list and rota in the project files; never take a rate
from the timesheet. Total regular and overtime hours per person.
Flag, do not fix:
- missing punches (hours unknown, never assumed)
- overtime above 125% of the person's last-4-week average
- hours more than 10% above the rota; any shift over 12 hours
- rate changes, starters, leavers, anyone missing since last week
Group flags by person. For each: date, both numbers, cost if known.
Then totals and the % change in gross against last week's run.

SAMPLE OUTPUT (illustrative)
Electrician 5
- Overtime 14.5h vs 4-week average 3.0h (+11.5h, about $500 extra
  at time and a half on $29.00). Real, or a punch problem?
Apprentice 2
- Thu 13 Nov: in 07:02, no clock-out. Hours unknown, cost unknown.
Electrician 7
- Rate $32.00 this week vs $29.50 last week. Effective date?
TOTALS: 512.0h (46.5h overtime). Gross $19,860 vs $17,940 (+10.7%).
Not calculated: tax, deductions, employer costs (from your provider).

Two things I would check in any answer like this. First, the overtime premium: if your overtime rule is not time and a half, or applies daily rather than weekly, say so in the project instructions, because Claude will otherwise assume one rule and should tell you which. Second, the last line. If Claude ever produces tax or deduction figures, ignore them and delete the line; mixing estimated tax into a prep sheet invites someone to key it in.

How do you stop Claude from ever submitting payroll?

Claude cannot submit payroll if no connected tool lets it, so block the payroll-submission tools, run payroll sessions in Manual mode and keep the submit step inside your provider. Each layer covers a gap in the others.

  • Block the run tool. Gusto's connector lists run_payroll, described in Gusto's documentation as calculating and submitting an unprocessed payroll. On a Team plan, an owner can go to Organization settings, then Connectors, open Gusto's Tools tab and set that tool to Blocked for everyone. Anthropic's help pages describe the same Blocked setting for individual tools, which denies the tool in every mode.
  • Watch the quieter writes. Intuit's QuickBooks connector can add an employee or set base pay, and Gusto's can record time. Set those to "Always ask" so nothing changes a rate or a punch without your say-so.
  • Use Manual mode for payroll. Anthropic is rolling out Auto as the default mode on Pro, Max, Ultra and Team plans, and in Auto, tools set to "Needs approval" can run without asking. Switch to Manual from the mode selector before a payroll session; Gusto's own documentation says you should be prompted to confirm every tool call.
  • Keep payroll out of schedules. A weekly reminder to upload timesheets is fine; a scheduled task with payroll tools enabled is not.

Connecting Gusto at all needs a Gusto account with full admin permissions, according to Gusto's documentation, which is one more reason to keep that connection on the owner's login only. For comparing what the payroll platforms themselves offer, Gusto, Rippling and Deel compared on real AI for payroll is the companion piece.

If you want help fitting payroll checks into how your business runs, I offer one-to-one AI consultations for small businesses.

What should the sign-off record say?

A payroll sign-off record should state six facts before you approve: people paid, total hours, total gross, cash leaving the account, the pay date and how many flags were left open. Add who approved it and when.

The plugin asks for exactly these before it stages anything, and the habit is worth keeping in a plain Project. Here is the contractor's record for the week above, saved in the project files so next week's run can compare against it:

ItemWeek ending 16 November (illustrative)
People paid12 (1 starter, 1 leaver removed)
Total hours512.0, of which 46.5 overtime
Total gross$19,860 (up 10.7%: call-out overtime and new starter)
Cash leaving the accountAbout $23,000 including employer costs, per the provider's preview
Pay dateFriday 21 November
Flags left open0 of 6
Approved byOwner, Tuesday 18 November, 10:40
Submitted inThe payroll provider, by the owner, after its own preview

One flag left open is sometimes fine: an employee on holiday who cannot confirm a punch until Monday can be paid the rota hours and corrected next run, if your policy allows. The record makes that a decision rather than an oversight.

Where payroll prep with Claude goes wrong

These mistakes come up repeatedly in small firms, and each looks harmless until payday.

"Looks good?" for all six flags at once

The owner reads the summary, types "all fine" and approves. One of the six was a missed clock-out that Claude had correctly left as unknown, and the provider paid zero hours for that day. Resolve flags one person at a time; the plugin refuses to batch them for this reason.

A rate taken from the timesheet

A supervisor typed rates into the timesheet app, and one was out of date. Tell Claude rates come only from the rate list or the payroll system. In QuickBooks Payroll, one person can carry many pay rates with most at zero hours; the plugin's notes say to total from the rate with hours against it, not the first one listed.

Leave estimated instead of read

Claude worked out a holiday balance from an accrual rule and approved two days that were not there. Leave balances come from your payroll provider, never from a calculation in a chat.

The period crossing a month end

A run for the week of 27 October to 2 November was compared with October's monthly report and looked short. Always give Claude the exact period dates and compare like with like: last week's run, not last month's totals.

Cash checked after approval

The run was approved on Tuesday and the cash checked on Thursday, after a large supplier payment had gone out. Put the cash check before approval. If cash is short, drafting and chasing invoices with Claude is the fastest lever, and a rolling cash budget built from last year's actuals shows the squeeze weeks before payday.

The underlying timesheet problems, such as people forgetting to clock out or supervisors approving late, are covered in catching timesheet errors before payday. Fixing those at source shortens every future check.

Questions owners ask about payroll prep with Claude

Can Claude run payroll for me?

It can on some setups, which is why you should stop it. Gusto's Claude connector lists a run_payroll tool and says eligible customers can run payroll in Claude, while the Small Business plugin's payroll skills stage the run and never press that button. Block the run tool in your connector settings, keep payroll sessions in Manual mode, and submit the run yourself in your payroll provider.

Does Claude calculate tax and deductions on a pay run?

It should not be your source for them. Claude can total hours and gross pay from the rates you give it, but tax, statutory deductions and employer costs come from your payroll provider, which is built and updated for your jurisdiction. Use Claude to check inputs and explain changes, then let the provider calculate and submit.

Which payroll systems work with the Claude for Small Business plugin?

Gusto and QuickBooks Payroll are the main ones, and Claude Academy also lists MYOB, NetSuite, Xero, Zoho Books and Notion as tools the payroll-prep skill can draw on. With Gusto, the plugin stages the approved hours in an unprocessed payroll for you to submit. With QuickBooks Payroll, or no connected system, you get a checked run sheet to key in yourself.

Is it safe to upload timesheets and payroll data to Claude?

It can be, if you strip what Claude does not need. Remove tax ID numbers, dates of birth, home addresses and bank details before uploading; hours, names and rates are enough for the checks. Team and Enterprise plans do not train on your content by default; on Free, Pro and Max, check the model-improvement setting under Settings, Privacy.

How long should payroll prep with Claude take each week?

For a team of ten to fifteen hourly staff, expect about 20 to 30 minutes once the routine is set up: a few minutes to export and upload, a few for Claude's run sheet, and the rest resolving flags one person at a time. If it takes longer, the timesheets themselves usually need fixing at source.

Further reads

Show 4 more related tutorials

Sources: Claude for Small Business plugin page, Claude Academy plugin tutorial and Anthropic's launch post; the plugin's published payroll-prep (with its anomaly rules) and plan-payroll skill files and README; the Gusto and Intuit QuickBooks listings in Claude's connector directory; Gusto's MCP developer documentation; Claude help centre articles on Cowork permissions, connectors and uploading files; Claude pricing page and privacy centre (all read 22 November 2026).

Ready to put AI on the jobs that eat your week?

No data team or big budget needed. In a 1:1 consultation I show small and medium-sized businesses and self-employed people which AI tools they can use, one workflow at a time.

Book a 1:1 call with me