HR Automation for Small Business: Start With the Record

HR automation for a small business starts with one employee record that everything else reads from. It does not start with an onboarding journey. Get the record right and attendance, leave, pay basis and the monthly report stop being retyped in three places. Get it wrong and whatever software you buy on top becomes the fourth place to retype it.

We run this across our own group. The useful part of the story is not the demo, it is which pieces we automated and which two we deliberately have not.

Why a small business never has enough HR to do this by hand

The median organisation has roughly two HR staff per hundred employees, which means a 40-person business has well under one. SHRM’s 2025 CHRO benchmarking survey of 2,371 members, fielded January to March 2025 and self-reported, put the median HR-to-employee ratio at 1.98.

That is an all-sizes median, not a small-business figure, and the smaller you are the more it understates the problem. The admin does not scale down as fast as the headcount does. The same SHRM data has HR expense at 2.4% of operating expense, double the 2017 ratio, with median HR budgets up 9.1% on 2024. The cost per head is climbing while the person carrying it is still part-time at the job.

Small businesses are about a tenth as likely to have automated any of it

ADP’s 2025 HR trends survey, published in its HR Trends and Priorities for 2026 guide, reports agentic AI in use at 4% of small businesses, 25% of midsized and 48% of large ones. Governance follows the same shape: 20% of small businesses have a generative AI governance process in place, against half of midsized and two thirds of large.

ADP does not publish the sample size or the fielding method for that study, and it sells the software category in question. Read those as ADP’s own attributed figures rather than neutral market data.

Sage’s HR and Payroll Leaders’ Report, published 10 March 2026 from a survey of 1,000 HR and payroll leaders at small and mid-sized businesses in the UK, Ireland and South Africa, found 71% saying their workload rose over the past year and 86% expecting AI to change the function. Self-reported, and Sage sells payroll software too.

One number we are not going to give you: hours per week a small business owner spends on HR admin. There is no defensible 2025 or 2026 source for it. What circulates is either unattributed or traces back to pre-2020 vendor research, so we left the hole open rather than filling it with something that sounds authoritative.

The record is the part the law already cares about

In the United States, the Fair Labor Standards Act requires an employer to keep 14 specified items for every non-exempt employee, including hours worked each day, total weekly hours, the basis of pay, the regular hourly rate, weekly overtime earnings, and every addition to or deduction from wages. That list is in the US Department of Labor’s Fact Sheet #21.

Payroll records are kept at least three years, and the records the wage computations rest on, such as time cards and work schedules, at least two. State law frequently requires longer, and this is US federal law only, so check your own jurisdiction before copying the retention periods.

Getting it wrong has a measured price. The Department of Labor’s Wage and Hour Division recovered more than $259 million in back wages for 176,957 employees in fiscal year 2025, roughly $1,465 per worker. That figure counts enforcement outcomes, so treat it as a floor rather than an estimate of how common errors are.

Fourteen mandated fields per employee is the argument for automation on its own. A spreadsheet can hold them. What a spreadsheet cannot do is tell you who changed one, when, and what it said before.

What we run across our own group

What we built is narrow on purpose. One record per employee: employee number, name, job title, section, identity and statutory fund numbers, join and resignation dates, pay basis, and the business unit they sit in. Several business units run inside the same system with per-unit reporting. There is no document vault in it and no career history, by choice.

Attendance is entered and corrected in a grid by an HR user, and every edit is attributed to the account that made it with a timestamp. A separate access log records logins and actions with the user, the action, the time, the IP address and the device.

Reports generate daily and monthly, there is a CSV export, and a daily summary posts itself into a chat channel broken out per business unit, so nobody has to open the tool to see the state of the day. Accounts are role-gated, with an admin role that manages users and an operator role that does the work, and sign-in runs through single sign-on.

A tomte, our word for one production agent with one defined job, is what pushes that daily summary. It decides nothing. It assembles what the records already say and puts it where people already are.

The two things we did not build

Onboarding does not exist as a flow anywhere in our stack. Our pre-hire recruitment tool drafts and sends an offer with the position, salary and start date filled in, and that same tool’s own onboarding and reporting screens are still unbuilt placeholders. The post-hire steps in our group are mostly one-off document collection and access grants, and we would rather run those from a checklist a person owns than ship a half-built journey that pretends to track them.

Employment letters and certificates for current staff are not generated either. Salary confirmations and employment letters are written by a person, every time. It is the most requested item on our own list and the most likely next build, and it is not shipped today.

Leave sits in a separate application. Staff log in themselves to request leave, check balances and view their own basic employment details, and the attendance system reads approved leave back and overlays it on the calendar. The join between the two runs on employee name. It works, and it is also the weakest link in the whole arrangement: a name-matched join holds for two years and then breaks the week somebody is entered with a middle initial. We know. It is on the list with the letters.

The attendance and leave half of this is written up in attendance management automation, and the pre-hire side, including the line a human has to hold on screening, is in AI recruitment screening automation.

What actually gets automated

Job	On spreadsheets and chat	In our group today
Employee record	A file per unit, versions by email	One record per person, scoped to a business unit
Correcting attendance	Overwritten, no trace of who	Every edit attributed to an account and timestamped
Who saw or changed what	Unknown	Login and action log with IP and device
Monthly report	Rebuilt by hand	Generates from the records, plus CSV export
Daily visibility	Somebody asks somebody	Summary posts itself, per business unit
Leave requests	A message thread	Staff self-service app, read back into the calendar
Onboarding a joiner	A checklist in someone’s head	Not built. Still a checklist, deliberately
Employment letters	Typed from scratch each time	Not built. Written by a person

If you only automate one thing, automate the written-down part

Rule-governed work automates well: a record with defined fields, a report with a defined shape, a retention period with a number attached to it. Judgment work does not. Whether a borderline absence counts as authorised, whether a resignation date is negotiable, what a performance conversation should conclude: those stay with a named person.

Automate the record and leave the judgment alone, and the person carrying HR in a 40-person company stops spending the week as a typist. That is the whole return. Anyone promising you a transformed function from a 4%-adoption baseline is selling ahead of the evidence, including the vendors whose surveys we quoted above.

What this would cost as a Sprint

Scoped the way we built it, an employee record with per-unit reporting, attributed edits, an access log, daily and monthly reports and a summary that posts itself, is one AI Workforce Sprint, from $9,500 fixed, delivered in four to six weeks, shipped into your own GitHub, database and hosting accounts. If you do not yet know how many spreadsheets and chat threads your HR admin currently spans, that inventory is what the AI Operations Audit produces first.

The logging and ownership discipline every one of these builds inherits, with the run and failure counts behind it, is on our agent fleet proof page.

If you want this running on your own HR admin, write down how many business units it has to cover, which employee records your jurisdiction requires you to hold, and who should receive the daily summary. Start async. You get a written reply within one business day, and there is no call to book.