Payroll Calculation AI in Three Layers — What Needs Cutting Is Not the Arithmetic

Payroll Calculation AI in Three Layers — What Needs Cutting Is Not the Arithmetic

Enquiries about bringing AI into payroll are becoming more frequent, but when you ask what is actually being expected, most of them come down to a wish to automate multiplying an hourly rate by hours worked. Off-the-shelf payroll software solved that arithmetic long ago. Where payroll calculation AI genuinely removes hours is not the calculation itself, but the collection of variable information that comes before it and the checking and correction that remains after it. This article evaluates investment in those two places, layer by layer, using an in-house model calculation built around a Japanese-owned factory in Thailand.

The opposite side of the same data — employees looking up their own leave balances and payslips — was covered in Automating Internal HR Inquiries. This article turns that around and looks at the work a payroll administrator does to arrive at a confirmed payment figure each month. Even where the same HR data is involved, a mechanism for looking things up and a mechanism for calculating them call for investment in entirely different places.

Where the Hours Actually Sit — Arithmetic Is Already a Solved Problem

The word payroll conjures up a chain of arithmetic. Multiply base pay by days worked, multiply an overtime rate by overtime hours, deduct social security contributions. That part, however, is very largely automated the moment a commercial payroll package is installed. Companies where a person still works out payment amounts on a calculator are already a minority even among Japanese-owned manufacturing sites in Thailand.

So why does the burden of closing payroll refuse to go down. In a survey of the realities of payroll processing that LayerX Inc. conducted on 12 and 13 February 2026, 67.4% of 500 payroll administrators at companies described the payroll close as a heavy burden. The breakdown is 20.4% answering “a very heavy burden” and 47.0% answering “a somewhat heavy burden”. Nearly seven in ten administrators report a heavy load in work that software is supposed to be doing for them.

The same survey also measured how far each process depends on particular individuals. Collecting and checking attendance information came in at 82.0%, collecting and checking variable information at 77.6%, calculating payments and deductions at 72.8%, and post-calculation follow-up at 76.0%, while only 33.8% of companies said the work was standardised. The point worth noticing is that calculating payments and deductions — the arithmetic itself — has the lowest individual dependency of the four at 72.8%. In other words, the calculation step is the one furthest along toward standardisation, and what sits on either side of it is handled far more personally.

There is one more decisive figure. According to the same survey, 71.6% of administrators spend at least half of their total working hours on checking, correcting and confirming after the calculation has been run. The largest single block of payroll work is not the calculation but the verification and correction that follow it.

Once that structure is clear, the targets for payroll automation narrow on their own. Automating the arithmetic has already been done by commercial software, so investing there produces nothing new. If an effect is to be had, it lies before the arithmetic and after it.

A Three-Layer Model for Payroll Automation — Layers B, C and D

To break payroll into units you can actually make investment decisions about, we split the process into layers. In this article the arithmetic itself is called Layer A and is excluded from the calculation, because there is no reason to invest again in territory that existing payroll software has already settled.

These three layers are a classification of function and, at the same time, a classification of cost. The model calculations that follow are built on the assumption that Layer C stacks on top of Layer B and that direct investment in Layer D stacks on top of both. The measurement thinking behind matching cost to effect layer by layer is set out in Measuring the Effect of AI Adoption and Designing ROI, which is worth reading alongside this article if you intend to perform the same decomposition on your own operation.

Layer B — Why Collecting and Reconciling Variable Information Is Heavy

Layer B is everything that happens before figures are entered into the payroll software. Gathering the information that changes every month, reconciling it and confirming it all belongs here.

The diagram above shows information from separate input sources — clock-in records from attendance terminals, shift rosters, leave application forms — converging on a single payroll screen. Layer B is heavy precisely because there is only one destination while there are several sources, each with its own format and its own update timing.

At a manufacturing site in Thailand, that convergence gets more complicated still, because overtime premiums differ by the type of working time. Under Thailand’s Labor Protection Act, overtime on a normal working day is paid at 1.5 times the ordinary hourly rate. For work on a holiday, monthly-paid employees are treated as already having holiday pay included in their monthly salary, so hours within normal working time are paid at 1 time, while daily-paid employees receive 2 times, and overtime worked on a holiday is paid at 3 times. The same two hours of overtime carry a different rate depending on whether they fall on a weekday evening, inside normal hours on a holiday, or beyond normal hours on a holiday. Totalling clock data as it stands does not produce correct overtime pay; a step that classifies the type of time is always required. If you are starting on overtime pay automation, the entry point is not the part that produces an amount but confirming whether your attendance system can output data split into weekday overtime, holiday normal hours and holiday overtime.

Judging allowances is also Layer B work. Pay packages for factory workers in Thailand add a commuting allowance to base pay, and because public transport is thin outside the central urban areas, a fixed monthly figure of around 5,000 baht is common. Housing allowances run from 5,000 baht to 10,000 baht, mainly for supervisory grades, and alongside them sit perfect-attendance allowances and meal or living-cost allowances for workers with heavy overtime or shift patterns. Of these, the perfect-attendance allowance carries the purpose of discouraging abuse of sick leave, so whether it is paid changes from month to month according to absences and late arrivals. Allowances are therefore not fixed amounts to be added up but items to be judged against each month’s attendance record.

The LayerX survey put collecting and checking attendance information at 82.0% and collecting and checking variable information at 77.6% — the two highest individual dependencies of the four processes — and the sheer number of these judgements is a plausible reason why. The rules for the judgement live in the administrator’s head and have never been written down.

Layer C — Compliance Checking Is Reconciliation Against the Rules

Layer C is the process of confirming that the results are consistent with legislation and internal rules. It is the core of what gets called payroll compliance. If Layer B is the work of gathering the figures, Layer C is the work of seeing whether the figures gathered fall inside permitted limits.

At a site in Thailand there are three representative checks in this layer.

• Weekly overtime ceiling. Under Section 26 of Thailand’s Labor Protection Act and the related ministerial regulations, overtime and holiday work combined are capped at 36 hours per week. Aggregate the figures only at month end and you discover after the fact that some week went over
• SSO (social security) contribution ceiling. 2026 is the year the rules change
• Provincial minimum wages. Minimum wages in Thailand differ by province, so you have to confirm that pay does not fall below the level for the province your site sits in

Of these, the SSO contribution ceiling is the item demanding attention in 2026. The Royal Gazette notification dated 12 December 2025 announced that the ceiling on monthly wages used as the basis for calculating Thai social security contributions will be raised in stages from 1 January 2026. In the first stage, running from January 2026 to December 2028, the ceiling rises from 15,000 baht to 17,500 baht and the monthly contribution ceiling rises from 750 baht to 875 baht for each of employer and employee. The contribution rate stays unchanged at 5% for each side. The second stage, from 2029 to 2031, sets a ceiling of 20,000 baht with a contribution ceiling of 1,000 baht, and the third stage, from 2032 onward, a ceiling of 23,000 baht with a contribution ceiling of 1,150 baht.

The problem arises where calculation logic or management ledgers built around the old ceiling of 15,000 baht are still sitting inside the company. If the ceiling alone is left at the old value, contributions are understated for every employee whose wage exceeds it. A year in which the rules change is also a year in which hard-coded values like this are easy to miss.

What the diagram above shows is how a Layer C check gate behaves. Processing that clears the pre-calculation gate carries straight on, while processing that turns out to be inconsistent with the rules is stopped at the gate and sent back. The important point is that this send-back happens before the calculation. Without the gate, the same inconsistency would be discovered after the calculation had finished and a payment figure had been produced. Even for an identical problem, changing where it is found changes how much work arises downstream.

Layer D — Verification, Correction and Exceptions Are the Biggest Block

Layer D is the post-calculation process. Verifying the results, responding to queries raised by employees, and recalculating where necessary all fall here.

Given that the LayerX survey put individual dependency in post-calculation follow-up at 76.0%, and that 71.6% of administrators spend at least half their total hours on checking, correcting and confirming after the calculation is run, there is no arguing with the conclusion that Layer D is the biggest bottleneck in payroll work.

The intuitive response at this point is to say, right, let us automate Layer D. Attacking the process that consumes the most hours is a natural instinct. In this article’s model calculation, however, that investment does not pay back. The numbers come later, but the logic can be stated up front. Most of what happens in Layer D is not the kind of work where applying a rule mechanically produces the answer. Reaching agreement with an employee, matching supporting documents, confirming individual circumstances — cases that require judgement remain.

One caveat. The view taken in this article, that most of the verification and correction arising in Layer D traces back to the rule inconsistencies Layer C detects, is not a sourced statistic. It is an interpretation derived logically from the individual dependency figures and the size of the checking and correction time reported in the LayerX survey. Treat it as an interpretation rather than an established fact, and test it against what actually happens at your own site.

A Payroll Error Does Not End With Lost Time

Assess investment in Layer C by looking only at hours removed and you will undervalue it, because errors in payroll carry legal consequences as well as rework for the administrator.

Section 9 of Thailand’s Labor Protection Act provides that where an employer fails to pay wages without reasonable cause, interest accrues on the unpaid amount at 15% per year. Furthermore, from the day seven days have passed since the due date, and every seven days thereafter, an additional 15% per year must be paid on top of that default interest. Separately, Section 144 of the same Act provides that a breach of the wage-payment provisions, including Section 70, carries imprisonment of up to six months, a fine of up to 100,000 baht, or both.

This article does not attempt to estimate the probability or the amount of such a breach occurring at the model factory, because there is no sourced data on incidence. What matters here is that investment in preventing payroll errors is not something to be judged solely on the benefit of reducing an administrator’s overtime. The existence of consequences like these in the legal framework is what justifies underwriting rule-based checking of the kind Layer C provides. In the model calculations that follow, this reduction in legal risk is not converted into money; payback periods are computed from hours removed alone. Read the resulting figures, in other words, as a conservative view that excludes the risk-reduction component.

In-House Model Calculation — Current Hours at a Japanese-Owned Electronic Components Plant in Thailand

From here on, the numbers. What follows is a model that TOMAS TECH assembled for illustration, not measured results from any particular company. Change the assumptions and the conclusions change with them, so read it with your own figures substituted in.

The 200 baht hourly cost is set as the fully loaded hourly rate for a payroll administrator, an HR department staff-grade role.

Next, the hours currently going into Layers B, C and D. The formulas are “monthly cases multiplied by minutes per case, divided by 60, equals monthly hours”, “monthly hours multiplied by 12 equals annual hours”, and “annual hours multiplied by 200 baht equals the annual labour cost equivalent”.

Taking Layer B as the worked example, 480 multiplied by 8 gives 3,840 minutes, 3,840 divided by 60 gives 64.0 hours per month, 64.0 multiplied by 12 gives 768.0 hours per year, and 768.0 multiplied by 200 gives 153,600 baht. For Layer C, 220 multiplied by 6 gives 1,320 minutes, 1,320 divided by 60 gives 22.0 hours per month, 22.0 multiplied by 12 gives 264.0 hours per year, and 264.0 multiplied by 200 gives 52,800 baht. For Layer D, 160 multiplied by 22.5 gives 3,600 minutes, 3,600 divided by 60 gives 60.0 hours per month, 60.0 multiplied by 12 gives 720.0 hours per year, and 720.0 multiplied by 200 gives 144,000 baht.

The totals are 64.0 plus 22.0 plus 60.0, or 146.0 hours per month, then 146.0 multiplied by 12 for 1,752.0 hours per year, and 1,752.0 multiplied by 200 for 350,400 baht.

How large is 146.0 hours against the combined scheduled monthly hours of five payroll administrators. Five multiplied by 176 hours gives 880 hours per month, and 146.0 divided by 880 is 16.6%. Roughly one sixth of the payroll team’s capacity disappears into the processes before and after the calculation.

Note that this 16.6% is an internal estimate derived under a different survey design from the LayerX population of corporate payroll administrators, and no direct comparison with external statistics has been made. Read it strictly as an in-house estimate. The case volumes of 480, 220 and 160, and the per-case times of 8, 6 and 22.5 minutes, are likewise values set for the TOMAS TECH illustrative model, not sourced measurements.

Hours Removed, Layer by Layer — Automation Rates Differ by Layer

In working out the reduction, a different automation rate is set for each layer. Applying the same coefficient to every layer would put work of different natures on the same footing and distort the conclusion.

For Layer B, 480 multiplied by 0.70 gives 336 cases, 336 multiplied by 8 gives 2,688 minutes divided by 60 for 44.8 hours per month, 44.8 multiplied by 12 gives 537.6 hours per year, and 537.6 multiplied by 200 gives 107,520 baht.

For Layer C, 220 multiplied by 0.85 gives 187 cases, 187 multiplied by 6 gives 1,122 minutes divided by 60 for 18.7 hours per month, 18.7 multiplied by 12 gives 224.4 hours per year, and 224.4 multiplied by 200 gives 44,880 baht. The high automation rate of 85% is set for Layer C because its judgements are rule-based comparisons of a number against a threshold, leaving less room for discretion than in Layer B.

The reduction in Layer D splits into two parts of different character. The first is reduction by spillover. It arises only where Layer C has been adopted, and does not arise in a Layer B only configuration. Here, 160 multiplied by 0.35 gives 56 cases, 56 multiplied by 22.5 gives 1,260 minutes divided by 60 for 21.0 hours per month, 21.0 multiplied by 12 gives 252.0 hours per year, and 252.0 multiplied by 200 gives 50,400 baht.

The mechanism behind that spillover comes from where the gate sits, as described in the Layer C section. Verification and correction in Layer D arise mainly not from arithmetic mistakes but from inconsistencies with the rules — an overtime ceiling exceeded, an old SSO ceiling still being applied — being discovered after the calculation. If Layer C detects them beforehand, some of the cases that would have reached Layer D never arise in the first place. What is falling here is not the time each response takes but the number of responses that occur.

The second part is the additional reduction from investing directly in Layer D. After spillover, 160 minus 56 leaves 104 cases, of which the share that can be automated directly is set at 104 multiplied by 0.25, or 26 cases. Then 26 multiplied by 22.5 gives 585 minutes divided by 60 for 9.75 hours per month, 9.75 multiplied by 12 gives 117.0 hours per year, and 117.0 multiplied by 200 gives 23,400 baht. The direct automation rate is held to 25% on the assumption that most of the 104 cases left in Layer D involve individual circumstances that rules cannot be applied to mechanically, such as reaching agreement with an employee or matching supporting documents.

The automation rates of 70%, 85%, 35% and 25% are likewise settings in an illustrative model rather than measured values. The 35% spillover rate for Layer C in particular is the variable that decides this article’s conclusion, so when you examine the case for your own site we recommend sorting the past year’s correction cases into those caused by rule inconsistencies and those caused by individual circumstances, and producing a real number.

Payback by Investment Configuration — Comparing Three Options

Costs are set as increments, on the assumption that Layer C stacks on Layer B and that direct investment in Layer D stacks on both.

These upfront and running costs are also values set for the TOMAS TECH illustrative model rather than results based on external sources. They will move with how easily existing systems can be integrated and how much content has to be loaded into the rules database at the outset.

On that cost basis, here are three investment configurations compared. Annual net benefit is annual value removed minus annual running cost, and simple payback is cumulative upfront investment divided by annual net benefit.

The diagram above lays out the payback periods of the three configurations as the lengths of three tracks running from a common starting point. The top track is Configuration 1 (one module stacked), the middle is Configuration 2 (two), and the bottom is Configuration 3 (three); the middle track is the shortest and the bottom track is conspicuously the longest. The payback period does not simply lengthen as the investment grows — it shortens once in the middle before extending again, and that shape is the point of this article.

For Configuration 1, the value removed is Layer B’s 107,520 baht alone, since without Layer C there is no Layer D spillover. Then 107,520 minus 54,000 gives 53,520 baht, and 260,000 divided by 53,520 gives 4.85799…, or 4.9 years.

For Configuration 2, upfront investment is 260,000 plus 190,000, or 450,000 baht; running cost is 54,000 plus 42,000, or 96,000 baht; and value removed is Layer B 107,520 plus Layer C 44,880 plus Layer D spillover 50,400, or 202,800 baht. Then 202,800 minus 96,000 gives 106,800 baht, and 450,000 divided by 106,800 gives 4.21348…, or 4.2 years.

For Configuration 3, upfront investment is 450,000 plus 150,000, or 600,000 baht; running cost is 96,000 plus 48,000, or 144,000 baht; and value removed is 202,800 plus direct Layer D 23,400, or 226,200 baht. Then 226,200 minus 144,000 gives 82,200 baht, and 600,000 divided by 82,200 gives 7.29927…, or 7.3 years.

Here is what is happening. Configuration 1 on its own pays back in 4.9 years. Configuration 2, which adds Layer C, improves the overall payback from 4.9 years to 4.2 years through the 50,400 baht of spillover into Layer D, even though Layer C’s own direct effect is a modest 44,880 baht. Configuration 3, which invests directly in Layer D — the biggest bottleneck of the three — sees annual net benefit fall from 106,800 baht to 82,200 baht and payback deteriorate from 4.2 years to 7.3 years.

Why Direct Investment in Layer D Does Not Pay Back — Reading It as an Increment

Cumulative figures obscure the difference in character between Layer C and Layer D, so here they are again as increments. Incremental payback is incremental upfront investment divided by incremental annual net benefit.

The incremental net benefit from 1 to 2 is 106,800 minus 53,520, or 53,280 baht, and 190,000 divided by 53,280 gives 3.56607…, or 3.6 years. The incremental net benefit from 2 to 3 is 82,200 minus 106,800, or -24,600 baht, an outflow of 24,600 baht every year, so no payback period can be calculated.

The asymmetry breaks down into two reasons.

First, Layer C is a layer whose spillover effect exceeds its own direct effect. Against an upfront investment of 190,000 baht, Layer C’s own reduction is only 44,880 baht, but adding the 50,400 baht of spillover into Layer D brings the total to 95,280 baht. Put Layer C forward on its own as an overtime ceiling alerting feature and the capital request shows 44,880 baht of value removed against 42,000 baht of running cost, leaving a net benefit of just 2,880 baht. Presented that way, it will not get approved. Layer C has to be assessed as an investment that reduces the number of downstream cases, not as one that reduces the hours it touches itself.

Second, direct investment in Layer D faces a population that has already been cut back before the investment is made. After Layers B and C are in place, 104 cases reach Layer D, of which 25%, or 26 cases, can be automated directly. Even though Layer D’s 22.5 minutes per case is the longest of the three layers, the value removed stops at 23,400 baht because the volume has fallen and the automation rate is low. Against that sit additional costs of 150,000 baht upfront and 48,000 baht a year to run.

The instinct to invest in the process that consumes the most hours, in other words, overlooks the fact that those hours divide into a part that can be solved by mechanical application of the rules and a part that cannot. The former is eliminated upstream by Layer C, and the latter is not a candidate for automation at all. Eliminating the cause of verification and correction upstream is a more efficient investment than pointing AI at the verification and correction process itself — that is the conclusion this model produces.

Sensitivity Analysis — How the Conclusion Moves With Layer B’s Automation Rate

The most uncertain input above is Layer B’s automation rate of 70%. That single figure is varied across 50%, 70% and 85%. Layer C’s automation rate of 85%, the Layer D spillover rate of 35% and the Layer D direct automation rate of 25% are all held fixed, as are the 200 baht hourly cost, the case volumes and the per-case times. Because no single coefficient is applied across every effect, the reductions from Layer C and from Layer D spillover — 44,880 baht and 50,400 baht — are identical in all three cases.

In the 50% case, 480 multiplied by 0.50 gives 240 cases, 240 multiplied by 8 divided by 60 gives 32.0 hours per month, 32.0 multiplied by 12 gives 384.0 hours per year, and 384.0 multiplied by 200 gives 76,800 baht. Configuration 1’s net benefit is 76,800 minus 54,000, or 22,800 baht, and 260,000 divided by 22,800 gives 11.4035…, or 11.4 years. Configuration 2’s value removed is 76,800 plus 44,880 plus 50,400, or 172,080 baht, its net benefit is 172,080 minus 96,000, or 76,080 baht, and 450,000 divided by 76,080 gives 5.91483…, or 5.9 years.

In the 85% case, 480 multiplied by 0.85 gives 408 cases, 408 multiplied by 8 divided by 60 gives 54.4 hours per month, 54.4 multiplied by 12 gives 652.8 hours per year, and 652.8 multiplied by 200 gives 130,560 baht. Configuration 1’s net benefit is 130,560 minus 54,000, or 76,560 baht, and 260,000 divided by 76,560 gives 3.39603…, or 3.4 years. Configuration 2’s value removed is 130,560 plus 44,880 plus 50,400, or 225,840 baht, its net benefit is 225,840 minus 96,000, or 129,840 baht, and 450,000 divided by 129,840 gives 3.46580…, or 3.5 years.

What to read out of this is the difference in spread. Whether Layer B’s automation rate lands at 50% or 85% splits Configuration 1’s payback between 11.4 years and 3.4 years, a range of more than three to one. Configuration 2, by contrast, stays within 5.9 years to 3.5 years. Deploying Layer C alongside is itself a hedge against Layer B failing to bed in as expected.

Note that in the 85% case the order reverses, with Configuration 1 at 3.4 years and Configuration 2 at 3.5 years. If Layer B is going to settle at an exceptionally high automation rate, the value of adding Layer C thins out. That, however, presumes you know in advance that Layer B will reach 85%, and few companies can hold that level of confidence before starting. Where the adoption rate cannot be predicted, choosing the configuration with the narrower spread is the more rational move.

Across Multiple Sites, Configuration 2 Looks Different Again

Configuration 2’s payback of 4.2 years fails for a company that applies a three-year threshold. For a group operating three sites of comparable size, however, the conclusion changes.

From the second site onward, design assets such as allowance rules, the rules database and the API integration specifications can be reused, so upfront investment is set at 40% of the first site. Running cost, on the other hand, is set at 70% from the second site onward, because collective agreements and regional minimum wages differ from site to site. These 40% and 70% figures are assumptions too, not measurements based on a source.

Upfront investment across three sites is 450,000 plus 450,000 multiplied by 0.40 multiplied by 2, giving 450,000 plus 360,000, or 810,000 baht. Annual running cost is 96,000 plus 96,000 multiplied by 0.70 multiplied by 2, giving 96,000 plus 134,400, or 230,400 baht. Annual value removed is 202,800 multiplied by 3, or 608,400 baht. Net benefit is 608,400 minus 230,400, or 378,000 baht, and payback is 810,000 divided by 378,000, giving 2.14285…, or 2.1 years, which clears the three-year threshold.

The same Configuration 2 investment halves from 4.2 years at a single site to 2.1 years across three. Rule-based investment of the Layer C kind shows a particularly large reuse effect, because the rules themselves are common at national level. The 36-hour weekly overtime ceiling and the SSO contribution ceiling are the same at every site inside Thailand. Only provincial minimum wages vary by location, and that amounts to swapping a parameter in the same mechanism. For Japanese-owned companies with several sites in Thailand, there is real value in designing the rules database for horizontal rollout from the very start.

How to Sequence the Rollout — Do Not Skip Layer C

Given the model, the sequence works out as follows.

• Record the last three months of payroll close work by process and measure the real case volumes and handling times for Layers B, C and D
• Sort the past year’s correction cases into those caused by rule inconsistencies and those caused by individual circumstances, and reset Layer C’s spillover rate to your own real number
• Get the allowance judgement rules out of the administrator’s head and into a document — perfect-attendance eligibility, who qualifies for housing allowance, the conditions for meal and living-cost allowances
• Confirm whether the attendance system can output overtime split into weekday overtime, holiday normal hours and holiday overtime
• Take stock of whether the SSO contribution ceiling has been updated to the 2026 values of 17,500 baht and 875 baht, and whether the old values survive anywhere
• Build the Layer B integration and measure the automation rate over three months
• Put the Layer C rules database and alerting on top of Layer B, and measure the change in the number of Layer D cases
• Decide on direct investment in Layer D only after Layer C’s spillover effect has been measured

The critical items in this sequence are the second and the seventh. The investment case for Layer C depends on the spillover rate, and the 35% used in this article is only an assumption. If rule inconsistencies account for merely one tenth of your correction cases, Layer C assesses lower than it does here. If they account for more than half, Layer C’s priority rises further still. This one figure is worth confirming against your own data before you begin.

Nor is there any need to strike direct investment in Layer D off the list from the outset. What the model says is that it fails to pay back specifically when the target is the 104 cases remaining after Layers B and C are in place. Invest in Layer D alone without Layers B and C and the population stays at 160 cases, so the value removed is larger — but that configuration leaves the cause of the inconsistencies untouched and keeps catching them downstream, which means the same problem recurs every time the rules change.

Frequently Asked Questions (FAQ)

What is payroll calculation AI?

It refers to building generative AI or automation into the payroll process to reduce the administrator’s workload. The part that calculates the payment amount, however, is already handled by existing payroll software, and putting AI there produces almost no new effect. The hours that remain in practice sit in the collection and reconciliation of variable information before the calculation — confirming overtime hours, reflecting shift differences, judging allowance eligibility — and in the verification, correction and exception handling that remain afterwards. Understand the investment target implied by the term payroll calculation AI as those two places on either side of the calculation.

How much does payroll calculation AI cost?

It varies with the scope covered. In this article’s illustrative model, one Japanese-owned electronic components plant in Rayong Province, Thailand with 750 employees, Layer B covering the collection and reconciliation of variable information is set at 260,000 baht upfront and 54,000 baht per year to run; Layer C covering compliance checking adds an incremental 190,000 baht upfront and 42,000 baht per year; and direct support for Layer D adds an incremental 150,000 baht upfront and 48,000 baht per year. These are settings in an illustrative model and will move with how easily your existing attendance system can be integrated and how complex your allowance rules are. Build the numbers you take into a capital request only after measuring hours by process at your own site.

Can payroll calculation AI eliminate errors completely?

No. Even in this article’s model the automation rate is set at 85% for Layer C and 70% for Layer B, neither of them 100%. Judgements that compare a number against a threshold, such as reconciliation against the rules, mechanise easily, while cases requiring agreement with an employee or the matching of supporting documents stay with people. The goal is not to automate every case but to detect the seeds of rule breaches before the calculation and reduce the number of cases that reach post-calculation correction. Note also that Section 9 of Thailand’s Labor Protection Act provides for interest at 15% per year where wages are not paid without reasonable cause, with a further 15% per year added from the day seven days have passed since the due date and every seven days thereafter, and that Section 144 of the same Act attaches imprisonment of up to six months, a fine of up to 100,000 baht, or both, to a breach of the wage-payment provisions. The consequences of an error are not confined to lost time.

What changes should a payroll system in Thailand watch for in 2026?

The ceiling on monthly wages used as the basis for social security (SSO) contributions is being raised in stages. According to the Royal Gazette notification dated 12 December 2025, the first stage running from January 2026 to December 2028 raises the ceiling from 15,000 baht to 17,500 baht and the monthly contribution ceiling from 750 baht to 875 baht for each of employer and employee. The contribution rate stays unchanged at 5% for each side. The second stage, from 2029 to 2031, sets a ceiling of 20,000 baht with a contribution ceiling of 1,000 baht, and the third stage, from 2032 onward, a ceiling of 23,000 baht with a contribution ceiling of 1,150 baht. Where calculation logic or management ledgers still hold the old ceiling of 15,000 baht as a fixed value, contributions will be understated for employees whose wages exceed it. In a year when the rules change, make taking stock of hard-coded values like this the priority. It also helps to have a mechanism that monitors the 36-hour weekly ceiling on overtime and holiday work combined, set by Section 26 of the Labor Protection Act and the related ministerial regulations, on a weekly basis rather than at month end, which cuts down on post-calculation rework.

Summary — Cut the Cause of the Rework, Not the Calculation

Here is the article in short.

The hours in payroll sit not in the arithmetic but on either side of it. In the survey LayerX conducted in February 2026 (n=500), 67.4% described the payroll close as a heavy burden, while individual dependency in calculating payments and deductions — the arithmetic — was the lowest of the four processes at 72.8%, against 82.0% for collecting and checking attendance information, 77.6% for collecting and checking variable information, and 76.0% for post-calculation follow-up. And 71.6% of administrators spend at least half of their total hours on checking, correcting and confirming after the calculation is run.

We modelled that structure as Layer B (collecting and reconciling variable information), Layer C (compliance checking) and Layer D (verification, correction and exceptions). Current hours in the illustrative model come to 768.0 hours a year for Layer B, 264.0 hours for Layer C and 720.0 hours for Layer D, a total of 1,752.0 hours a year, or 350,400 baht in money terms.

Comparing investment configurations gives 4.9 years for Configuration 1 (Layer B only), 4.2 years for Configuration 2 (Layer B plus Layer C) and 7.3 years for Configuration 3 (Layer B plus Layer C plus direct Layer D support). Investing directly in Layer D, the biggest bottleneck, produces an incremental annual net benefit of -24,600 baht and never pays back. Layer C, by contrast, improves the overall payback through 50,400 baht of spillover into Layer D despite its own direct effect being a modest 44,880 baht, and pays back in 3.6 years as an increment.

Eliminating the cause of verification and correction upstream is a more efficient investment than pointing AI at the verification and correction process itself — that is what the asymmetry means. In the sensitivity analysis, Configuration 1’s payback splits between 11.4 years and 3.4 years depending on whether Layer B’s automation rate lands at 50% or 85%, while Configuration 2 stays inside 5.9 years to 3.5 years. Deploying Layer C alongside also hedges the risk of Layer B failing to bed in as hoped. Rolled out across three sites, the same Configuration 2 pays back in 2.1 years.

The calculations in this article rest on case volumes, handling times, automation rates, upfront investment and running costs that are all values set for the TOMAS TECH illustrative model rather than measured results. The 35% spillover rate for Layer C in particular decides the conclusion, so we recommend sorting your own correction cases by cause and checking it.

Where the time in your own payroll close is disappearing, whether your correction cases lean toward rule inconsistencies or individual circumstances, and whether the new SSO contribution ceiling has been reflected in your existing calculation logic. Sorting out those three points before you request quotations will cut down on rework later. We are equally happy to talk with companies still at the early stage of asking whether they should step into Layer C at all. Get in touch through our contact form and we can cover the allowance structures specific to Thai sites and the design for rolling out across several of them.

References

1. Survey on the realities of payroll processing

Source for the findings that 67.4% described the payroll close as a heavy burden (“a very heavy burden” 20.4% plus “a somewhat heavy burden” 47.0%), that individual dependency by process was 82.0% for collecting and checking attendance information, 77.6% for collecting and checking variable information, 72.8% for calculating payments and deductions and 76.0% for post-calculation follow-up, that only 33.8% of companies had standardised the work, and that 71.6% of administrators spend at least half of their total hours on checking, correcting and confirming after the calculation is run. The survey ran on 12 and 13 February 2026 as an internet survey with 500 corporate payroll administrators responding.

Survey on the Realities of Payroll Processing | LayerX Inc. (PR TIMES)

2. Overtime premiums and the overtime ceiling in Thailand

Source for the points that overtime on a normal working day is paid at 1.5 times the ordinary hourly rate and that overtime on a holiday is paid at 3 times: confirmed to agree across both sites below. The distinction between 1 time for monthly-paid employees within normal holiday hours and 2 times for daily-paid employees, and the Section 26 cap of 36 hours per week combining overtime and holiday work, rest on Kaipro alone; Avance Legal Group’s article does not state them.

Overtime in Thailand | Kaipro

Thai Labour Law and Practice | Avance Legal Group

3. Default interest under Section 9 and penalties under Section 144 of Thailand’s Labor Protection Act

Source for the points that interest at 15% per year accrues on unpaid wages where an employer fails to pay without reasonable cause, and that a further 15% per year must be added from the day seven days have passed since the due date and every seven days thereafter (Section 9, Avance Legal Group). The penalty under Section 144 for a breach of the wage-payment provisions, including Section 70 — imprisonment of up to six months, a fine of up to 100,000 baht, or both — is confirmed by Juslaws & Consult’s guide below. This article cites these as evidence that the legal framework exists, and does not estimate the probability or amount of a breach at the model factory.

Thai Labour Law and Practice | Avance Legal Group

Labor Litigation in Thailand: A Complete Guide | Juslaws & Consult

4. Increase in the wage ceiling for Thai social security contributions

Source for the points that the ceiling on monthly wages used as the basis for social security (SSO) contributions will be raised in stages from 1 January 2026, that this was announced in the Royal Gazette dated 12 December 2025, that the first stage from January 2026 to December 2028 raises the ceiling from 15,000 baht to 17,500 baht and the monthly contribution ceiling from 750 baht to 875 baht for each of employer and employee, that the contribution rate stays unchanged at 5% for each side, that the second stage from 2029 to 2031 sets a ceiling of 20,000 baht with a contribution ceiling of 1,000 baht, and that the third stage from 2032 onward sets a ceiling of 23,000 baht with a contribution ceiling of 1,150 baht.

Thailand to Raise the Wage Ceiling for Social Security Contributions in Stages | JETRO

5. Pay structure and allowances for factory workers in Thailand

Source for the points that a commuting allowance is paid on top of base pay at a fixed monthly figure of around 5,000 baht because public transport is thin outside the central urban areas, that housing allowances run from 5,000 baht to 10,000 baht mainly for supervisory grades, that the perfect-attendance allowance carries the purpose of discouraging abuse of sick leave, and that meal and living-cost allowances are provided for workers with heavy overtime or shift patterns.

Salaries and Allowances at Factories in Thailand | Kuno Yasunari CPA Office