Internal Inquiry Automation 2026｜Four Version Drifts That Age Your Answers

Internal inquiry automation almost always starts as a counting exercise — how many of this month’s questions can we hand to a machine. At a Thai site, though, what decides whether it works is not the count. It is how long the answer stays correct. The authoritative source of an answer is split across four places, and the legally correct answer genuinely changes in December 2025 and again in October 2026. Deploy AI without deciding which version of an answer is authoritative, and automation simply distributes the outdated answer faster and wider. This article uses a Japanese-owned manufacturing site in Thailand with 520 employees as a model case, sorts its 1,240 monthly inquiries into three tiers, and shows — formula by formula — where 716,604 baht a year of return comes from, set against a three-year TCO of 1,820,000 baht.

Internal inquiry automation is not a volume problem｜it is a version problem

Proposals for internal inquiry automation are, almost without exception, written from volume. How many questions arrive each month, what share a machine could answer, how many hours that frees. It is an easy story to tell, and it usually gets approved. What it does not do is deliver.

There are two reasons. The first is that the deflection estimate is set too high. The second, and the more damaging one, is that the return comes from somewhere else entirely.

The core claim of this article fits in one line.

The return on internal inquiry automation does not come from the questions the machine answered. It comes from the questions a human should answer, reached sooner by a human.

Here are the model-case numbers up front. First-line workload reduction is worth 288,378 baht a year. Against that, the three-year TCO is 1,820,000 baht and the annual running cost is 360,000 baht. Look only at first-line reduction and you get 288,378 − 360,000 = −71,622 baht a year. That is a loss, every year. Stack it for three years and 288,378 × 3 = 865,134 baht — barely half of 1,820,000 baht.

That structure is why a proposal written from volume gets approved and then fails to deliver. The gap is closed by two other effects — 173,826 baht from eliminating rework caused by version drift, and 254,400 baht from reaching tier 3 questions faster. Together that is 716,604 baht a year, or 2,149,812 baht over three years, which is the point at which the 1,820,000 baht is finally cleared.

And both of those effects come from one thing only — deciding which version of an answer is authoritative. Rework disappears because the source document and the validity period sit in one place. Humans reach tier 3 sooner because someone drew the line between questions the machine may answer and questions it must not.

At a Thai site this is not an abstraction. The authoritative source of an answer is split across four places — the head office rules written in Japanese, the Thai-language work rules, statutory amendments already in force, and individual employment contracts. On top of that, maternity leave entitlement changed on December 7, 2025, and a new payroll deduction starts on October 1, 2026. A question whose answer flips on a date is something an AI with no concept of a validity period cannot handle on its own.

Start with where you actually stand｜73.0% report heavier load while 88.5% already use generative AI

Before evaluating anything, it is worth looking at what is happening inside corporate IT departments. A survey published by Canon Marketing Japan on July 22, 2026 covered 111 corporate IT staff at companies with 300 to under 1,000 employees. Fieldwork ran from June 12 to 15, 2026.

The thing to read in this table is that 73.0% and 88.5% stand at the same time. Generative AI is already almost everywhere. The workload is going up anyway. Tool adoption and workload relief are not connected — that is the single most important fact this survey shows.

The 62.9% figure deserves a second look as well. The most commonly cited use case is automatic FAQ creation. In other words, most companies are investing in the direction of producing answers faster and in greater volume. Produce answers faster, and you have more answers. Have more answers, and you have more contradictions between versions. Part of the reason the workload never falls is right there.

At a Thai site this contradiction bites harder. The Japanese head office holds rules in Japanese, the Thai site holds work rules in Thai, and there is no guarantee that the two reflect the same amendment at the same time. Speed up answer production alone and you get contradictory answers multiplying in two languages at once.

What this article does not cover｜product selection and retrieval accuracy live elsewhere

Let me draw the line first. This article deals with the version of an answer and nothing else. Three neighbouring articles cover the adjacent ground, so if your question is a different one, read those instead.

Product types, the overall cost picture and channel design are covered in multilingual chatbot implementation cost and rollout. Scripted or LLM-based, on Teams or on a LINE official account, how to split Thai, Japanese and English. Everything you need at the product-selection stage is there.

How to actually get retrieval accuracy is a technical question, covered in building RAG on factory knowledge. Chunking, embedding model choice, reranking, how to build an evaluation set. If the same internal documents produce poor answers, the cause usually sits in that territory.

Deciding how far AI may be used at all is covered in the generative AI usage policy for 2026. What information may be entered, how far output may drive business decisions, how violations are handled. If you are writing an internal policy, start there.

This article is none of those three. It is about one question only — how long the answer the AI returns stays correct. However good the product, however accurate the retrieval, however tidy the usage policy, if the version of an answer is undecided the AI will return the outdated one with complete confidence.

Sorting internal inquiries into three tiers｜what you may automate and what you must not

Here are the model-case assumptions. A Japanese-owned manufacturing site in Ayutthaya province, Thailand. 520 employees, of whom 8 are Japanese expatriates and 512 are Thai nationals. The administrative side is 6 people in general affairs and HR, plus 3 in corporate IT.

Average monthly salary for administrative staff is 38,000 baht. Applying a loaded-cost factor of 1.4 to cover social security and bonuses gives an annual employment cost of 38,000 × 12 × 1.4 = 638,400 baht per person. Annual working hours are 8 hours × 240 days = 1,920 hours, so the hourly rate is 638,400 ÷ 1,920 = 332.5, roughly 333 baht. Every figure below uses 333 baht. The 1.4 factor is our own assumption. The make-up of overhead differs by company, so when you read the amounts below, substitute your own loaded cost and recalculate. Every formula is written out in the text.

Current inquiry volume

Average handling time per inquiry is 11 minutes. That excludes time spent waiting for confirmation and counts only the working minutes from intake to answer. 1,240 inquiries × 11 minutes = 13,640 minutes a month = 227.3 hours. Annualised that is 2,728 hours, or 2,728 × 333 = 908,424 baht.

908,424 baht a year. Part of the payroll of nine administrative staff is being consumed by answering questions. That much shows up in any evaluation.

How to split the three tiers

The problem starts here. Treat 1,240 as a single number and you will make the wrong call, because three different kinds of question are mixed into it.

Tier 1 is what you may automate. Password reset steps, printer setup, how to use the expense system, the application flow for company housing. The answer is the same for everyone and it is not going to change any time soon. Putting AI here works exactly as advertised.

Tier 2 is what you may automate with conditions. How many days of maternity leave am I entitled to, how does annual leave carry over, how much social security is deducted. The answer is company-wide, but the version moves with statutory amendments and rule revisions. Automating this requires a mechanism that attaches a validity period to each answer first. Deploy without a ledger and you distribute outdated answers.

Tier 3 is what you must not automate. My remaining leave balance, my grade, my allowances, how a work injury is treated, the settlement on resignation. Conditions differ by employment contract and individual agreement, and judgement is required. Let AI answer here and a wrong answer maps straight onto someone’s money.

The point worth stressing about tier 3 is that not automating it is not a defeat. The value in tier 3 is handing the question to a human reliably and quickly. As you will see below, more than a third of the model-case return comes from exactly that.

How many inquiries become self-service in year one

Here is a deliberately conservative first-year deflection estimate.

394 inquiries is 394 ÷ 1,240 = 31.8% of the total. Time saved is 394 × 11 minutes = 4,334 minutes a month = 72.2 hours a month. Annualised that is roughly 866 hours, worth 866 × 333 = 288,378 baht.

288,378 baht a year. That is the first-line reduction. We will set it against cost later, but the conclusion is already visible — it does not even cover the annual running cost of 360,000 baht.

Four ways answers drift out of date at a Thai site

Why does automating tier 2 require a ledger? Because at a Thai site the authoritative source of an answer is split across four places. Here is each drift path in turn.

Path 1, head office rules in Japanese versus work rules in Thai

Under Section 108 of Thailand’s Labour Protection Act, an establishment with 10 or more employees is required to prepare work rules in Thai. The deadline is within 15 days of the day the headcount reaches 10.

In other words, the authoritative document for labour matters is the Thai-language work rules, not the Japanese rulebook held by head office. In practice this distinction matters enormously. Feed the AI the Japanese internal rulebook that was assembled for Japanese expatriates and it will return answers that carry no legal weight for 512 Thai employees.

What makes it awkward is how naturally the mistake happens. The Japanese versions tend to be the better-organised internal documents, and the more digitised ones. It is not unusual for the Thai work rules to exist only on paper and never to have been placed on the internal file server in their latest form. The document that is easiest to feed an AI is the one that is not authoritative.

Path 2, the gap between an amendment in force and work rules nobody revised

An amendment to the Labour Protection Act took effect on December 7, 2025. The main changes are as follows.

What this produces is a state where the law has changed but the work rules have not been revised. Revising work rules takes procedure and time, so a gap between the effective date and the revision date is unavoidable.

At that moment, the AI reads the work rules. The work rules say 98 days. So the AI answers 98 days. It returns the outdated figure, accurately, and with confidence. That is what makes AI dangerous here. A human officer would hesitate, half-remembering that this changed last year. An AI simply returns what is written, as written.

Without automation, that error stops at a handful of cases a month. With automation, the same error is distributed to every affected question among 430 tier 2 inquiries, around the clock, in multiple languages.

Path 3, a scheme not yet in force that flips on a date

The Employee Welfare Fund (EWF) starts on October 1, 2026. It applies to establishments with 10 or more employees, with contributions of 0.25% from the employer and 0.25% from the employee. From October 1, 2030 both rise to 0.5%. Companies enrolled in a provident fund are exempt.

This scheme is a textbook example of version drift because the correct answer today and the correct answer from October 1, 2026 are different. Asked today whether the welfare fund is deducted from pay, the correct answer is that it is not deducted yet. From October 1, the correct answer is that 0.25% of wages is deducted.

An answer with no validity period is guaranteed to get this class of question wrong. Say 0.25% is deducted before October 1 and employees are confused. Say it is not deducted after October 1 and the answer contradicts the payslip. Both generate more inquiries.

Handling answers that flip on a date requires the answer to carry a valid-from and valid-until. This is not a retrieval accuracy problem, it is a data structure problem. Which is why switching to a more capable model does not fix it.

Path 4, individual agreements that are not company-wide

The fourth is the territory where a company-wide answer does not exist in the first place. Employment contract terms, collective agreements, individual arrangements. My remaining leave balance, my grade, my allowances.

If the AI returns the company-wide answer here, it produces an answer that ignores conditions that differ person by person. An employee who negotiated a specific allowance on joining receives the standard allowance table. A collective agreement with job-specific terms is ignored and a generic answer is given instead. Because the error maps onto money, both the cost of correcting it and the damage to trust are large.

Draw the boundary between questions the AI may answer with a company-wide answer and questions it may not, before anything else. That is what separating tier 3 is for. Announce “ask us anything” without drawing that boundary and tier 3 questions flow into the same entry point as tier 1, where the machine answers them.

Answers change twice around 2026｜120 days of maternity leave and the EWF on October 1, 2026

Let me put paths 2 and 3 together on a timeline. The answers at a Thai site change twice around 2026.

What this table shows is that 2026 is a year in which the version of the answer moves. And it moves in two different ways.

The first is a version that has already moved. It is in force, so the correct answer as of today is 120 days. If the work rules still say 98 days, the work rules are the ones that are wrong. What this case needs is a diff exercise between the rules and the statute, and a shift of what the AI references toward the statutory basis.

The second is a version that is about to move. The EWF is not yet in effect today, so not deducted is correct. But on October 1 the correct answer to the same question inverts. What this case needs is a design where the answer carries a validity period and a different answer is returned automatically once the switchover date passes.

These two need different treatment. The first is settled by a diff exercise, the second requires a mechanism. Most implementation projects do only the first, conclude that the rules have been tidied up, and then have an incident on October 1.

One practical caution. Because companies enrolled in a provident fund are exempt from the EWF, establish first whether your own site is in scope. Announcing that 0.25% will be deducted from October when you are exempt makes employees brace for a deduction that will never happen. The worst configuration of all is letting the AI answer while your own scope is still undecided.

How far helpdesk automation actually gets｜the industry average of 20-30%

The part of a proposal that most often breaks is the deflection estimate. Vendor documents sometimes carry figures like a 70% reduction. The figures that are actually measured look different.

The industry average is 20-30%. The 40-60% band belongs to best-practice organisations and is not a first-year target. Proposals get written on a first-year assumption of 50% anyway. And they get approved. And then they do not deliver.

The model case sits at 31.8% because that is slightly above Gartner’s 20-30% industry average while still inside the 20-35% L1 average from HDI / MetricNet. It is nowhere near the 40-60% of best-practice organisations. 55% in tier 1, 25% in tier 2, zero in tier 3. Putting tier 3 at zero may look excessively conservative. It is deliberate. Keeping tier 3 away from the machine is the design principle of this whole article.

One clarification on how to read 31.8%. It does not mean that 31.8% of inquiries disappear. It means 31.8% are resolved without passing through a human. The remaining 68.2% are still handled by people. And the bulk of the return comes from how you handle tier 3 inside that 68.2%.

The real harm in overestimating deflection is not that you miss the target. It is that deflection becomes the only justification for the investment. If deflection is the sole basis, the project is judged a failure the moment deflection falls short. Add version control and faster tier 3 reach to the basis, and the investment still stands up at 25%.

Three ledgers to build before you deploy an internal FAQ chatbot

Now to turn all of this into a practical sequence. Before you deploy an internal FAQ chatbot or an internal policy search AI, build three ledgers. The order is fixed too.

Ledger 1, the inquiry ledger

Write out every inquiry from the last three months, one row per inquiry. Columns are the date received, the intake channel, a summary of the question, who answered it, how long handling took, and which of the three tiers it belongs to.

Request a quotation without this ledger and the vendor will come back with a configuration scoped to everything. A quotation that assumes AI answers all 1,240 inquiries. Once you know the split into three tiers, the scope narrows to the 950 inquiries in tiers 1 and 2.

Building ledger 1 requires no system. Three months of email and chat history is enough. What matters more is that you do not outsource this work. The classification judgement itself determines the design that follows.

Ledger 2, the answer version ledger

This is the most important of the three and the one most often skipped. Columns are the question ID, the text of the answer, the source document the answer rests on, the version of that document, the valid-from date, the valid-until date, and the next review date.

For maternity leave, you record that the basis is a specific section of the Thai work rules, when that document was revised, that the statutory basis is Section 41 of the Labour Protection Act, and that it is valid from December 7, 2025. For the EWF, you record a valid-from date of October 1, 2026, and hold a separate answer of “not in scope” for the period before that.

Ledger 2 works on its own, even without AI. In the model case, 13.5% of the 430 tier 2 inquiries — 58 a month — end in rework or re-confirmation because answers contradict each other. Each one takes 45 minutes, counting the officer’s re-research, the confirmation with head office, and communicating the correction to the employee.

58 × 45 minutes = 2,610 minutes a month = 43.5 hours. That is 522 hours a year, worth 522 × 333 = 173,826 baht. Those 173,826 baht start disappearing the moment ledger 2 exists. It has nothing to do with the accuracy of the AI.

Read the other way round, without ledger 2 the AI does not work. Train on as many answers as you like — if they carry no basis and no validity period, what comes out is a plausible-sounding outdated answer. Choosing the tool first fails precisely because it inverts this order.

Ledger 3, the escalation ledger

This ledger defines who receives a tier 3 question, through which route, and within what time. Columns are the question type, the first-line owner, the destination, the target time to reach that destination, and a flag for whether money is at stake.

The money-at-stake flag is the heart of this ledger. Resignation settlements, work injury claims, payroll errors. In those three types, the amount grows the longer they sit. In the model case, 41% of the 290 tier 3 inquiries — 119 of them — sit at first-line intake for 48 hours or more, and 17 a month among those are questions where leaving them alone moves money.

Whether a human reaches those 17 quickly is the source of the 254,400 baht we come to below. The amount that can be read with confidence covers only the 4 payroll-error cases a month inside that group. The role of the AI is to keep those 17 away from the machine and pass them to the right person immediately. Not answering becomes the job.

Keep the ledgers in order

Build ledger 1, then ledger 2, then ledger 3. Without ledger 1 you cannot separate tier 2 from tier 3. Without knowing tier 2 you cannot scope ledger 2. Without knowing tier 3 you cannot write ledger 3.

And do not finalise a product until all three exist. With the ledgers in hand, the required feature set narrows itself. Without them, you end up choosing on the length of the feature list. Lining up candidates once ledgers 1 and 2 are done is fine, but sign nothing until ledger 3 is complete.

The effect differs by department｜how it shows up in HR, general affairs and IT

The same internal helpdesk AI behaves completely differently by department. Explain it as if the effect were uniform and some department will end up disappointed.

The 480 corporate IT inquiries are a high tier 1 area. Password resets, printer setup, ERP operating steps. The answer is the same for everyone and statutory amendments do not touch it. So the self-service rate comes out high. This is also usually where people feel it working immediately after go-live.

On the other hand, what comes out of corporate IT is mainly first-line workload reduction. That is inside the 288,378 baht, and on its own it does not cover the 360,000 baht annual running cost. The department where the effect is visible and the department where the money is are not the same one.

The 610 general affairs and HR inquiries are the substance of this investment. Leave, payroll, insurance, company housing, rules. This is where tier 2 questions with moving versions concentrate, and where tier 3 individual judgement concentrates too.

Nearly all of the 173,826 baht of version-drift rework occurs on the general affairs and HR side. Getting the maternity leave figure wrong, checking with head office, issuing a correction. That 45-minute round trip essentially never happens in corporate IT. The same holds for the 254,400 baht of faster tier 3 reach, because payroll errors, resignation settlements and work injuries are all HR territory.

So of the 716,604 baht of return, 428,226 baht (173,826 + 254,400) comes from the general affairs and HR side. Budget this through the corporate IT department alone and you produce a proposal in which 60% of the benefit is never counted.

The 150 shop floor inquiries are an area dominated by the channel problem. Equipment requests, purchasing procedures. The volume is small, but many of these employees do not use a PC routinely, so putting the assistant on Teams does not reach them. A LINE official account or shop floor terminals are the options to consider here, and the detail of channel design is covered in multilingual chatbot implementation cost and rollout.

Here is the departmental summary.

PDPA and inquiry logs｜employee personal data lands in them

Thailand’s Personal Data Protection Act (PDPA) came fully into force on June 1, 2022. The thing to register here is that employee personal data is in scope too. This is not only about customer data.

Personal data lands in internal inquiry logs naturally. Names, employee numbers, department, and the text of the question. The text of a question carries health conditions, family circumstances, salary amounts, attendance records. A single line such as “my wife is expecting, how many days of leave can I take” contains both a family circumstance and an identifier for the person.

Tier 3 in particular is territory where information easily becomes sensitive. Work injury claims, harassment consultations, leave of absence for health reasons. Those logs are not ordinary operational records.

So draw three lines before you deploy AI.

The retention line. How many months do you keep inquiry logs? Accumulate them indefinitely and you cannot respond when a deletion request arrives later.

The access line. Who can read the logs? Developers reading logs to improve answer quality is an entirely natural practice that will emerge on its own, but do not leave tier 3 logs readable under the same permission.

The secondary-use line. Decide that logs collected to handle inquiries will not be used for performance evaluation or attendance management. This is a policy question, not a technical one.

How to actually write the policy that governs how far AI may be used at work is covered in the generative AI usage policy for 2026. This article only points out that the lines have to be drawn.

In practice the efficient moment to draw these three lines is while you are building ledger 1, the inquiry ledger. Classifying three months of inquiries makes it obvious which question types carry sensitive information.

Estimating cost in five layers

Quotations slice line items completely differently from vendor to vendor. To make them comparable, break the cost into five layers. Note that these layers are cost layers and have nothing to do with the three answer tiers above. To avoid confusion, they are referred to below as cost layer 1 and so on.

Initial cost is 120,000 + 180,000 + 350,000 + 90,000 = 740,000 baht. Annual cost is 216,000 + 144,000 = 360,000 baht. Three-year TCO is 740,000 + 360,000 × 3 = 1,820,000 baht.

What to notice in this table is the 300,000 baht that cost layers 1 and 2 add up to. Roughly 40% of the 740,000 baht initial cost goes to ledgers rather than to the AI itself.

In most quotations those two layers do not exist. Only cost layer 3, the AI platform, and cost layer 4, the channel connections, are listed, and ledger preparation is written up as something for the customer to arrange. It then materialises as your own staff hours, outside the quotation. Cost that is not in the quotation has not vanished, it has only moved.

Cost layer 5, operations, works the same way. The 18,000 baht a month is the labour of updating versions monthly, reviewing the questions that went unanswered, and retraining where needed. Skip that work and the answer versions stay frozen at go-live for a year. Maternity leave may become 120 days and the EWF may start, but the AI keeps answering 98 days and telling people nothing is deducted.

When you compare quotations, lay out on one sheet which of these five cost layers contains what. A comparison table of amounts alone is almost certainly comparing different things.

Payback does not come from the number of questions the machine answered｜the three effects

Now to build up the benefit side. There are three effects.

Effect 1, first-line reduction of 288,378 baht a year

As calculated above, first-year self-service is 394 inquiries a month, or 31.8%. Time saved is 866 hours a year, worth 288,378 baht.

This is the only effect that comes from the number of questions the machine answered. And it does not reach the 360,000 baht annual running cost. 288,378 − 360,000 = −71,622 baht a year. That structure is why a proposal written from volume alone never delivers.

Effect 2, eliminating version drift rework, 173,826 baht a year

13.5% of the 430 tier 2 inquiries, or 58 a month, end in rework because answers contradict each other. Each takes 45 minutes. That is 2,610 minutes a month = 43.5 hours, or 522 hours a year, worth 522 × 333 = 173,826 baht.

To repeat, this is the portion that disappears by building the cost layer 2 ledger, and it has nothing to do with AI accuracy. Conversely, deploy AI without ledger 2 and those 58 cases do not fall. They may well rise, in proportion to how widely the machine distributes outdated answers.

Effect 3, faster tier 3 reach, 254,400 baht a year

Of the 290 tier 3 inquiries, 41% — 119 of them — sit at first-line intake for 48 hours or more. Among those, 17 a month are questions where leaving them alone moves money.

The most readable amount inside that group is payroll errors. Average time to discovery was 19 days. Build the escalation ledger so a human reaches them sooner, and that becomes 3 days.

The number of months requiring retroactive correction falls from an average of 1.6 to 0.3, and the administrative cost per correction falls from 8,500 baht to 3,200 baht. The difference is 8,500 − 3,200 = 5,300 baht. This applies to 4 cases a month, 48 a year, so 48 × 5,300 = 254,400 baht.

Not one baht of that 254,400 comes from the AI returning an answer. It comes from the AI blocking a tier 3 question and passing it to the right person immediately. It is an effect produced by not answering.

Adding the three together

716,604 baht a year. Over three years, 716,604 × 3 = 2,149,812 baht, which exceeds the three-year TCO of 1,820,000 baht.

The annual net gain is 716,604 − 360,000 = 356,604 baht. Simple payback on the 740,000 baht initial cost is 740,000 ÷ 356,604 = 2.08 years.

What it looks like on first-line reduction alone

For comparison, here is the same calculation using effect 1 only.

Against a three-year TCO of 1,820,000 baht, a three-year benefit of 865,134 baht covers only about half.

The same investment becomes either a rejected case or an approved one purely by changing how the benefit is counted. Neither version is dishonest. But a proposal that counted volume alone has no answer when someone says during execution that the effect is not showing. Write in the effects produced by ledger 2 and ledger 3 from the start, and the investment still holds even if deflection lands below plan.

A 90-day rollout

Here is the rollout arranged over 90 days. The key point is that no product discussion appears at all in the first 30 days.

Weeks one to four are just counting

There is no product evaluation and no vendor approach in the first 30 days. The work is counting and mapping. Candidates start being lined up from day 31 when ledger 2 begins, and a contract is signed from day 60 once ledger 3 is complete.

Skip those 30 days and request a quotation, and the vendor returns a configuration scoped to all 1,240 inquiries. A quotation that assumes AI answers even the 290 tier 3 inquiries. With ledger 1 in hand, the scope narrows to 950.

Why days 61-75 launch tier 1 only

The reason for launching with tier 1 alone is to build the operating rhythm in territory where the version does not move. What you want to verify here is not answer accuracy but whether tier 3 questions go to a human instead of being answered by the machine.

Take tier 1 and tier 3 through the same entry point and tier 3 questions will get mixed in, without fail. “Tell me about annual leave” is a tier 1 question in general terms and a tier 3 question when it means the employee’s own balance. Verify that this routing works before you add tier 2.

Days 76-90 verify across a date boundary

The last slot is the core of this article. You verify that an answer carrying a validity period changes correctly across the switchover date.

Concretely, take an item with a fixed start date such as the EWF, set the system date either side of the switchover, and confirm the answer changes. An implementation that has not passed this test will start distributing wrong answers on October 1, 2026. Working and staying correct indefinitely are two different things.

Common failure patterns

1. Skipping ledger 2 and starting from the AI platform

The most common failure. The 180,000 baht of cost layer 2 is easy to cut because the deliverable looks unglamorous. But without ledger 2, not one baht of the 173,826 baht of version drift rework disappears. Worse, the AI starts distributing answers that carry no basis and no validity period, so outdated answers spread faster. That is the whole reason choosing the tool first fails.

2. Feeding the AI only the Japanese rulebook

The pattern where nobody knows the Thai work rules are the authoritative document, and the Japanese rulebook sitting on the internal file server gets used for training. Under Section 108 of the Labour Protection Act, establishments with 10 or more employees are required to prepare work rules in Thai. Answers based on the Japanese rulebook alone carry no legal weight for 512 Thai employees. And because the AI answers confidently, it takes a long time before anyone points out the error.

3. Assuming a 50% self-service rate in year one

Gartner’s industry average is 20-30% and the HDI / MetricNet L1 average is 20-35%. The 40-60% band belongs to best-practice organisations and is not a first-year target. A proposal built on 50% gets approved, then misses. And once it misses, the whole project — including the version control operation — is judged a failure.

4. Letting the AI answer tier 3 as well

The pattern where you announce “ask us anything” and the AI ends up answering individual leave balances and allowances. Because the error maps onto someone’s money, both the correction effort and the loss of trust are large. On top of that, if the AI absorbs tier 3, the 254,400 baht of faster reach never appears. In tier 3, the value is created by not answering.

5. Budgeting through the corporate IT department alone

Of the 716,604 baht of return, 428,226 baht comes from the general affairs and HR side. Write the proposal on the corporate IT budget alone and 60% of the benefit goes uncounted while the entire cost lands on IT. Naturally it fails to get approved. Even if it does, there is no basis for asking general affairs and HR to carry the operating workload, so the monthly version update never happens.

6. Cutting the 18,000 baht a month of operations

Cost layer 5 gets cut first because its deliverable is invisible. But that labour is precisely the monthly work of updating versions and reviewing the questions that went unanswered. Cut it and the answer versions freeze at go-live. The EWF starts on October 1, 2026 and the AI keeps telling people nothing is deducted. Build ledger 2 and it still goes stale within a year if nobody updates it.

Frequently asked questions

Where should internal inquiry automation start?

Not from comparing products, but from writing out the last three months of inquiries one row per inquiry. That work requires no system. Once written out, classify them into three groups — routine, company-wide and stable in version (tier 1), routine and company-wide but with a moving version (tier 2), and individual with judgement required (tier 3). In the model case, 1,240 inquiries a month split into 520, 430 and 290. Request a quotation before that classification exists and you get back a configuration scoped to everything, tier 3 included.

How much does an internal FAQ chatbot cost?

It depends on scope, and on whether you build the ledgers in-house or contract them out. In the model case, initial cost was 740,000 baht, annual cost 360,000 baht, and three-year TCO 1,820,000 baht. The breakdown is 120,000 baht for building the inquiry ledger, 180,000 baht for the version ledger, 350,000 baht for the AI platform, 90,000 baht for channel connections, plus annual costs of 216,000 baht for operations and 144,000 baht for licences. When comparing quotations, always check whether the two ledger layers are included. A quotation without them is not cheaper — that portion has simply moved outside the quotation and into your own staff hours. The spread of costs by product type is covered in multilingual chatbot implementation cost and rollout.

How much do inquiries actually fall with helpdesk automation?

Gartner’s industry average is 20-30%, best-practice organisations reach 40-60%, and the HDI / MetricNet L1 average is 20-35%. The model case assumes 31.8% in year one, made up of 55% in tier 1, 25% in tier 2 and zero in tier 3. The thing to watch is that 31.8% is not the rate at which inquiries disappear, it is the rate at which they are resolved without passing through a human. The remaining 68.2% are still handled by people, and most of the return comes from how those are handled.

Should we automate general affairs inquiries or corporate IT inquiries first?

The realistic split is to launch with corporate IT and to justify the investment on general affairs and HR. The 480 corporate IT inquiries are heavily tier 1 with stable answers, so the self-service rate comes out well and the area suits building an operating rhythm. The money, however, is in the 610 general affairs and HR inquiries — essentially all of the 173,826 baht from removing version drift and the 254,400 baht from faster reach comes from that side. Confuse launch order with budget justification and you end up with 60% of the benefit missing from the proposal while the entire cost sits on corporate IT.

Can an internal policy search AI just read the Japanese rulebook?

For questions about labour matters at a Thai site, that is not sufficient on its own. Under Section 108 of the Labour Protection Act, establishments with 10 or more employees are required to prepare work rules in Thai, within 15 days of the day the headcount reaches 10. The authoritative document for labour matters is the Thai version, not the Japanese rulebook at head office. Feed it only the Japanese rulebook and it will confidently return answers that carry no legal weight. For areas outside the scope of the Thai work rules, such as internal procedures aimed at Japanese expatriates, the Japanese rulebook is fine.

How do we measure the effect of reducing corporate IT inquiries?

Do not measure by volume alone. There are three metrics to track. The first is the number of self-service resolutions, which corresponds to the 288,378 baht of first-line workload reduction. The second is the number of rework cases caused by contradictory answers, which in the model case is 58 a month heading toward zero. The third is the number of tier 3 items sitting for more than 48 hours — 119 a month, of which 17 are cases where money moves. Measure all three before go-live, or you will reach day 90 and conclude that nobody can tell whether it worked.

How should the AI handle the EWF starting in October 2026?

By giving answers a validity period. There is no other way. The EWF starts on October 1, 2026, applies to establishments with 10 or more employees, and takes contributions of 0.25% from the employer and 0.25% from the employee. From October 1, 2030 both rise to 0.5%. So the correct answer to the same question changes twice, on October 1, 2026 and again on October 1, 2030. An answer with no validity period is guaranteed to get this class of question wrong. Note also that companies enrolled in a provident fund are exempt, so establish first whether your own site is in scope.

Do inquiry logs carry PDPA risk?

Yes. Thailand’s PDPA came fully into force on June 1, 2022, and employee personal data is in scope. Inquiry logs naturally accumulate names, employee numbers, health conditions, family circumstances and salary figures. Tier 3 inquiries about work injuries, harassment and health are particularly prone to becoming sensitive information. Draw the three lines — retention, access and secondary use — before you deploy AI. How to develop that into a written policy is covered in the generative AI usage policy for 2026.

Will changing the model fix the problem of AI returning outdated answers?

No. This is a data structure problem, not a retrieval accuracy problem. If the work rules say 98 days of maternity leave, even the most capable model will answer 98 days. Knowing that the December 7, 2025 amendment made it 120 days requires the answer to be linked to a source document, a version and a validity period. Raising retrieval accuracy itself is covered in building RAG on factory knowledge, but improving accuracy leaves the version problem untouched.

Summary

Internal inquiry automation is not a question of how many inquiries you hand to a machine. It is a question of how long the answer stays correct. At a Thai site the authoritative source of an answer is split across four places, and the legally correct answer genuinely changes — maternity leave went from 98 to 120 days on December 7, 2025, and the EWF starts on October 1, 2026. Deploy AI without deciding the version and automation distributes the outdated answer faster and wider.

Sort inquiries into three tiers. The model case split 1,240 inquiries a month into 520 in tier 1 with a stable version, 430 in tier 2 where the version moves, and 290 in tier 3 requiring individual judgement. Tier 1 may be automated. Tier 2 may be automated once a version ledger is attached. Tier 3 must not be automated, and its value lies in handing questions to a human reliably and quickly.

Realistic deflection is the industry average of 20-30%. The 40-60% band belongs to best-practice organisations and is not a first-year target. The model case assumes 31.8%, saving 866 hours a year, worth 288,378 baht.

Then the numbers. Three-year TCO is 1,820,000 baht. On first-line reduction alone, 288,378 − 360,000 = −71,622 baht a year, a loss, and a three-year benefit of 865,134 baht covers only half the TCO. Add 173,826 baht from eliminating version drift rework and 254,400 baht from faster tier 3 reach, and only then do you get 716,604 baht a year and 2,149,812 baht over three years, clearing 1,820,000 baht. The annual net gain is 356,604 baht, and simple payback on the 740,000 baht initial cost is 2.08 years.

The return on internal inquiry automation does not come from the questions the machine answered. It comes from the questions a human should answer, reached sooner by a human. So build three ledgers, in this order. The inquiry ledger, the answer version ledger, the escalation ledger. Choosing a product comes after that.

Talking to us while you are still evaluating

Internal inquiry automation starts with counting, not with choosing a product. How many inquiries arrived over the last three months, how many of them are questions whose version moves, how many require individual judgement. That classification needs no system, and you do not need outside help to produce it.

In practice, though, plenty of sites start collecting quotations before that table exists anywhere internally. Ask a vendor in that state and you get back a configuration scoped to everything. Unsure where to draw the classification lines, unsure how far to take the diff between work rules and statute, wanting to see how the payback moves once your own numbers go in — those are the conversations we have regularly. Even if it is only a review of your inquiry ledger, well before any decision to implement, feel free to contact us while you are still evaluating.

References

• 2026 Survey on Internal Helpdesk Operations in Corporate IT Departments — Canon Marketing Japan, July 22, 2026. 111 corporate IT staff at companies with 300 to under 1,000 employees. Fieldwork June 12 to 15, 2026
• Amendments to the Labour Protection Act in Thailand — GVA Law Office, December 26, 2025. In force December 7, 2025. Maternity leave from 98 to 120 days, employer-paid portion from 45 to 60 days, and other changes
• Thai Labour Law and Practice vol.11, Rules and Practice on Work Rules — GVA Law Office, August 30, 2021. Section 108 of the Labour Protection Act. 10 or more employees, Thai language, within 15 days
• Introduction of the Employee Welfare Fund — Asahi Business Solutions Thailand, published August 27, 2025 and updated March 9, 2026. Starts October 1, 2026, 10 or more employees, 0.25% each from employer and employee, 0.5% each from October 1, 2030, companies enrolled in a provident fund exempt
• Overseas Legal Systems, Kingdom of Thailand — Personal Information Protection Commission, updated January 25, 2022. PDPA fully in force June 1, 2022
• AI Service Desk Deflection Rates 2026 Benchmarks Reconciled — servicedeskagents.com, last checked April 2026. Gartner industry average 20-30%, best practice 40-60%, HDI / MetricNet L1 average 20-35%