Back-Office AI | Where to Start in Accounting, General Affairs, and HR

Invoice processing in accounting. Contract checks in general affairs. Payroll-related calculations in HR. The term back-office AI turns up everywhere now, and yet the practical question stays unanswered: which of our own processes should we actually start with? Plenty of administrative departments at Japanese manufacturers in Thailand are stuck at exactly that point. This article lays out, on a single map, where AI is genuinely useful across the indirect functions, so you can find the case closest to your own situation and move on to the detailed article that covers it.

What Back-Office AI Is — Systems That Carry Out or Support Indirect-Function Work

Start with the words themselves. Leave this vague and people inside the company end up talking past each other very quickly.

The definition is a system that carries out or supports indirect-function work

Back-office AI is the umbrella term for systems that use a combination of generative AI, AI agents, and RPA to carry out or support the work of indirect functions such as accounting, HR, labor relations, general affairs, legal, and sales administration. These are the functions that generate no revenue directly but without which the company cannot run.

The kinds of work in scope look like this. Reading the contents of an incoming invoice and entering it into the accounting system. Pulling out the risky clauses in a draft contract sent over by a trading partner. Answering the questions employees keep asking, such as how many days of leave they have left or where a particular application form lives. Turning a recording of a meeting into minutes. Calculating severance pay and various allowances exactly as the internal rules specify.

None of it is difficult for the person doing it. All of it eats time. And at most sites this work has collected around one specific individual, so when that person takes leave, it stops. That structure is what back-office AI is really trying to solve.

RPA is the hands and feet, AI is the brain

When back-office automation comes up, RPA is usually the first thing people picture. Plenty of Japanese manufacturing sites in Thailand already run it. So what does adding AI on top actually buy you?

An explanatory article published by Persol Business Process Design in March 2026 frames it as a division of labor in which RPA is the hands and feet and AI is the brain. That contrast matches how the work actually feels.

RPA repeats a recorded procedure without the slightest deviation. Open the specified file in the specified folder, take the value in the specified cell, put it in the specified field on the specified screen. For that kind of work it is faster than a person, more accurate, and it never gets tired. But it can do nothing that is not in the procedure. Change the format and it stops. It will never notice that the digits on an invoice look wrong.

What AI, and generative AI or AI agents in particular, handles is precisely the part that cannot be reduced to a procedure. Reading the required fields out of invoices whose layouts are all different. Interpreting what a passage means and finding the internal rule it relates to. Detecting something exceptional and routing it to a human. The presence of judgment is the decisive difference from RPA on its own.

So the two do not compete. AI takes on the part that requires judgment; RPA carries out the fixed operations. Only in that combination does a real process connect end to end.

The decisive difference is whether exceptions can be handled

Push a little further and the gap between RPA alone and back-office AI comes down to a single question: how are exceptions dealt with?

Here is a pattern that shows up constantly at sites that automated with RPA alone. Eight cases in ten flow through the robot automatically. The remaining two — the trading partner with an unusual format, the document with a handwritten note in the remarks column, the case that goes through a different approval route than usual — all come back to a human. The result is that the job changes from steadily processing transactions into adjudicating nothing but exceptions, and the perceived workload can actually go up.

The value of adding AI sits on that smaller share. By referring to past processing history and internal rules, it works out which known pattern an exception most resembles and presents it to a person together with a proposed response. Even when the final call stays with the human, the effort of researching it from scratch disappears.

Which is why, when evaluating back-office AI, asking how many exceptions come back to a person tells you far more about reality than asking how many transactions can be processed.

Why Back-Office AI Is Drawing More Attention in 2026

The term is not new. What is worth examining is why evaluation has accelerated in 2026, and published survey data gives some shape to that. One caveat first: the studies below are global and weighted toward North America and Europe. None of them is a statistic about Thailand or about Japanese manufacturers specifically.

66% of organizations plan to invest in AI over the next three years

According to Deloitte’s 2025 Global Business Services Survey, 66% of GBS organizations reported that they plan to invest in AI over the next three years.

GBS refers to the organizational model in which indirect work such as accounting, HR, procurement, and IT is consolidated and delivered as a service. It is the familiar shared services center arrangement that global companies use to pull the indirect work of their country sites together. So that 66% figure describes organizations whose entire specialty is back-office work, all moving toward AI investment at once.

Japanese manufacturing sites in Thailand rarely have a GBS operation on that scale. But the underlying problem is identical: indirect work has to run efficiently. In fact, the smaller the headcount, the wider the range of work each person carries and the more heavily it depends on individuals.

Nearly 90% say routine work is already changing

A more recent data point comes from The Hackett Group’s 2026 GBS Key Issues Study, in which nearly 90% of GBS leaders said that AI is already changing routine work.

The tense matters. This is not a prediction that things will change; it is a statement that they already are. In the same study, 63% reported experiencing early gains. That means more than half of these organizations are past the inconclusive pilot stage and have something tangible to point to.

Yet only 25% expect broad deployment

The same Hackett Group study carries a second figure worth holding onto: only 25% of organizations expect broad AI deployment during 2026.

Nearly 90% acknowledge the change and 63% feel early gains, and still only one organization in four expects to reach company-wide rollout. That gap is what best describes where back-office AI actually stands in 2026.

The reason for the gap is almost certainly on the business side rather than the technical side. Results in one process are one thing; the moment you try to widen the scope you run into procedures that differ by department, integration between systems, and rules governing authority and approval. Trying something is easy. Spreading it is hard.

Turned around, that means the question facing a company starting now is not whether to adopt. Most organizations have already reached the entrance. The difference shows up in how far they managed to spread it.

Automation rates have stalled mid-range, and attention has shifted to agentic AI

One more set of figures, from an SSON survey cited by Auxis: 56% of shared services organizations remain at a middling automation rate of 25 to 50%, and 65% rank agentic AI as their top investment priority.

An automation rate in that band means, in plain terms, that the accessible work has largely been automated and the rest is out of reach. It overlaps exactly with the structure described in the previous chapter, where exceptions come back to people.

And the direction of interest has moved away from AI that responds to one-off instructions and toward agentic AI that can carry several steps forward on its own. Read the figures as the shop floor’s own judgment showing through: lifting an automation rate that has stalled mid-range requires something that can move across process boundaries.

Vendors are turning toward the back office as well

Product roadmaps point the same way. In April 2026 Salesforce announced Agentforce Operations, an AI agent product aimed specifically at back-office work. The company states that it shortens business cycle times by up to 70% and reduces manual effort such as data entry by up to 80%.

Those two figures, however, are vendor-published claims about the vendor’s own product. Neither the conditions assumed for the work nor the conventional method used as the comparison baseline is disclosed. There is no guarantee of the same effect in your own environment, so if you quote them in an internal document, always attach the source and never carry them across as a target.

That single example is the only place this article names a specific product. Comparing which product is better serves no purpose before the business requirements are settled.

Which Work Back-Office AI Covers — Five Areas

This is the core of the article. The back office divides into five areas, and what falls in scope in each is set out below. Four of them are functional areas; the fifth is the automation foundation that spans departments. Where you find something close to your own situation, follow the detailed article linked at the end of that area.

Area 1, accounting — invoice processing and journal entries

This is the most advanced area. An invoice arrives from a supplier, the amount, the counterparty, and the accounting month are extracted, and the record is created in the accounting system. That sequence resisted full automation by RPA alone precisely because the formats vary, so manual work persisted in it for years.

Add AI-based reading and the fields can be identified even when layouts differ. On top of that you can layer processing such as proposing account codes by reference to past journal history, or raising a warning when the amount or supplier name does not match the purchase order data.

At Thai sites the variety is greater than in Japan, not smaller, because invoices issued only in Thai, documents prepared in English for the Japanese parent, and Japanese-language paperwork from Japanese suppliers all flow in at once. Validating reading accuracy against a real stack of the documents your own company actually receives is not optional.

The practical sequence for this area, how to think about touchless processing, and the cost picture are covered in detail in invoice processing automation and touchless processing. If accounting is your likely starting point, begin there.

Area 2, general affairs — contract review and internal inquiries

General affairs has two entry points, and they are quite different in character.

The first is contract review. Going through master trading agreements, non-disclosure agreements, and service contracts to surface clauses that disadvantage your company or that differ from your established standard wording is work that general affairs usually absorbs at sites with no dedicated legal function. AI can take this as far as reading the contract and presenting candidate issues. The final judgment and the negotiation stay with people, but the effort of narrowing down where to look drops sharply. The practicalities of adoption are covered in implementing AI contract review.

The second is internal inquiries. Work rules, expense claim policies, where to find each application form. The answer is written down somewhere, but asking is faster than searching, so the same questions land on the general affairs desk over and over. Letting an AI that has ingested the internal documents provide the first-line answer fits multilingual sites particularly well, since it makes it possible to ask a question in Thai about a rule written in Japanese and get an answer back. The design essentials are collected in automating internal help desk responses.

Area 3, HR and labor — payroll-related calculation and recruitment admin

In HR and labor, calculation work governed by written rules is the representative candidate.

At Thai sites the heaviest of these is severance pay calculation. It has to account for the payment rate corresponding to years of service, the scope of wages that count, and the relationship between statutory requirements and internal rules, and getting it wrong leads straight to a labor dispute. The volume is not high, but each case carries a long list of things to check, and accuracy tends to drop whenever the responsible person changes. How to fit AI into this specific process is covered separately in using AI for severance pay calculation.

Recruitment administration is also in scope: organizing applications against the requirements, drafting the correspondence to arrange interview schedules. What should be avoided is any design that hands the hiring decision itself to AI. Being unable to explain the basis for a decision creates problems in both labor and ethical terms. Keep it positioned strictly as a tool for first-pass organization.

Area 4, company-wide — meeting minutes

The classic example of work that arises regardless of department and belongs to nobody’s job description is minute-taking. Transcribe what was said from the recording, separate decisions from open actions, distribute to the participants. None of it is hard, but it recurs with every meeting, so the cumulative hours are impossible to ignore.

Multilingual sites carry a heavier version of this, because the content of a meeting that ran in a mix of Japanese and Thai has to be shared in both languages. The operational design for this area is covered in putting automated meeting minutes AI to work.

Area 5, a shared automation foundation across departments

The fifth is not a process at all. It is the ground the other four stand on.

Introduce AI separately in accounting, then general affairs, then HR, and before long the tooling multiplies haphazardly, one set of tools per process: reading handled by AI here, approval by an existing workflow there, system registration by RPA somewhere else. Let that continue and the maintenance owners and the configuration settings scatter along with it, producing a fresh form of individual dependency in which nobody dares touch anything once the original person transfers out.

The answer is to settle on one company-wide shape for the division of labor in which AI judges and RPA executes. This idea recurs in the later chapters, so for now hold onto one point only: automating an individual process and building the foundation that connects those processes are two different conversations.

How to Spot Work Where AI Pays Off — Three Tests

Looking at the five areas will not tell you where to start. What follows are three tests you can apply to an individual process.

Test 1, are the rules written down?

Does a document exist that states the criteria for judgment in this process? Work rules, accounting policies, approval authority rules, standard operating procedures. When the criteria are written in that form, AI can refer to them and use them as the basis for a judgment.

Conversely, a process whose criteria live only in a veteran employee’s head needs one step before anything else. That step is not installing AI. It is writing the criteria down. It looks like overhead, but it doubles as handover material for the day the role changes hands, so it has value whether or not AI ever enters the picture.

Test 2, how often do exceptions occur?

What share of total volume departs from the standard procedure? A process where exceptions stay around one case in ten will show the benefit of automation cleanly. A process where close to half the cases are exceptions calls for a different first move: sort out why the exceptions are being created.

Most exceptions have accumulated as the residue of accepting a different way of working for each trading partner. In that situation, aligning the practices themselves before introducing AI produces a far better return on the investment.

Test 3, is the input data structured?

Is the thing being processed data emitted by a system, or is it paper, a PDF, or a free-text email?

This one is easy to misread, so a clarification. Unstructured data is exactly where AI earns its keep, and handling it is the point of AI. But the more unstructured the input, the more effort goes into validating reading accuracy and designing the fallback path for exceptions. For a first project, a process where the data is already reasonably tidy makes it easier to confirm the effect quickly.

Here is how to use the three tests together.

A process that meets all three is ideal, but in practice meeting two is enough to make it a serious candidate.

The same process can produce very different results depending on the design

Choosing a process that passes the three tests does not by itself determine the outcome. Within the same process, the quality of the design moves the result a great deal.

There is a clear illustration of this. According to the accounts payable benchmark study Ardent Partners published in 2025, the touchless rate for invoice processing, meaning the share handled automatically without human intervention, averages 32.6%, while best-in-class companies reach 49.2%.

That is a gap of roughly 17 points inside the same activity. It cannot be explained by whether AI is in use. Standardizing formats with suppliers, designing the match against purchase order data, working out who adjudicates exceptions and how: that accumulation of process design is what shows up as the difference.

How to read the touchless rate properly, and the concrete steps for raising it, are covered in the invoice processing article linked above. Here the point is only the general rule that design creates a gap within identical work.

Considerations Specific to Thai Operations — Three Languages, Labor Law, and the Information Gap

Every figure so far comes from global studies weighted toward North America and Europe. Japanese manufacturers operating in Thailand work under conditions those studies do not capture. What follows is not statistics but a set of practical considerations.

Consideration 1, internal documents mixing Japanese, Thai, and English

The documents an administrative department in Thailand handles span at least three languages. Rules and notices from the Japanese parent arrive in Japanese; filings to local authorities and notices to Thai employees are in Thai; exchanges with the regional headquarters and overseas suppliers are in English. It is common for documents with identical content to be managed as separate files per language.

For AI adoption this situation is both a burden and a source of benefit. A burden because reading accuracy has to be validated separately in all three languages, which is more work than a site in Japan would face. A source of benefit because translation and cross-site distribution already consume real labor, so the reduction can be substantial once it runs well.

One practical warning. If you still hold internal documents whose Japanese and Thai versions say different things, feeding them to AI produces inconsistent answers. Decide which language version is authoritative before you start.

Consideration 2, Thai labor law and PDPA

Using AI in HR and labor means working within Thai labor legislation and the personal data protection law, generally known as PDPA.

Payroll, attendance, evaluation, and health information all demand careful handling. Whether such data may be fed into an external AI service is not a technical question; it is a question of internal policy and the terms of your service agreement. Whether the setting that excludes your input from training is enabled, which country’s servers process the data, and how employees are informed and give consent: settle those three before the process design begins.

The realistic approach is to start with work that contains no personal data. Contract review and internal inquiries are easy candidates for a first project on exactly that basis.

Consideration 3, the information gap between Japanese management and local staff

The third is organizational. At Thai sites there is a structural tendency for the Japanese managers who make the decisions and the local staff who do the daily processing to hold different pictures of what the work actually involves.

Japanese managers see the output of the process, but not who is absorbing how much rework upstream of it. Local staff who can see the inefficiency often have limited routes for turning that into a proposal that reaches the top. Evaluate AI adoption in that state and you end up selecting a process that was never particularly burdensome while the genuinely congested work stays untouched.

The remedy is simple. Before deciding what to start with, ask the people actually doing the work how long a single case takes and how many exceptions come up in a month. That interview also becomes the baseline you will need later, when you want to confirm the effect in numbers.

How to Roll Out Back-Office AI — From a Single-Process PoC to Horizontal Expansion

The approach itself holds no surprises. Keeping to the order is the part that matters.

Stage 1, run a PoC on one process only

The first job is to narrow the scope to a single process. Draw up candidates using the three tests from the previous chapter and pick the one that satisfies them most fully.

The temptation to run several processes at once is strong, and best resisted. The reason is measurement. Move several at the same time and, whether the result is good or bad, you lose the ability to isolate where it came from.

The purpose of a PoC is not to deliver a benefit. It is to find out what gets in the way when the idea meets your own operations. Reading accuracy, the real proportion of exceptions, the connection to existing systems, acceptance on the floor. Obstacles you cannot see until you try will always turn up.

Stage 2, confirm the effect in numbers

Measure the same indicators before and after the PoC. Which indicators depend on the process, but the following set is easy to work with.

That last line is there for a reason. The effect of back-office AI cannot be captured by time savings alone. Work only one person could handle becomes work others can run too, and at a site with a small headcount that change is worth more than the hours saved.

Once you have the numbers, they become the input for the next investment decision. Without them, the case for adoption turns into a matter of impressions, and that is where evaluation stalls.

Stage 3, expand into adjacent work

Once the first project has produced something tangible, widen it to adjacent work. Adjacent means either similar processing inside the same department, or the steps immediately before and after in the flow of the same document. Results in invoice processing point toward payment processing; results in contract review point toward contract storage and expiry management.

This is the stage where the automation foundation from the earlier chapter starts to matter. Bolt on a different tool for each process and maintenance eventually falls behind. Designing the division of labor in which AI judges and RPA executes as one common shape, then loading processes onto it, turns out to be the easier route to expansion. The technical integration methods, and where to draw the line between what AI takes and what gets passed to RPA, are covered in connecting RPA and generative AI, which is worth reading when you reach the point of planning a second project.

The gap described in the second chapter, where nearly 90% acknowledge the change but only 25% expect broad deployment, is created precisely at this stage. Treat succeeding with the first project and spreading it sideways as two different kinds of difficulty.

What to Watch Out For with Back-Office AI — Three Walls

Having set out what can be expected, here are the walls worth understanding before you begin.

Wall 1, hallucination, meaning incorrect output

Generative AI assembles plausible output from the patterns it has learned. As a result it will cite clauses that do not exist, invoke criteria that appear in no internal rule, and mix numbers up. And because the output is well formed, it looks correct to anyone who does not already know the answer.

In back-office work that property translates directly into risk. Amounts, dates, payment rates, statutory requirements. Every one of them causes real damage when it is wrong.

The countermeasure sits in operations. Always treat AI output as a first draft, and have a person verify anything touching numbers, legislation, or internal rules against the primary source. The essential part is building that verification into the workflow as a step, rather than leaving it to individual diligence. Listing rework and correction volume among the indicators above serves partly to show whether this check is working.

Wall 2, handling confidential information that cannot leave the company

Most of what the back office handles cannot go outside the company. Trading terms, figures tied to cost, employee personal data, unannounced HR decisions.

You need to decide what may be entered before people start using the tool. In practice the easiest path is to begin with a simple rule that highly classified information is never entered, then widen the scope as the business need becomes clear, checking the terms of the service agreement as you go. At a Thai site, work through the PDPA points above at the same time.

None of this is a reason to stop evaluating. There are several processes you can start with using low-sensitivity documents, such as reviewing standard contract templates or answering questions about internal rules.

Wall 3, integration with existing core systems

The third is the least glamorous and the one most likely to trip you up.

Even if AI reads the invoice correctly, the work is not finished until the result lands in the accounting system. Does the existing system expose a way for an external system to write to it? If not, can screen operation substitute? And will that substitute survive the next system update? Leave this unresolved and, however accurate the AI is, a person ends up retyping at the end.

Include this check within the PoC. Validating reading accuracy alone and declaring a pass leaves you searching for an integration method right before production rollout, which is exactly where projects stop.

Checklist — Which Entry Point Fits Your Operation

Finally, a table for narrowing down where to start from your current situation. The row where the most items apply is your first candidate.

If several rows apply, compare them using the three tests from the earlier chapter: whether the rules are written down, whether exceptions are infrequent, and whether the data is tidy. And if every row applies, the recommendation still stands. Start with one.

Frequently Asked Questions

What is back-office AI?

It is the umbrella term for systems that use a combination of generative AI, AI agents, and RPA to carry out or support the work of indirect functions such as accounting, HR, labor relations, general affairs, legal, and sales administration. The clearest way to frame it is a division of labor in which RPA, repeating fixed procedures, is the hands and feet, while AI, which supplies judgment, is the brain. The difference from RPA on its own lies in whether exceptions requiring judgment can be handled.

Which process should we start with?

Apply three tests. Are the criteria for judgment written down as documents? Is the frequency of cases departing from the standard procedure low? Is the data being handled reasonably tidy? A process meeting two or more of the three is a first candidate. Starting with work that contains no personal data also keeps you from having to retrofit compliance with the Thai personal data protection law.

How much does back-office AI cost to implement?

Cost varies enormously with the scope of the target process, the method of integration with existing systems, and the number of document languages involved, so no single market rate can be quoted. What informs the decision is less the price than the effect measured in the PoC. Time per case, the share of exceptions returned to people, and the number of cases handled by someone other than the usual owner. If you can compare those three before and after, you have what an investment decision requires. Comparing prices alone without those numbers gives you no way to assess whether the price is reasonable.

We already use RPA. Do we also need AI?

The volume of exceptions currently coming back to people answers that. If RPA is carrying most of the transaction volume yet the workload on the operator has not fallen, time is most likely going into handling the exceptions that remain, and adding AI to supply the judgment is worth doing. Where a process generates almost no exceptions, RPA on its own is sufficient.

How should we measure the effect of back-office AI?

The key is not to measure time savings alone. Alongside time per case, look at the share of exceptions returned to people, the number of cases corrected after registration, and the number of cases handled by someone other than the usual owner. The last of these carries more weight than time savings at a site with limited headcount. None of them can be compared without a pre-start baseline, so begin measuring the moment you decide to proceed.

Summary

Here are the key points of this article.

Back-office AI is a system that uses a combination of generative AI, AI agents, and RPA to carry out or support the work of indirect functions such as accounting, HR, labor relations, general affairs, legal, and sales administration. It can be framed as a division of labor in which RPA is the hands and feet and AI is the brain, and what separates it from RPA alone is whether exceptions requiring judgment can be handled.

The attention it draws in 2026 shows up in the data. Deloitte’s GBS survey puts the share of organizations planning AI investment over the next three years at 66%. The Hackett Group’s 2026 study reports that nearly 90% of GBS leaders say routine work is already changing, with 63% experiencing early gains, while only 25% expect broad deployment during 2026. The question has moved from whether to adopt to how far it has been spread.

The work in scope sorts into five areas: invoice processing and journal entries in accounting, contract review and internal inquiries in general affairs, payroll-related calculation in HR and labor, meeting minutes across the whole company, and the automation foundation that connects them. Choose your starting process using three tests, namely written rules, exception frequency, and how tidy the data is. Touchless rates in invoice processing show that design creates a gap within identical work, averaging 32.6% against 49.2% for best-in-class companies.

Thai sites add three further conditions: internal documents mixing three languages, compliance with labor law and PDPA, and the information gap between Japanese management and local staff. The sequence is a PoC narrowed to one process, confirmation of the effect in numbers, then expansion into adjacent work. Understanding three walls before you begin, namely hallucination, the handling of confidential information, and integration with core systems, will save you a good deal of backtracking.

It is entirely fine to be at the stage where you cannot yet say which process should come first. TOMAS TECH works from the realities of administrative departments at Japanese manufacturers in Thailand, and we are happy to start by taking stock of the current work together. If you would like to talk through what a realistic entry point might look like for your own operation, get in touch through our contact page.

References

• AI Assembly Lines, State of AI in Shared Services 2026 Benchmarks, the reference source for the figures from the Deloitte 2025 Global Business Services Survey, The Hackett Group 2026 GBS Key Issues Study, the SSON survey, and the Ardent Partners AP Benchmark 2025
• Japan AI, report on Salesforce announcing Agentforce Operations, an AI agent product for back-office work, published 11 June 2026, with the original announcement made by Salesforce in April 2026
• Persol Business Process Design, explanatory article on automating back-office work with generative AI, published 4 March 2026