- The cost of manual work doesn't sit in a single line of the accounts. It's spread across wages, rework and customers kept waiting.
- Before approving an AI budget, calculate this number first, so you know whether the project really pays off or only feels like it should.
- Count the AI side in full too: the system, its upkeep and the time people spend checking its work.
1. Manual costs are spread across many accounts, so nobody sees them
Ask your accounting team how much manual work costs you and they won't be able to say. That's no reflection on their skill: this cost has simply never been recorded in one place. It hides in wages, rework and deals lost to slow replies.
Companies see salaries by department, but a single process crosses several departments. Taking an order might involve sales, admin, the warehouse, accounting and a supervisor, yet there is no account called “order-taking cost.” So the owner only sees that everyone is busy and assumes more people are needed, without knowing the total Touch Time one order takes.
Manual work also creates queues as volume grows. People start batching work to save time, so customers wait longer. Data copied several times leads to errors, and managers lose time chasing status updates. All of this is cost that never shows up in any one employee's hours.
2. A formula for the real labor cost of one process
The simple version is the number of times per month, multiplied by the time each one takes, multiplied by the hourly labor cost. Remember to add the time of the people who check and fix things afterward, because that is the part most people leave out.
Start with the number of items per month, multiply by the Touch Time of every role, then multiply by the labor cost per minute. Labor cost is more than salary. It should include benefits, bonuses, employer taxes, equipment, space and management overhead. If you don't have the data, start with a multiplier of about 1.25 to 1.5 times salary and adjust it for your company.
Formula: Monthly Labor Cost = number of jobs × minutes per role × cost per minute, added up across all roles. Suppose there are 2,000 orders. Admin spends 8 minutes on each, the warehouse 3 minutes, accounting 2 minutes and a supervisor 1 minute, for a total of 14 minutes per order, or 28,000 minutes a month. At an average cost of THB 4 per minute, the direct labor cost is THB 112,000 a month.
Time at least 30 to 50 real items and separate typical cases from difficult ones. Don't let staff estimate from memory, because people rarely count the time spent searching for files, opening systems or coming back to fix things later. Use the median for typical work and the P90 for planning capacity during complex periods.
| Data | Source | Watch out for |
|---|---|---|
| Job volume | CRM, ERP, Email, Ticket | Don't count a job twice when it's handed on |
| Touch Time | Timing or observing the work | Include searching and rework |
| People cost | Payroll + Overhead | Don't use net salary |
| Peak periods | Daily or weekly data | Averages hide queues |
3. Calculate the cost of errors and rework
Find the error rate from real data, such as the percentage of quotes that get revised, orders with wrong information, returned goods or reports whose numbers don't match. Then multiply by the average cost of fixing each one, including labor, shipping, credit notes, compensating discounts and managers' time. If errors damage your reputation or cause customers to leave, treat that as a separate risk. You don't need to put a number on it beyond what the evidence supports.
For example, if 3 percent of 2,000 orders are wrong, that's 60 orders. At an average fixing cost of THB 450, that comes to THB 27,000 a month. If AI checks for completeness and brings the error rate down to 1 percent, the benefit is THB 18,000 a month, on top of the staff time saved.
That said, AI can create new kinds of errors, such as misreading documents, giving wrong answers with confidence, or doing the same work twice. So you need to estimate the residual error after the new system goes in, and the cost of monitoring. Don't assume automation brings errors down to zero.
4. The cost of delay: a waiting customer still costs you
Measure time from business events as well as working time: from lead to response, from request to quote, or from order to confirmation. Then check how these relate to conversion, cancellations and complaints. If leads answered within 15 minutes close at a higher rate than those answered after two hours, delay has a revenue cost you can estimate.
The basic formula is number of opportunities × difference in conversion × profit per deal. Suppose you get 500 leads a month, faster replies could raise conversion by 2 percentage points, and the average profit per deal is THB 6,000. The opportunity is worth THB 60,000 a month. Use experiments or historical data so you don't overstate the case.
5. The opportunity cost of talented people stuck on low-value work
A sales manager who spends ten hours a week compiling reports costs you more than ten hours of wages. Those are hours they should be spending coaching the team, working on big deals or meeting customers. Opportunity value is harder to estimate than wages, so build low, medium and high scenarios and don't use the highest figure as the base case.
Ask what the company would do with 100 hours a month handed back, with a clear owner and KPI. If the only answer is “more important work,” don't count it as a financial benefit until there's a plan to redeploy the capacity, such as having the team make 800 more follow-up calls or develop two more trial products.
Another kind of benefit is the avoided hire: as volume grows, the company doesn't need to recruit. Calculate it from the workload the system can absorb compared with the capacity of the role, and the date you expect you would have had to hire. This benefit is more credible than claiming hours saved when no real expense goes down.
6. The full cost of AI you need to include
- Discovery and Process Design: time to gather data, design the steps and clean up the rules
- Development and Integration: connecting the CRM, ERP, email, databases and permissions
- Model/API: usage charges based on the number of messages, documents, images or agent calls
- Infrastructure: hosting, database, queue, monitoring, backup and logs
- Quality & Security: testing, evaluation sets, data checks and incident response
- Change Management: training users, adjusting roles, and the dip in productivity while people learn
- Maintenance: fixes when source systems or models change, plus updates to data and prompts
Split costs into one-time and recurring, then work them out month by month over at least 12 to 24 months. Build scenarios based on work volume and service pricing. If the system uses multi-step agents, the cost per item can vary a lot, so set budget alerts, rate limits and a cheaper backup model for simple tasks.
Include the cost of manual fallback and downtime, especially for critical processes. A cheap system that often goes down may force you to keep a full backup team, which means it produces no real savings.
7. Calculating ROI, payback and the value that actually materializes
Annual Net Benefit = labor costs actually reduced + error costs reduced + additional profit + hiring avoided − annual system costs. Then ROI = (net benefit − initial investment) ÷ initial investment, and Payback = initial investment ÷ monthly net benefit.
For the development team · Technical detail
Separate Hard Saving, Avoided Cost and Soft Benefit. A hard saving happens when spending actually falls, for example when a contract isn't renewed or overtime goes down. An avoided cost means not adding staff as you grow. Soft benefits, such as satisfaction or faster decisions, have value but shouldn't be used to prop up a business case that doesn't add up.
Set criteria for stopping and reviewing. For example, if after 60 days the Auto Rate is below 30 percent or the cost per item goes over the ceiling, fix it or stop. Accepting a sunk cost early protects the budget for projects that pay off better.
8. A 7-day assessment before approving the budget
- Day 1: choose the business unit and set the scope of a single process
- Day 2: pull job volumes and cycle times going back at least one month
- Day 3: time the Touch Time of every role using real examples
- Day 4: calculate errors, rework, delay and overtime
- Day 5: estimate the share of work that can be automated, reviewed or handled as exceptions
- Day 6: get estimates of development cost and monthly running costs at three levels
- Day 7: build a low, medium and high business case with KPIs and stopping criteria
Have Finance check the assumptions, and have the process owner confirm how the savings turn into real results, such as less overtime, not replacing people who leave, or more sales capacity. Without a plan to capture the value, the hours saved may simply turn into scattered idle time that never shows up in profit.
In short: Evaluating AI has to start from the economics of the process: real labor costs, errors, delays and opportunities, together with the cost of the system over its lifetime. Use scenarios and evidence instead of excitement. When owners know the cost per job and how to capture the value, they can confidently choose projects that reduce headcount or avoid new hires.
Turn the cost of manual work into numbers Finance will accept
The figures for wages, the cost of errors and the value of delayed work sit with your accounting team and line managers. What's usually missing is a repeatable way to collect the data, so every assessment produces different numbers. DNA Maker helps set up one standard way of measuring: counting work volume from real systems, recording the time spent per item and defining what counts as rework. Once the method is stable, comparisons between departments and between periods start to mean something.
From rough estimates to data you can decide on
What we build is usually an internal calculation tool tied to real data from the systems you already use. It shows the cost per item, the cost of errors and estimates after the process changes, with editable assumptions so Finance can check the logic themselves instead of taking the numbers on trust. We don't guarantee returns before measuring, but we make measuring easy enough to do before every budget approval. If you're about to propose an AI project to the executive team, this set of numbers belongs on the first slide.
SOFTWARE ENGINEERING GLOSSARY
Software engineering glossary
These terms are about measuring the costs and returns of automated work.
| Term | What it is | A simple example | What executives should ask the development team |
|---|---|---|---|
| Cost per Transaction | The average cost of completing one item of work | The cost of issuing one invoice | What costs does this figure include, and what does it leave out? |
| Rework | Doing work again because of errors or incomplete information | Revising a quote because the price was wrong | How do we count rework, and what percentage is it now? |
| ROI | The return compared with the investment | How much is saved compared with the cost of development and upkeep | Over how many years is this ROI calculated, and on what assumptions? |
| Opportunity Cost | The value lost by putting talented people on work a system could do | An analyst spends half a day combining files | If we gave this time back to the team, what would they do instead? |
| Baseline | The starting figures before a project begins, used to compare results afterward | The average time per item in the month before starting | Have we recorded a complete baseline before we start? |