By KYN AI Advisory Team — AI implementation specialists, Singapore
A process audit answers one question: which of your recurring tasks costs the most for the least judgement? This worksheet has 20 questions in five groups: volume, steps, people, data and risk. Answer them for one process at a time, score each answer, and rank your processes. The highest scores are the first candidates for automation, and the lowest are the ones to leave alone. Allow 30 minutes per process and one person who actually does the work.
How do you use the worksheet?
Pick a process with a clear start and end, such as "supplier bill to posted entry" or "enquiry to quote". Do not pick a department.
Have the person who does the task answer, with the manager present. Managers describe the process as designed. The doer describes it as it runs, and the difference is the finding.
Score each answer 0, 1 or 2 as shown. A high total means: frequent, rule-based, data already digital, low risk if wrong. Those are the safe first candidates.
Group 1: Volume
- How many times does this process run each month? 0 = fewer than 10, 1 = 10 to 100, 2 = more than 100.
- How many minutes does one run take a person? 0 = under 5, 1 = 5 to 20, 2 = over 20.
- Is volume growing? 0 = shrinking or flat, 1 = slowly, 2 = faster than headcount.
- Does it peak? 0 = no, 1 = monthly peak, 2 = peaks that cause overtime or backlog.
Multiply the answers to questions 1 and 2 to get hours a month. Under 20 hours a month, automation rarely pays for itself on time alone. See our automation ROI calculator for the arithmetic.
Group 2: Steps
- Can you write the steps as a list of fewer than 15 lines? 0 = no, 1 = with effort, 2 = yes.
- Do the steps follow rules that do not change by person? 0 = no, depends on who does it, 1 = mostly, 2 = yes.
- How many exceptions occur? 0 = more than a third of runs, 1 = a few, 2 = rare.
- Is the output a defined thing? A posted entry, a sent quote, a report. 0 = vague, 2 = defined.
- Does the process need a decision a person must own? 0 = yes, at every run, 1 = at exceptions only, 2 = no.
Questions 6 and 7 matter most. Rules without exceptions are the easiest to automate. Rules with many exceptions need a design that routes exceptions to a person.
Group 3: People
- How many people touch one run? 0 = four or more, 1 = two or three, 2 = one. Fewer hands make the change simpler to design and to test.
- Does the task depend on one person's memory? 0 = no, 2 = yes. Dependence is a risk and a reason to write the process down, so it adds to the score.
- Would the team accept a change? 0 = no, 1 = unsure, 2 = they asked for one.
- Is there a named owner of the outcome? 0 = no, 2 = yes.
Question 12 decides success more often than the technology. A team that asked for the change adopts it. A team that did not needs the reason explained first.
Group 4: Data
- Is the input already digital? 0 = paper, 1 = photos and PDFs, 2 = structured data.
- Is the data in one place? 0 = scattered across systems, 1 = two places, 2 = one.
- Is the data correct? 0 = often stale or wrong, 1 = mostly, 2 = trusted. Automation exposes bad data faster than a person does.
- Can you export from the systems involved? 0 = no, 2 = yes.
If question 16 scores 0, fix the data before anything else. An agent acts on what it reads, and it will act confidently on stale records.
Group 5: Risk
- What does a wrong output cost? 0 = cash or a legal consequence, 1 = a customer-facing slip, 2 = an internal delay.
- Can a wrong output be reversed? 0 = no, 1 = with effort, 2 = easily.
- Does the process send messages, move money, file a return or delete records? 0 = yes, 2 = no. A "yes" is not a bar. It means a named person must approve that step.
Our rule for any agent: outbound messages, money, filings and deletions wait for a named person. Question 20 tells you where to put that gate. For a method see how to set approval gates for high-risk agent actions.
How do you read the score?
Add the 20 scores. The maximum is 40. Use the total only to rank processes against each other, not as a pass mark.
| Rank | What it usually means | Action | |---|---|---| | Top third | Frequent, rule-based, digital data, low risk | Automate first, with review | | Middle third | Valuable but with exceptions or data issues | Fix data, define exceptions, then automate | | Bottom third | Rare, judgement-heavy or risky | Leave with a person; document it |
Then check the top three against a second test: if the system is wrong, who finds out, and how fast? If the answer is "a customer, in a month", add a check before you build.
What should you record from each audit?
Keep one row per process with the score, hours a month, owner, systems touched and the biggest exception. After auditing five to ten processes you will see shared causes, for example the same missing data feed behind three tasks. One fix then serves several processes, which is cheaper than automating each one alone.
Re-run the worksheet every six months. Volumes change, and so do the tools around them. Our article on where a Singapore SME should start with AI adoption covers sequencing.
Worked example with illustrative answers
An enquiry-to-quote process at a 20-person trading company, scored as the team described it. The numbers are invented to show the method.
- Volume: 80 runs a month (1), 25 minutes each (2), growing faster than headcount (2), no peak (0). That is about 33 hours a month.
- Steps: nine lines (2), rules mostly the same by person (1), a few exceptions (1), output is a sent quote (2), a person decides price only at exceptions (1).
- People: two people touch it (1), one person remembers the special terms (2), the team asked for a change (2), the sales manager owns the outcome (2).
- Data: enquiries arrive as WhatsApp messages and emails (1), in two places (1), price list trusted (2), exportable (2).
- Risk: a wrong quote is customer-facing (1), reversible with effort (1), the process sends messages (0).
The total is 26 of 40. The finding is not the score. It is the single point of memory and the gate: drafting the quote can be automated, a person approves before it goes out, and the special terms get written into a rule table that anyone can read.
Run two or three processes this way and compare. The ranking will often differ from what the team expected, which is the value of scoring before arguing.
Where KYN fits
The 14-day review KYN runs before any build uses a worksheet like this one. We inventory your stack, interview the people who do the work, time a sample of runs and rank the processes. We then say which to automate, which to fix without software and which to leave alone.
Builds ship in under 90 days. Each system runs in a cloud account in your name, with named accounts, an audit log of every action and a gate for anything outbound, financial or irreversible. For a manufacturer the result was eleven systems reduced to three on one AWS account. See operations and case studies.
Frequently asked questions
How long does a process audit take?
About 30 minutes per process with the person who does the work present. Ten processes take a day, including the ranking.
Who should answer the questions?
The person who runs the task, with their manager listening. The doer knows the exceptions, and the manager knows the goal.
Does a low score mean the process is unimportant?
No. A low score means it is a poor candidate for automation now. A judgement-heavy process can be among the most valuable in the company.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
