By KYN AI Advisory Team — AI implementation specialists, Singapore
A results ledger records four numbers for each automated workflow every month: hours returned to the team, errors caught, cost per unit processed and minutes spent in review. It gives you evidence to renew, narrow or stop a workflow. This article sets out the ledger, how to fill it in and the unit costs from KYN's own systems. It uses only measured values from live systems and gives no savings claims for a workflow you have not measured.
Why keep a ledger at all?
Automation projects drift. A year after launch, nobody can say whether the reconciliation agent saves anything, or whether the reports are still read. A ledger fixes this with one page per workflow and one row per month.
It serves three decisions:
- Renew or stop. If the numbers do not move, the workflow is not worth its retainer.
- Widen. If review time is low and errors are caught, you can allow more actions to run unattended.
- Narrow. If errors rise or review time climbs, tighten the rules.
The ledger does not prove that automation is good. It shows whether this workflow, in your business, earned its place this month.
What are the four columns?
| Column | Definition | Where it comes from | |---|---|---| | Hours returned | Hours the previous manual process took, less hours of review and exception handling now | Baseline time log, then review log | | Errors caught | Mistakes the system flagged before they left the building: duplicates, mismatches, missing fields | The audit log, filtered by flag type | | Cost per unit | Running cost divided by units processed | Cloud and model invoices; unit counts from the log | | Review minutes | Time people spent approving per 100 items | Approval timestamps |
Add a fifth column if useful: errors that got through, found later by a person. This is the honest counterweight. A ledger that records only wins is advertising.
How do you set the baseline?
Before the workflow goes live, measure the manual version for two weeks. Use a time log with the same tags every week. See our method for auditing where admin hours go. Record:
- minutes per unit (per invoice, per lead, per report);
- units per week;
- error rate from the last quarter, if you have one: corrections, credit notes, re-sent reports.
Without a baseline, "hours returned" is a guess. With one, it is subtraction. Keep the baseline on the ledger page for good, so the comparison does not move.
What do unit costs look like in practice?
Narrow tasks cost cents. These are from KYN's live systems, and they depend on the model and the text length, so treat them as a guide to the order of magnitude, not a quote.
| Task | Approximate running cost | |---|---| | Morning report to one department on WhatsApp | S$0.05 per report | | Outreach email with one checkable first line | S$0.02 per email | | Reply sorted by the reply desk | S$0.01 per reply | | Group benefits comparison to a client PDF | S$0.03 per comparison | | Article from the article engine | S$0.06 per article | | AI search check | S$0.05 per check |
These cover the model and compute for the task. They exclude the build fee, the retainer and people's review time. A complete cost per workflow adds all three: running cost, a monthly share of the retainer, and review minutes at your loaded hourly rate.
What does an error ledger look like?
Errors are where a ledger earns trust. Two examples show how to record them.
Finance. A manufacturer's ledger on our system holds 3,309 reconciled rows with zero duplicates. A zero is only meaningful next to the check behind it. Record the matching rule, how many candidate duplicates were flagged and how many a person confirmed. Both numbers matter. A system that flags nothing may be missing things.
Sales. In an outreach workflow where each first line cites a source the reviewer can open, the reviewer rejects lines whose source does not support the claim. Record the rejection rate. A rising rate means the research step is drifting and needs attention before the next batch.
For every workflow, define in advance what counts as an error, who confirms it and where it is recorded. Without the definition, the count is whatever someone chose to write down.
A worked layout
One page per workflow, in this order:
- Header. Workflow name, owner, start date, what counts as one unit, what counts as an error.
- Baseline. Minutes per unit, units per week, error rate before launch, with the dates measured.
- Monthly rows. Month, units processed, hours returned, errors caught, errors that got through, running cost, review minutes per 100 items.
- Decisions. Date, decision (widen, hold, narrow, stop), reason, who decided.
Hours returned for a month is a formula: units processed times baseline minutes per unit, divided by 60, less review minutes and exception handling. Write the formula on the page. A reader who disagrees with the baseline can change one cell and see the effect, which is better than arguing about the total.
For a morning report, a unit is one report. For reconciliation, a unit is one statement row. For outreach, a unit is one email drafted and approved. Pick the unit once and keep it, or the months will not compare.
How often do you review the ledger?
- Weekly, five minutes. The owner of the workflow glances at errors that got through and review minutes.
- Monthly, thirty minutes. The owner and a manager read the page, decide to widen, hold or narrow, and write the decision in a notes line.
- Quarterly. Compare the total across workflows with the retainer. Drop any workflow that has not earned its place.
Keep the decisions on the page. In a year, the notes tell you why a rule exists.
What should the ledger never claim?
- Savings you did not measure. If you have no baseline, leave the cell blank.
- Hours as headcount. Eleven hours returned across ten people is not a salary. It is eleven hours.
- Zero errors as proof of quality. Show the number of checks as well.
- Comparisons with other firms' averages. Your workflow, your baseline.
Where KYN fits
KYN builds the ledger into the systems it delivers: every action is logged with the actor, so errors caught, review times and unit counts are queries, not a reconstruction. The workflows run in a cloud account in your name, with named accounts and row-level security. Outbound messages, money, filings and deletions wait for a named person, and the approval timestamps feed the review column.
The retainer, at least twelve months, covers changes, monitoring and fixes, and the ledger is how both sides judge whether it is working. See pricing, case studies and the related piece on automated bank reconciliation in Singapore.
Frequently asked questions
Which number matters most?
Errors that got through. Hours returned can rise while quality falls. A low count of errors that got through, with healthy review times, is the sign of a sound workflow.
Can I build a ledger in a spreadsheet?
Yes, to start. One tab per workflow, one row per month. It becomes a query when the system already logs actions with the actor.
Do unit costs include the retainer?
No. The unit figures cover model and compute only. Add the retainer share and review time for a full cost per workflow.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
