Internal operations automation

Overnight checks, AI meetings, morning reports and settlement verification, all handled by automation.

Checking eight sites three times a day, reading the results to pick out changes, and writing an eight-section management report: all automated. This includes collecting settlement statements from 26 clients and preparing the issuance ledger, a task cut from a full day for one person to one hour.

The only task left to a person is reading the report that arrives at 8 a.m. and setting priorities. The other five of the six stages finish automatically overnight.

We run an automation system with the same structure we deliver to clients in our own operations, every day. We test it on ourselves first because failed automation can still look normal on screen. When records and lists look unchanged but no new data is coming in, you cannot tell from the screen alone.

We check not “Was it installed?” but “Did it arrive today?” Each run records success or failure in one line, and a stopped run is flagged first in the next morning’s report. We apply the same structure to the systems we build for clients.

This automation is designed and run for our internal work, so it is presented separately from client cases.

From checks to decisions, the system runs one cycle every day.

Each system is not a separate job but part of a single cycle. Check → judge → collect → report is one cycle, and the cycle restarts at the same time every day.

Each stage also shows its “risk”: the problem that arises if the stage is not designed this way. Real incidents and figures are covered in the three sections below the cycle.

Legend: an empty square marker means rules; a filled marker means the judging party

• Rules Process data by predefined rules and always give the same result for the same input
• AI Handles qualitative judgments that rules cannot: the verdicts in the early-morning meeting and the one sentence at the top of the morning report
• People Read the report and set work priorities

AI handles two of the six stages. Checks, progress monitoring, collection and report assembly are all handled by rules, because these stages must give the same result for the same input. AI is used only for qualitative judgments that rules cannot handle: the verdicts in the early-morning meeting and the one sentence at the top of the morning report. AI does not write the report. Rules fix the figures in all eight sections, the split between completed and recurring items, and the priority marks.

In one cycle, only the last stage involves a person. Checks, judgment, collection and sending do not wait for anyone. Fixes are not automated. The checks only read and never change their targets, because if the checker changes what it checks, later figures cannot be trusted. So what arrives in the morning is not a fixed result but a report listing what needs fixing.

Changes made by a person are picked up in the next check. When we added a search engine site ownership tag to one site, it appeared in the next check’s “changed items” list without anyone reporting it (2026-08-09).

Additional jobs outside the cycle

• 07:30 Client settlement statement collection and verification
• 09:00 · 15:00 · 21:00 Grant program announcement collection

With the cycle’s seven runs and the three jobs above, the system runs 10 times a day. At 07:30, meeting result collection and settlement verification run in the same batch.

VROF rechecks AI’s “done” reports every day.

AI gives answers every day, but it does not tell you whether those answers were acted on. Because output piles up daily without errors, it can take months before anyone asks about results.

VROF uses a mechanism that automatically marks whether a proposal was actually acted on, and reviews the previous day’s proposals every day. Beyond the proposals, the next run also checks whether those marks are correct.

If you only add a mechanism that marks items “done,” no one checks whether the mechanism itself is right. Items marked as done but never actually applied get mixed in, creating a state where unfinished work looks finished. With no errors or warnings, the logs look perfectly normal. In our own system, five of nine items were wrong, and the review of the previous day’s proposals caught them in the next run.

• Reviewing the previous day’s proposals before the meeting

  Past proposals and what changed since are placed at the front of the meeting materials, so no meeting starts without knowing the previous discussion.

• Structured meeting minutes

  Free text cannot be checked the next day, so each proposal is recorded in fixed fields: what, who, why and how it will be verified.

• Log-based evaluation instead of scorecards

  Proposals, comments and processing times are logged, and results are calculated from the logs. A scorecard distorts work toward raising the score, and AI is even more prone to this.

• Verdicts only when there is a past proposal

  If there was no proposal the previous day, the verdict field stays empty. Forcing the field to be filled makes AI invent unsupported content.

The rule is “completion is judged only by verified values,” and the log records whether each proposal met it. Proposals that did not are recorded too, so results cannot be overstated later.

A recent meeting record includes the following. About a figure that had dropped, the meeting wrote that it did not read it as success or failure and classified it as “no record of cause.” About a figure that had risen, it wrote that it did not conclude it was a “search success”, because traffic sources had not yet been separated. VROF designed the system to record “unknown” as it is. That way, a report of real improvement can be trusted.

The August 23, 2026 meeting was the 12th, with 4 new proposals and 27 verdicts on past proposals. Past proposals received more than six times the attention of new ones, and the number of completions recorded by that meeting itself was 0. When AI cannot mark items “done” on its own, the completed column shows 0, and that is the accurate figure. Compared with a log showing “12 completed,” a log showing 0 is more trustworthy.

Rules fix the figures in the management report.

AI writes sentences as instructed, but in the process numbers can change. For example, the report alone cannot show whether “shortfall of ○○ won” means the amount needed on a specific date or the total for the whole billing cycle.

In one of our own reports, a total was read as the amount needed on a specific date, overstating it by nearly four times. The solution was not “a better model.” The lesson was that the same problem recurs unless judging and transcribing are separate steps. VROF sets this separation before building anything.

• Classification based on source data

  Completed items, recurring items processed automatically each month, and groupings that are not actual work are separated using fields in the existing data. Every item in the report can be found as-is in the source data.

• Reports go out even if some lookups fail

  Items that could not be checked are flagged at the top of the report. An error in one source does not stop the whole day’s report.

The original instructions given to the AI

• Do not create or change any numbers or items
• Copy it exactly without changing a single character. Do not summarize, merge or omit anything
• Do not recalculate amounts, invent funding methods or state that a transfer was made
• Do not add news, market forecasts or business ideas that are not in the source data

The instructions contain one more line. It is the most important sentence in this system.

If there is a “lookup failure,” treat the figures in the section that item feeds as “not checked,” not “none.”

“Nothing to pay” and “could not check what is due” mean completely different things. If both states look the same, you cannot tell whether a quiet morning is truly safe. The distinction must be made when the report is built.

The one sentence AI writes answers “What should be handled first today?” We provide the priority rules with it: work with fixed external deadlines comes first, work that incurs penalty tax if late goes in the morning, work that brings in cash comes next, and development without external deadlines comes after. This judgment is hard to define as rules, because the answer depends on what overlaps that day.

We also check whether the checking system itself has stopped.

A check can only report results when it runs. So on a morning when the whole server is down, every check is silent. With no warnings, both the screen and the report look normal. On one day, six morning hours were missing entirely. Every job in that window was skipped, yet nothing reported a problem.

“No alerts arriving” and “nothing to alert” looking the same is the problem companies with automated checks discover last.

• Run counts by time slot

  Apart from individual checks, we count runs per time slot to find periods when the system stopped. If a count cannot be read, it is shown as “unreadable,” not 0.

• Stating what the report covers

  The following note appears at the end of that section.

This section cannot detect a failure of the 08:00 report itself. A day with no report at all is that signal.

For your company, a missing morning report is itself the most important warning. Knowing this, you will not misread a quiet morning.

Key internal automation cases

The first stage of the cycle and the 07:30 job outside it are each presented as a case. Because these are internal systems, we show screens recreated with mock data and figures recorded by the tools themselves, instead of public URLs.

We apply automation proven in-house to client operations.

Get figures scattered across multiple systems in a single report every morning: warehouse stock, receivables, the previous day’s sales, overnight automation runs and more. The targets differ, but the structure is the same.

When you hand work to AI, proposals pile up, but whether they were carried out must be checked separately. What VROF builds is a structure that reviews the previous day’s proposals. It scales the same way whether you cover five areas or ten, and prevents reports that show only favorable figures.

Also, AI-generated numbers never enter management decisions, because rules fix all figures, classifications and selections. We do not build reports that cannot answer “Where did this number come from?”

Scope of internal operations automation

The systems on this page are internal tools designed for VROF’s own work processes. Products for sale are introduced on the In-house products page. This system is evidence that the same approach can be applied to client operations.

By area, data collection is covered under Web data collection, standardizing and aggregating collected data under Data analysis, and judgments that rules cannot handle under AI adoption.

We check whether your automation is actually working.

Automation you believe is running may actually have stopped. We check it based on the last date new data came in.

Reply within 1 business day