Read this for your industry
Optional. Choose your industry and this piece opens with a situation from your world and closes with a first move for you. Your choice carries across the guides and analysis on this site.
You lead advice quality across a licensee
The Royal Commission's final report set out how charging fees for no service became a systemic failure across the largest advice licensees, and the compensation that followed reached about $3.7 billion at the four major banks alone. The harm built up because the same failure repeated across thousands of client files and nobody fixed it at the source.
Your file reviewers see a smaller version of that pattern every quarter. They find the same missing step in Statements of Advice at several practices, the finding goes into a report, and the next quarter it appears again. Once agents start preparing advice documents, a correction made to one draft helps one client. The same correction written into a rule helps every client the agents touch from then on.
What this means for you
- Treat every finding that appears in more than one file review as a candidate rule for the agents that prepare advice documents.
- Keep one correction log for the whole licensee, owned by the head of advice or compliance, so a lesson learnt in one practice reaches all of them.
- Review the log each quarter against the latest file review results, and retire any rule that no longer earns its place.
You run a territory where the same problem keeps appearing
When the Federal Court ruled against Woolworths and Coles in September 2025, the problem it described had been repeated across stores for years: rosters kept informally and overtime that nobody recorded, as SBS reported. More than 27,000 salaried managers were covered by the case.
Area managers see this kind of pattern on every visit cycle. One store fixes an issue after you point it out, and three weeks later another store makes the same mistake. Agents that draft rosters or summarise trading will repeat errors in exactly the same way until someone writes the correction down where every store's agent can read it.
What this means for you
- Keep one correction log for your territory, and add a rule the second time you see the same problem in any store.
- Share the log with your store managers, so they can see what changed and why it matters.
- Bring the log to your regional meetings, because a rule that holds in your territory will often hold in your colleagues' territories too.
You lead clinical quality across a group of homes or clinics
The aged care Royal Commission found that substandard care often occurred in routine areas such as food, medication management and skin care, the everyday tasks that repeat many times a day. In general practice, the RACGP warns that AI scribes rarely fill in structured fields such as allergies and blood pressure, so records can be incomplete unless someone checks them.
If you lead a group of homes or clinics, these are the gaps a correction log can close. A clinician who adds a missing allergy to one AI-drafted note has fixed one record. A rule that tells every agent to look for allergies and record them in the structured field fixes the records that follow, at every site.
What this means for you
- Collect repeat corrections from clinicians and from audit findings into one log that every site uses.
- Write each rule in plain clinical language, and have a senior clinician approve it before any agent works from it.
- Tell staff when a rule they suggested goes into the log, so they keep reporting what they see.
You lead engineering or manufacturing quality
The 2025 DORA report found that AI adoption is linked to higher delivery throughput and, at the same time, to lower delivery stability. More change is flowing through the same checks, and DORA's advice to leaders includes strengthening safety nets such as automated testing and fast feedback.
Manufacturers have long handled repeat defects through corrective action: find the cause, change the process and confirm that the defect stops. When code review catches the same agent mistake for the third time, or a supplier comparison keeps leaving out freight costs, the same discipline applies. The fix goes into the agent's instructions and into a test that catches it next time.
What this means for you
- Run a corrective action process for agents, with a named owner, a cause and a change recorded for each repeat defect.
- Pair every new rule with an automated check, so a regression is caught even when someone forgets the rule.
- Share the log across teams, because the same agent mistakes tend to appear in every codebase or supplier category that uses the same tools.
You run an installer whose crews keep making the same labelling faults
When the Clean Energy Regulator inspected 1,278 battery installations between July 2025 and April 2026, the most common problems were missing or incorrect warning labels, unlabelled backup circuits and badly placed energy storage labels. These are repeat faults, made by many crews in the same way.
A correction log works like a job pack that keeps learning. Each time an inspector, a customer or your own reviewer finds a fault, the fix becomes a line in the checklist that every crew and every review agent works from. The next crew starts the job with the lesson the last crew paid for.
What this means for you
- Add every inspection finding to the log on the day you receive it, together with a photo that shows what correct looks like.
- Load the log into the agent that reviews job photos, so it checks for each known fault before the crew leaves site.
- Talk through new rules at the morning toolbox meeting, because crews follow rules they understand.
The same conversation, three times
My coordinating agent had one job when I asked for research: gather the material and brief me, so I could decide what to do with it. It kept writing the finished document instead.
I corrected it: “Research and brief me. I’ll decide what to build.” It apologised and did it properly. Two days later, in a new session, it wrote the finished document again. The following week it did the same.
I was frustrated, and then a little embarrassed, because I had seen this problem many times before. I had just never seen it in software.
The book behind the counter
Most shops keep a communication book behind the counter. The morning shift writes that the delivery is due at two, that the EFTPOS terminal on register three keeps dropping out, and that a customer will call back about a lay-by. The afternoon shift reads it before they start. Without the book, every shift starts from scratch and the customer gets asked the same questions twice.
An AI agent starting a new session is the afternoon shift with no book.
Why agents forget
A model carries nothing from one session to the next unless something puts the lesson back in front of it. Researchers have worked on this for a while. The MemGPT paper, for example, treats an agent’s context like computer memory, moving information in and out of tiers, because by default each session starts fresh.
Across an organisation the cost adds up. Fortune’s write-up of MIT NANDA’s State of AI in Business 2025 report described the core problem as a “learning gap” for both tools and organisations: generic tools stall in enterprise use because they do not learn from or adapt to workflows. Picture fifty people correcting the same agents in fifty private chats. Each of them is writing in a book nobody else can read.
What a correction log looks like
A correction log is the communication book for your agents: a short, plain file of rules the agent loads every time it starts. Mine began with three lines, each from a mistake I had corrected at least twice.
- When asked to research, return findings and options, and leave the final document to me.
- Finish with the answer or the next step, and skip “let me know if you have any questions”.
- When asked to test, report the test results, and keep review comments for a separate pass.
I then added a second file for each agent, covering its role, what it can decide alone and when it should stop and ask.
The rule I wrote after the third research slip worked where three conversations had failed. The agent was no smarter. The lesson was simply there at the start of every shift.
How to run one
Write down repeats. Fix a mistake in conversation the first time it happens. Write it down the second time.
Describe the behaviour you want. “Be concise” is a wish. “Lead with the answer in two sentences, then give the detail” is a rule an agent can follow.
Load it every time. The rules work only if the agent reads them at the start of every task. Most agent tools now support a standing instructions file for this.
Point to the rule when it slips. Saying “that’s rule three” is quicker than explaining again, and it tells you whether the rule needs rewording.
Share it. A fix the finance team writes for over-long answers will help the product team too. One shared book is worth more than ten private ones.
Why it matters beyond one person
Writing corrections down is the step from operator to manager in the agent boss stages. Microsoft’s Work Trend Index expects agents to move from assistants to digital colleagues who take on tasks at human direction. Colleagues need onboarding, and a correction log is onboarding in writing.
It is also the cheapest piece of AI governance you will ever build: a file anyone can read that shows what each agent has been told and why.
If you have ever opened a shop on a Monday and found a clear note from Sunday’s shift, you know how much easier the day goes. Leave your agents the same kind of note.
Next in the series: the delegation trap, on why capable agents keep doing tasks they should hand on.
Take your last two quarters of file review findings and pick the three that appear most often. Write each as a one-line rule, load them into the agent that drafts advice documents, and check whether those findings return in the next review cycle.
List the last five things you corrected on store visits. Any that appeared in more than one store becomes a rule this week, written in one line that a store manager and an agent can both follow.
Ask your clinical leads for the three corrections they make most often to AI-drafted notes or care documents. Turn each into a rule, have it approved clinically, and check a sample of records a month later to see whether the error has stopped.
Pull the last month of code review comments or supplier comparison corrections. Group the repeats, write the top three as rules with a matching test, and track whether they come back over the following month.
Write the three labelling faults the regulator found most often into your job pack this week, with a photo of a correct example for each. From then on, check every job against them before sign-off.
Questions readers ask
Can't the model just learn from my corrections?
Some tools now keep a memory of your preferences. It is usually personal, partial and hard to inspect. A written rule file is visible, easy to change and can be shared, which matters as soon as more than one person relies on the same agent.
What goes in a correction log?
Corrections you have had to make more than once. For each one, write what the agent did, what it should do and why. Keep the file short enough that the agent can load all of it every time.
Who should own the rules in an organisation?
Name one owner for each agent or workflow. They approve new rules, retire old ones and make sure a fix one team finds reaches the others.
Sources
- MIT report: 95% of generative AI pilots at companies are failing, Fortune, 18 August 2025
- MemGPT: Towards LLMs as Operating Systems, arXiv (Packer et al.), October 2023
- 2025: The year the Frontier Firm is born (Work Trend Index), Microsoft WorkLab, April 2025
The views here are my own, not those of my employer or its clients.
More in The agent boss
What is an agent boss? How to manage AI agents like a team
An agent boss briefs, checks and manages AI agents the way a good manager runs a team. What the role involves, the four stages it grows through and where people stall.
The agent bossThe delegation trap: why capable people and agents keep doing the work
Strong performers struggle to delegate because doing a task themselves is faster this week. AI agents fall into the same habit. How to spot it and how to break it.
The agent bossThe confidence trap: when an AI agent is certain and wrong
AI agents sound equally sure whether they know or are guessing, and people tend to trust that fluency. Two standing rules that make agents say what they don't know.
The agent bossCulture comes before agents: what AI exposes about how you work
AI agents do exactly what they are told, so they find every habit a team never wrote down. Why culture comes first with agents, and how to make it explicit.