- Seven ways to use AI to decide what to fix first, when every tool you own hands you a longer list than you can work through.
- Severity labels come from whoever built the tool. They have no idea which of your pages earns money, so the ranking you are shown is not a ranking of your problems.
- Thirty broken things are usually three causes, and the fix that closes the most rows is rarely the one at the top of the list.
- The real output of triage is the will-not-fix pile, written down and dated, so you stop re-reading the same forty warnings every month.
- The build item is the weekly digest that names one action, so the reports you already receive get read for you instead of piling up unopened.
Using AI to decide what to fix first is a different job from using AI to find things that are broken, and the second job is solved to death. Your analytics tool emails anomalies. Your site scanner flags ninety issues. Your accounting software marks twelve unmatched transactions. Your project board shows four overdue items and a blocked one. None of these tools knows about the others, so each hands you a list sorted by its own idea of importance and leaves the hard part to you.
The hard part is ranking, and the evidence that people are bad at it under volume is not anecdotal. Vectra AI's 2026 research, a survey of 1,450 security practitioners, found teams receiving an average of 2,992 alerts a day with 63 percent of those alerts going unaddressed, even as the daily volume fell year over year. The same pattern shows up in medicine, where a 2024 meta-analysis in Health Informatics Journal put the average rate at which clinicians override drug interaction warnings at around 90 percent. Both groups are highly trained. Past a certain length, people stop triaging a list and start dismissing it wholesale, which is what you do every time you close the site health tab.
So none of the ideas below detect anything. They turn a list into an order, and an order into one thing you do this week.
1. Re-Rank the Warning List by What It Costs You
Every dashboard sorts by its own severity scale, which is a guess made by a product team who has never seen your business. A critical error on a page nobody visits is not critical. A warning on your pricing page might be the most expensive line on the screen. The tool cannot tell them apart, because it does not know which pages make you money.
You do, and that fits in ten lines. So stop reading the list in the order you were handed it. Export it, paste your own facts above it, and ask for a re-ranking against those facts with a reason per row. Google's own Search Console documentation makes the same point: the question is not how many errors you have, it is whether the pages you care about are affected, and most sites carry unindexed pages that will never matter.
Priya runs a two-person pottery supply shop and had been working down a site audit from the top for three weeks, starting with missing meta descriptions on tag pages. Nineteen rows further down sat a broken checkout redirect on a discontinued kiln model that customers still searched for by name. Re-ranking against her own sales report took fifteen minutes. Fixing the redirect took four.
Start today: Export your issue list to CSV. Above it, paste ten lines of your own context: your five highest-earning pages or products, your busiest day, and what one lost customer is worth. Then ask for the list re-sorted by estimated cost to you this month, with a one-line reason per row and an explicit "no cost found" verdict where that is the honest answer.
2. Find the One Cause Under Thirty Symptoms
Long problem lists are repetitive in a way that is invisible when you read them one row at a time. Thirty flagged items are frequently three causes wearing thirty labels. One expired certificate produces failures in four tools. One renamed column breaks six reports. One page template produces the same error on two hundred URLs.
This is where pasting the entire list at once beats working through it, which is the opposite of how dashboards want you to behave. Ask which entries are probably downstream of the same cause, then ask which single fix closes the most rows. The answer reorders your week, because the fix that clears twenty children is rarely the row sorted to the top. It also exposes the quiet ones: a job that stopped running months ago shows up as a scatter of unrelated-looking gaps rather than an alarm, which is the failure mode behind the automation that quietly stopped running.
Start today: Take the longest warning list you own and paste all of it into one prompt, not a sample. Ask for the items grouped by probable shared cause, each group named by the cause rather than the symptom, with a count of rows it would close. Look only at the biggest group.
3. Split the List by Who Can Really Fix It
Half of what lands on your plate is not yours to fix. The payment processor's settlement delay, the supplier's broken feed, the slow server, the colleague who owns the spreadsheet. People burn whole afternoons staring at items they have no power over, then run out of time for the three they could have closed before lunch.
Sorting by ownership takes two minutes and almost nobody does it, because the list arrives undifferentiated and you process it the way it arrives. The second half of the move is the useful part. For an item that belongs to someone else, the correct action is not a note to yourself, it is one message to the one person, with enough specifics that they can act without a reply. Draft it while the context is in front of you, not next Tuesday when you have forgotten which of the four errors you meant.
Start today: Paste your list and ask for three buckets: yours to fix, somebody specific's to fix, and nobody's because it is a property of a tool you chose. For the middle bucket, ask for a short message per owner naming the item, the evidence, the impact on you, and the one decision you need. Send them today.
4. Hand Off the Weekly One-Thing Digest
Everything above is a session you have to remember to run, which means it happens twice and then stops. This is the item to stop doing by hand, because the inputs already arrive on a schedule without any help from you. The analytics summary, the uptime report, the error digest, the unmatched-transactions notice, the overdue-task email. Five messages a week you archive unread, each holding a piece of a ranking nobody is assembling.
A standing job can read all five, apply the ranking rule you wrote once, and send one message: here is the thing to fix this week, here is what it costs you, here is why it beat the other eleven candidates, and here is what it beat so you can disagree. Not a digest of everything, which is the same overload in nicer wrapping. One named action with its reasoning exposed. That is a reporting workflow you hand off once and stop thinking about, and the honest part is that the rule has to come from you, because the model holds no opinion about what your business cares about until you give it one.
Start today: Write your ranking rule in four sentences. What counts as expensive, what counts as urgent, which pages or customers or accounts outrank the others, and what you will never be told about. Four sentences is the entire specification, and everything downstream of it is assembly.
5. Find the Item Somebody Else Is Waiting On
Cost to you is not the only axis, and on a team it is not the main one. An item that costs you nothing can be holding up three other people, and that cost is invisible on your own list because it lands on somebody else's week. The approval nobody chased. The spec with one open question. The access request that has to clear before anybody else can start.
Devon coordinates a six-person print shop and used to rank the morning list by what was loudest. One Tuesday he fed the list in alongside the week's open threads and asked which item, if finished, would unblock the most people. It was a one-line answer to a question a press operator had asked on Friday, sitting under eleven more dramatic items for three working days. Two minutes of work had been costing two people their afternoons. This is the blockage that makes chasing approvals across email and chat such a reliable time sink, and it never looks urgent from the inside.
Start today: Paste your open list alongside the last week of messages from whoever you work with, and ask which single item unblocks the most other people, naming who is waiting and for what. Do that one before anything on your own cost ranking.
6. Write the Will-Not-Fix List and Put a Date On It
Triage produces two outputs and everybody keeps only one. You take the top item and leave the rest in a pile you will read again next month, and the month after, spending the same forty minutes rediscovering that the same thirty-five things do not matter. The dropped pile is the real deliverable, and it is worthless unless you write it down.
So make the decision explicit and durable. One file, one line per item, each with a reason and today's date: will not fix, because this page gets eleven visits a year. Will not fix, because the tool reporting it is wrong about our setup. Will not fix until we change payment providers, at which point it disappears anyway. Next month you read only what is new, usually three rows instead of forty. Most of this needs no AI at all, which is the point. What AI adds is forcing the pile into explicit sentences and then comparing next month's list against the dated one so new items surface on their own. Same instinct as using AI for getting more done, where the win comes from deleting work rather than performing it faster.
Start today: Create a plain text file called will-not-fix.md. Paste your current list, ask for a one-line "will not fix, because" sentence for every item you are dropping, then keep the ones you agree with. Put today's date at the top. Next month, paste both and ask only what is new.
7. Ask for the Stabilizing Fix, Not the Correct One
Once you have picked the item, there is one more ranking to do inside it. Most problems have a proper fix that takes a week and a crude one that takes twenty minutes and removes most of the damage. Triage in medicine is not treatment, it is stabilization, and the same split holds for a broken redirect, a failing import, or a report nobody trusts. A redirect you add by hand is not the right answer. It stops the loss today.
Ask for both and name the gap out loud, so the cheap version is a decision rather than something you settle for and quietly forget. Note the proper fix with the date you chose to defer it, which is the same habit as the item above. The logic shows up in how performance targets get set: web.dev's write-up of how the Core Web Vitals thresholds were chosen lands on seventy-five percent of page views hitting a good score rather than every single one, because perfect was never the target.
Start today: Take the item you picked and ask for three things: the twenty-minute version, the right version, and what stays broken if you only do the first. Do the twenty-minute version now and write the right version on the deferred list with today's date.
Pick One
Do not run all seven at once, or this becomes one more list you are triaging. Number six pays back fastest, because the will-not-fix file is the only one that makes every future pass shorter. If you would rather feel it this week, run number two and find out how few real causes hide behind your longest list.
Sunday we report what changed in AI, Wednesday The Loop builds one workflow end to end, and this column hands you the ideas in between. If one of these turns into something you would rather never think about again, the agent workflow directory is where that version lives.
Let something else read the reports and tell you what to fix
Pick the ranking rule you just wrote down and BYOBot turns it into an agent that applies it to every report you receive, every week, and hands you one decision instead of five inboxes.
Frequently Asked Questions
- Start with a rough bracket rather than a number. For each item, ask whether it costs you nothing, under an hour a month, a few hundred dollars, or a customer. You will be wrong on individual items and still get the order roughly right, which is the whole point. An AI model is useful here because it will ask you the questions that turn a vague worry into a bracket, and because it will do it for forty items without getting bored. Once a few items have real brackets, the rest sort themselves by comparison.
- No, and the distinction matters. The model does not know what your business cares about, so it cannot rank anything until you tell it. What it can do is read a hundred warnings you have not opened, group them, trace them to shared causes, and apply the ranking rule you gave it consistently across every row. The decision stays yours. The reading, grouping, and consistent application is the part you hand over.
- Then the list is lying to you, because severity labels come from whoever built the tool and they have no idea what your week looks like. Force a single ordering: make the model rank every item against every other item and refuse to award ties. A forced ranking you disagree with is more useful than a pile of red badges, because arguing with a specific order tells you where your own judgment sits. Our installment on support ticket triage walks through the same move on a queue of incoming tickets.
- Yes, and that is the build item in this edition. The reports already arrive on a schedule, so a standing job can read them, apply your written ranking rule, and send you one message naming one thing to fix and why it beat the others. That is the shape of a reporting workflow you hand off once. The rule is the part only you can write, and it takes about ten minutes.
