In partnership with
Stop losing deals in between meetings.
Aligned is changing how B2B teams run their most important deals.
Most reps guess where a deal really stands. They forecast on a gut feeling and the buyer’s word vs what's actually happening. Then they realize the deal slipped through only after it's too late. Aligned surfaces what buyers are actually doing inside every deal: the risks, the openings, and the exact next step to keep things moving.
Start increasing your close rates and stand up your first deal room in minutes.
iPrompt
DEEP DIVE · iPROMPT #158 COMPANION
How to triage AI findings
By the end you’ll have a way to turn one AI review into three queues (fix now, fix later, won’t fix), and checks for the two ways triage goes wrong: throwing out real problems, and rating them too low.
6 MIN READ · FICTIONAL WORKED EXAMPLE AND CHECKLIST
TRIAGING AI FINDINGS / WORKED EXAMPLE
287 findings, ten fixed first
Ashby & Rowe is fictional, and so are its figures. It’s a 25-person design consultancy with 64 live client and supplier contracts. Its operations lead can give contracts four hours a week, and its solicitor charges £260 an hour. Before an AI reviewed the contracts, she wrote a three-level scale: HIGH if it could cost money, a client or legal exposure within a year; MEDIUM if real but able to wait for renewal; LOW for the rest.
The review returned 287 findings in an afternoon, with the contracts attached so each finding could point to its evidence. At 45 minutes each, working the raw list in order would take about 215 hours, more than a year of her contract time. That’s the cost of working it mechanically, not a measured saving over careful review. The bigger risk was order: what mattered most would wait its turn.
Step | Time | Result |
1 Merge duplicates | 40 minutes | 164 left. 123 were repeats, or the same issue under two headings |
2 Evidence: a quote and where it sits or, for a missing term, where it should be | 1 hour | 128 supported. 36 unverified, kept in their own pile rather than deleted |
3 Rate all 164 against the scale | 1 hour | Supported: 10 HIGH, 36 MEDIUM, 82 LOW. Unverified: 5 serious if true, 31 not. The AI drafted; she decided |
4 Read every HIGH and every serious-if-true in full | 2½ hours | 9 HIGH confirmed; 1 down to MEDIUM, as a liability cap sat in a schedule. 1 unverified was real and HIGH; 4 closed |
5 Check 10 of the other 31 unverified, picked at random | 30 minutes | 2 real, both LOW; 8 closed. Rule: one serious result means reading all 31. None was |
6 Check 20 of the 121 MEDIUM and LOW, picked at random, for reality and rating | 40 minutes | 18 real; 2 not (LOW, deleted). 1 LOW re-rated MEDIUM. Rule: a rating two levels out means re-checking that category. None was |
7 Fill three queues | 1 hour | 10 fix now, 38 fix later, 81 won’t fix; 21 unverified parked |
Fix now. Ten HIGHs: four client contracts with uncapped liability, three supplier contracts that auto-renew for 24 months or more, and three freelancer agreements with no assignment of the intellectual property in the work. One of those three came from the unverified pile: the AI had flagged it without saying where it looked. A rule that deleted every finding without a quote would have lost it, because a missing clause can’t be quoted.
Capacity. Triage took 7 hours 20 minutes. The fixes need about 25 more hours of her time and six of the solicitor’s (£1,560). That’s about eight weeks at four hours a week; the managing director wanted the HIGHs closed by 4 December, six weeks away. She raised the gap, and he approved two extra hours a week.
Fix later. 38 MEDIUMs. Updating the two standard templates stops 21 of them recurring in new contracts, but it doesn’t change the live contracts that carry them. So all 38 are diarised against each contract’s renewal date, with a side letter where renewal is more than a year away.
Won’t fix. 81 LOWs, each with a reason that names consequence, likelihood and alternative, such as: ‘Notices must go by post. If it bites, a termination is delayed by days; both parties are UK-based, so it’s unlikely; we’ll change it if the contract is renegotiated.’ The 21 unchecked unverified findings are parked, low-stakes if true, for the next review.
What the checks can’t tell you. Eighteen real out of 20 is encouraging, but a random sample that small leaves a wide margin: the true share could be anywhere from about two-thirds to nearly all. One re-rating in 20 doesn’t rule out others. So checking followed the stakes: everything rated HIGH or serious-if-true was read in full, and each sample came with a rule for what a bad result would trigger.
The record. ‘Contract review, 21 October. 64 contracts; 287 AI findings; 164 after merging; 128 supported, 36 unverified. Scale written before the run. 10 HIGH, each read in full, due 4 December with two extra hours a week approved; 38 MEDIUM at renewal; 81 LOW won’t fix, reasons logged; 21 unverified parked. Random samples: 18 of 20 real, one re-rated; 2 of 10 unverified real, both LOW. Approved by the managing director.’
TRIAGING AI FINDINGS / THE PARALLEL
What curl’s new policy does and doesn’t show
curl’s change isn’t a won’t-fix queue, and Daniel Stenberg says as much: true vulnerabilities can’t be ignored, and the work per report ‘does not shrink by doing them in public’. From 26 October, LOW and MEDIUM reports go to public GitHub issues so more people can help assess and fix them; only HIGH and CRITICAL stay private.
What it does show is a team sorting by severity because the volume left no choice: over 90% of submissions, he writes, could quickly be judged below HIGH. Bank of America answered the same pressure by buying capacity, 1,000 contractors. Most organisations can do neither at that scale.
TRIAGING AI FINDINGS / THE SETUP
Carry it over
Last week’s companion built checks for AI doing the work. This one is for AI checking the work.
Scale and hours first. A scale written after you’ve seen the list bends to fit it.
Park the unproven. Ask for a quote and its location or, for a missing term, where it should be. Findings without that go in their own pile, read in order of what they’d cost if true.
Check by stakes. Read every HIGH in full. Sample the rest at random, for rating as well as reality, and decide in advance what a bad sample triggers.
Escalate overflow. If the HIGHs don’t fit your hours, say so and get a decision: more time, more help or a later date.
Queue | What goes in | Who signs | Revisited |
Fix now | HIGH, plus a capacity decision if it overflows | Whoever answers for the risk | Weekly until clear |
Fix later | MEDIUM, each with a trigger | Whoever owns the documents | At the trigger, usually renewal |
Won’t fix | LOW, with consequence, likelihood and alternative | The person who’d answer for it | When the facts change, or yearly |
The strongest objection
‘Knowingly leaving problems unfixed is worse than not knowing.’ With 287 findings and four hours a week, most of the list waits either way. Triage decides whether the order is chosen and written down, or set by whichever finding happened to be opened first. A recorded reason is a judgement you can defend or revisit. Where a finding touches a legal duty, ask whoever would answer for it: a won’t-fix reason is a decision, not a defence.
THIS WEEK’S FORECAST
By 30 June 2027, a single curl release will fix more than 24 vulnerabilities, beating 8.23.0’s record of 14 October. The record fell twice this year, to 18 in June and 24 in October, and Stenberg does not ‘anticipate this flood to decrease within the next several months’.
Counting. It counts if curl’s release announcement or security page lists 25 or more fixed in one release before the deadline, as curl counts them. Void if curl stops publishing per-release figures.
YOUR MOVE
Take one contract or one policy, ask an AI to review it, and sort what comes back into the three queues before fixing anything. Reply with one line: how many findings, and how many made FIX NOW. Leave out names and confidential details.
The list isn’t the work. Choosing from it is.
New to iPrompt? Every Wednesday, the week’s AI news turned into one thing you can do. Subscribe free
Disclosure: the original draft used Claude, made by Anthropic, whose Mythos model appears in the newsletter.
iPROMPT / 158 /