Finds the order by number, or the signed-in customer’s latest order. Returns status, carrier, tracking and dates. Refuses orders that belong to someone else. Flags a parcel as lost when tracking has stalled 7 days (US) or 10 (international).
Demo store: Alder & Ash is fictional; the model runs are real.
Your support bot says it resolved the ticket. Did it?
After our rescue, the bot resolved 73.3% of tickets correctly, with 0 wrong answers. Before, it resolved 22.5%, and 45.2% of the tickets it closed were wrong.
For your customers: the right answer, or a person. Never a confident wrong one. Every ticket that needed a person reached one. Three fresh rounds, each on 356 new tickets, after an initial test (how we got here).
One square per ticket the bot never saweach reply checked against the order record
- Resolved 261
- Right handoff 93
- Needless handoff 2
- Missed escalation 0
- False resolution 0
Real resolutioncorrect, no person needed
73.3%was 22.5%
73.9% is the most a bot can solve; the rest need a person
Dashboard would show
73.3%was 41.0%
False resolutions
0.0%was 45.2%
Missed escalations
0.0%was 19.4%
Needless handoffs
0.8%was 51.3%
Fresh round 3 of 3: 261 of 263 resolvable tickets resolved, 0 wrong answers, 0 of 93 missed escalations. 2 tickets went to a person when the bot could have answered (safe). One of them, a promo-code question, the old bot had answered: the one ticket that got worse. Both are listed below.
The scorecard
Vendor dashboards count every chat the bot didn’t hand off as resolved. We count a ticket as resolved only when the reply is correct and complete against the store’s policy and order records.
356 new tickets the bot never saw, from fresh round 3 of 3 (the final test). Same AI model and graders in every round. How we got here: every round
Real resolution by topic
BeforeAfterRead the tickets, before and after
Every reply below came from the same model. Pick a ticket to see the customer’s message, the true order record, both replies, the code checks and the judge’s verdict.
What we changed
The order we work in: find why tickets fail, fix the knowledge base, then give the bot the order lookups and handoff rules it was missing. Then test on tickets nobody has seen, and fix what that test shows.
Why the default bot failed
1 Knowledge base
2 Actions on the store’s systems
Applies the return policy in code: 30-day window, final-sale items, 90-day defects. Starts the return or issues the refund itself, up to $150. Above that it asks for approval and hands the customer to a person.
≤ $150REFUNDED
> $150HELD FOR A PERSON
Reads the customer’s charges for an order. Tells a temporary card hold (it drops off in 3-5 business days) from a real double charge, and refunds the double charge under the same $150 limit.
3 Handoffs to a person
The AI reads every ticket for signs it needs a person, and a fixed rule checks the same signs, so a ticket the AI misreads still reaches one. Handoffs that depend on the order (a lost parcel, an address change before shipping, a refund over $150, a defect after 90 days) are decided from the order record, not by the AI.
What is a point of resolution worth to you?
Your numbers, one formula. The free check measures your real starting point.
tickets × (after − today) × cost per ticket a person handles
= 2,000 × (50% − 40%) × $5.00
$1,000per month
$12,000 per year · 200 tickets a month fewer for your team
How we measured
The same method we run on your tickets. Every verdict on this page is published, so you can check any of them.
- 356 new tickets in each fresh round, six topics, one fictional store. Refunds, shipping, account, billing, bugs and how-to, written from the store’s policy and order records. On your help desk we use your own exported tickets.
- A typical first deployment as the baseline. A bot that looks up a stale help center by keyword, with a fair default prompt: how most vendor bots start before anyone tunes them.
- Two graders, both must pass. Code checks catch stale facts, wrong numbers and wrong order numbers; an AI grader checks the answer against policy. One fail and the ticket fails.
- Every fix tested on tickets nobody had seen. Three fresh rounds, each on 356 new tickets, after an initial test. Same AI model and graders throughout, every round published (how we got here: every round). Final round: 0 wrong answers in 261 resolved tickets; at 95% confidence, the true rate is under 1.5%.
Fix My Support Bot: this scorecard, run on your tickets.
Support Rescue is the demo; Fix My Support Bot is the service. Your AI support bot answers correctly, hands off to a person at the right time, and stops embarrassing you. Measured on your real tickets, the way this page measures Alder & Ash.
$1,500
Quick fix
1 week. Your top 10 failures found on your tickets and fixed.
$3,500
Full rescue
2 weeks. Everything this demo shows, on your help desk.
What the full rescue includes
- A test set built from your own tickets, scored like this page: real resolution next to the dashboard number, false resolutions, missed escalations
- Knowledge-base cleanup: duplicates merged, stale and conflicting articles fixed, gaps filled from failed tickets
- Actions on your systems where the bot needs order or billing data, with a person approving wherever money moves
- Routing rules, so tickets that need a person reach one
- Also checked: a policy-answer test, an attack test, an EU AI disclosure check and the cost per answer
Start here
- Free 48-hour AI check$0Send your bot’s link or a ticket export with names removed. You get a one-page report within 48 hours on your top 3 problems. No call needed.
Other ways to work with us
- Build an AI Feature$3,500No bot yet? A support bot shipped in 2 weeks, hitting one agreed number. You pay when the agreed metric passes.
- AI on call$1,500/moAfter the fix: weekly tests on real traffic, alerts when a model update breaks something, a monthly cost and quality report. Priority $2,500/mo. Cancel any month.
Guarantee: the resolution rate goes up on your own tickets, or we keep working free.
Or write to us at hello@khlabs.dev. We reply within 24 hours. All offers