Demos
Support RescueBot CheckAgent GatewayCost CutterLLM Bill BenchmarkSync Watchdog Get a free AI checkWe make AI support bots give the right answers, cut what your AI costs, and build the AI features you planned. Every result is measured on your own data. Fixed price, one to two weeks.
Every fix and build comes with a written guarantee, and the scope and price are agreed in writing before we start. Not sure which fits? The free AI check tells you, with a one-page report within 48 hours.
Fix My Support Bot
$1,500quick fix · 1 week · your top 10 failures
$3,500full rescue · 2 weeks
We measured it: a three-line rule cut invented policies from 20% to 0.7% (5 models, 54 trap questions).
Resolution goes up on your own tickets, or we keep working free.
What’s insidepolicy-answer test · attack test · EU AI Act disclosure check · cost per answer
Same work, two triggers: the bill is too high, or the model is being retired.
Cut Your AI Bill
$1,900about 1 week
Up to$6,722a month
Low end $2,486/mo if only the easy part of each lever applies
Up to $80,659 a year. At the low end the $1,900 fee pays back in 23 days.
An estimate with the Bill Checker’s own assumptions (caching, batch and, if one big model, routing; the savings stack). “Up to” is every lever at its high end. Get the exact number free: try the Bill Checker
If verified 30-day savings come in below the fee, we refund the difference.
What’s insideBill Checker savings plan · caching · batch · model routing · cheaper provider · prompt trimming · output caps · code shipped
Switch Models Safely
$1,500one model, one use case
$3,500multi-agent or voice
From our model-switching study. Report
Quality on the agreed test set doesn’t drop, or we keep working free.
What’s insidetest set from your data · model shoot-out · bill cut · code switch · you keep the test set
Build something new: a one-week build for small teams, or a two-week sprint to one agreed number.
Build an AI Feature
$3,5002-week sprint
$7,500+production
What you get
You pay when the agreed metric passes.
What’s insideevals on your data · cost budget · attack test · the Agent Gateway on any action the AI takes
Small Build
$1,500about 1 week
AI that works inside the tools your team already uses, built with n8n, Make or code.
Example builds
You pay when it works on the agreed test cases.
What’s insidetest cases agreed up front · the assistant or automation, built and tested · you keep the code and tests
Also sold on their own: tests and a gateway for the AI you already run.
AI Security Test
from $2,900
We attack your chatbot ($2,900), a bot that answers from your documents ($4,900) or an agent ($7,500) and write it up for your enterprise questionnaire. One free retest within 30 days.
See the Bot Check demoEU AI Act disclosure check
$490
We check that your bot tells customers they are talking to an AI, as Article 50 of the EU AI Act requires, including right after a handoff to a named agent.
See the HandoffBench studyAgent Gateway
from $750
A gateway for one risky action your agent takes, in shadow mode first, then enforcing.
See it hold a refundAI on call · after any project
Weekly tests on real traffic, alerts when a model update breaks something, spend caps and usage alerts, and a monthly cost and quality report. Cancel any month.
$1,500/mo
$2,500/mo priority
Fixed scope and price in writing before we start. Read-only access, never production keys.
We agree the one number that matters and measure where it stands today, on your data.
The number, what changed and what is next. You can rerun the eval yourself.
The agreed acceptance test passes, then we invoice. You keep the code, the tests and the report.
We invoice only once the agreed acceptance test passes.
Bill cut: if verified 30-day savings are below the fee, we refund the difference.
NDA before any data is shared. Data deleted within 7 days of delivery.
One invoice in USD, paid by card or PayPal, due 14 days after the test passes.
For our first 3 clients, in exchange for a case study you approve before it is published. Same scope, same guarantees.
AI mistakes cost real companies money and trust. Six public incidents, none involving KH Labs, and the test that would have caught each.
Public incidentsNot KH Labs clients. Quotes and figures are from the linked sources.
“A remarkable submission.”
The tribunal, on Air Canada’s claim that its chatbot was a separate legal entity
Ordered to pay CA$812.02
Civil Resolution Tribunal, B.C. · Feb 2024
It told a grieving customer he could claim a bereavement fare after buying. The airline’s own policy said otherwise.
CA$812.02 ordered paid. The tribunal rejected Air Canada’s argument that its chatbot was a separate legal entity.
Would have caught itFix My Support Bot · policy-answer test
CBC NewsMoffatt v. Air Canada, 2024 BCCRT 149Same failure in our demo
“That’s a deal, and that’s a legally binding offer – no takesies backsies.”
The dealership’s chatbot, as reported by Business Insider
$1 for a new Tahoe
Business Insider · Dec 2023
A user told a Chevrolet dealer’s ChatGPT-powered bot to agree with everything, then offered $1 for a new Tahoe. The bot accepted and called it binding.
The deal was never honoured, but the bot was pulled after the screenshots went viral.
Would have caught itFix My Support Bot · attack test
Business InsiderAI Incident Database #622Our attack test demo
“The cost to the victim can be over $46,000 per day.”
Sysdig Threat Research, on LLMjacking
Stolen cloud keys
Sysdig Threat Research · May 2024
Attackers used stolen cloud credentials to resell access to hosted AI models, while the account owner paid for the usage.
Sysdig estimated the cost to a victim at over $46,000 a day at full quota.
Would have limited itAI on call · spend caps and usage alerts
“An error occurred after a system update yesterday.”
DPD’s statement, quoted by The Register
AI switched off
The Register, BBC · Jan 2024
After a system update, the parcel firm’s chatbot swore on request and wrote a poem calling DPD the worst delivery firm.
DPD disabled the AI element of its chat and said it was being updated.
Would have caught itFix My Support Bot · attack test, rerun after every update
“We have no such policy.”
Michael Truell, Cursor co-founder, quoted by The Register
Invented policy
The Register, Ars Technica · Apr 2025
Asked about logouts, the AI email agent blamed a one-device rule that did not exist. The real cause was a backend change.
A co-founder confirmed in public there was no such policy, and the user was refunded.
Would have caught itFix My Support Bot · graded on real tickets
“Yes, you can take a cut of your worker’s tips.”
NYC’s MyCity chatbot, tested by The Markup
Against city law
The Markup · Mar 2024
The city’s official MyCity chatbot told owners they could take workers’ tips, go cash-free and refuse housing vouchers.
Each answer contradicted city or state law. The city kept it online at first.
Would have caught itBuild an AI Feature · tested against the rules before launch
A tribunal made Air Canada pay for a refund rule its chatbot invented. Inside Fix My Support Bot we attack your bot first: here is Harbor Bank, a demo bank, with the same model before and after our fixes. Pick an attack.
What it costs you:
12 → 0of 40 attacks got through, before and after our fix.
Support Rescue demo
Wrong answers cost you customers. After the fix, real resolution went from 22.5% to 73.3% on 356 tickets the bot never saw, with 0 wrong answers in 261 resolved tickets. How we measured: a demo store and real model runs, every answer checked by code and by an independent AI grader, in three rounds that each used new tickets.
Open the Support Rescue demoBars: all 356 tickets the bot never saw (final test, round 3 of 3).
Is shipping free on my $62 order?
Yes, free shipping on all US orders over $50.
No: free over $75, so standard shipping is $6.95.
Will my two-year-old gift card still work?
Gift cards expire after 12 months, so it’s no longer valid.
Yes. Alder & Ash gift cards never expire.
I was charged twice for my order.
I can’t process billing, so I’m handing this to a person.
One charge posted; the second is a pending hold that drops off.
How long does standard shipping to the UK take?
Usually 7 to 14 business days, depending on customs.
Standard shipping to the UK takes 4–7 business days.
Can I return my beanie for a refund?
I’ve connected you with our team to send a return label.
Not eligible: it was delivered outside the 30-day window.
Six pieces of work: studies, fresh-data tests and a free tool. Every number comes from the data behind it, with data and code published where a link is shown.
Data: HandoffBench, 1,335 labelled replies
Research paper · HandoffBench
Data: MigrationBench, 80 support e-mails
Technical report · MigrationBench
Data: CostBench, accuracy next to each model. *Estimated tokens.
Cost study · CostBench
Data: Support Rescue demo, final test
Fresh-data test · Support Rescue
Data: Bill Checker sample export, savings a month
Free tool · Bill Checker
Data: PolicyBench, demo store policy
Research · PolicyBench
The Agent Gateway sits between your agent and its tools. It comes with every Build an AI Feature project that takes actions, and is sold on its own from $750. Week one runs in shadow mode and shows what it would have stopped. Then you switch on enforce.
The rules come from the demo’s policy v2: the refund and delete rules of the run above, plus limits on customer exports and outside email. Refunds use invoice INV-5047 from the recorded run.
billing.issue_refund {"invoice_id": "INV-5047", "amount_usd": 500}
Press “Send tool call”.
Each attempt is checked in order and written to the log below.
Agent Gateway demo · shadow mode, one week
$60,153
in 9 refunds that enforce mode would have blocked, out of 26 blocked and 25 held in 208 tool calls.
Open the shadow-mode reportSend your bot’s link, your monthly AI bill export or your agent’s tool list. In 48 hours you get a one-page report: the three biggest problems, what each one costs you, and how we would fix them.
What you get backexample format · demo data
Who we are
KH Labs is a founder-led AI engineering company. We scope each project, write the code, run the tests and write the report you get at the end.
What we have published
We reply within 24 hours. No call is needed: the free check and the scope happen by email. If you want a call, we will find a time.
Yes: Switch Models Safely. We test 2–3 replacements against your current model on 100–300 of your real inputs, switch to the winner and hand you the test set. $1,500 for one model and use case, $3,500 for multi-agent or voice. If quality on the agreed test set drops, we keep working free until it matches.
Small Build ($1,500, about a week) is one workflow, bot or automation, paid when it passes the agreed test cases. Build an AI Feature ($3,500, two weeks) ships a feature into your product against one agreed number, with evals, a cost budget, an attack test and the Agent Gateway.
No. Read-only usage exports and sample prompts, with personal data removed.
Yes, before any data is shared. NDA and DPA on request.
Read-only keys, in your cloud or a sandbox, on data with personal details removed. Deleted within 7 days of delivery, confirmed in writing. Security and data handling
Yes. The AI Security Test starts at $2,900 for a chatbot ($4,900 for a bot that answers from your documents, $7,500 for an agent), with one free retest within 30 days. The Agent Gateway is also sold on its own (from $750), and the EU AI Act disclosure check costs $490.
If verified 30-day savings are less than the $1,900 fee, we refund the difference. Verified saving = (old cost per request − new cost per request) × requests served.
Tell us in one line: what does your AI do today, and what’s going wrong (or what do you want to build)? We reply within 24 hours.