TRIZAI

中文 Talk to us

Your team
already uses AI.
Reviewing it
is now the work.

Triza AI is an AI evals studio. We use Triza Loop so one job your team does every day can go to AI and stay off your desk — like bringing a new colleague up to speed, the loop reviews it every week and improves it every week, until you stop checking every one.

Tell us about one job See how it works

QUALITY CONTROL✦ ONE JOB AT A TIME✦ AI AGENTS✦ WORKFLOW AUTOMATION✦ EVALS✦ RAG KNOWLEDGE BASE✦ FITS DAILY WORK✦ WEEKLY ITERATION✦ HUMANS STAY IN CONTROL✦ QUALITY CONTROL✦ ONE JOB AT A TIME✦ AI AGENTS✦ WORKFLOW AUTOMATION✦ EVALS✦ RAG KNOWLEDGE BASE✦ FITS DAILY WORK✦ WEEKLY ITERATION✦ HUMANS STAY IN CONTROL✦

Sound familiar?

Most AI transformations die between the demo and the daily work

Not because AI isn’t smart enough, but because nobody brought it up to speed: it never learned how your team works, and nobody checked what it produces on a bad day.

Getting AI to the point where it is genuinely useful — that is our work.

A familiar story

Stunning in the demo. Unused six months later.

The AI that wows a boardroom and the AI that gets the job done on an ordinary Tuesday afternoon are two different things. We build for that Tuesday.

Right on Monday. Wrong on Tuesday.

AI is not like normal software. Ask the same question twice and you can get two answers. Old project rules cannot handle that. Our method can.

Three colleagues reviewing AI eval results together on screen

Nobody wants to be the person who trusted it wrongly.

Teams will not hand real work to something they cannot verify. We give them a ruler they can measure with themselves.

You bought a roadmap, not a result.

Six-month plans and hundred-page decks do not get AI into daily work. Small steps, a weekly review and a weekly improvement do.

New AI models keep launching. Which one? When to switch?

You do not need to keep up. Every time a new model ships, the system runs the review again, adds what it can now do to your workflow, and tells you what changed.

How it works

Triza Loop: one review a week, so AI keeps getting closer to how you work.

Hire a new colleague and on day one they do not know how your company does things. You guide them, they get things wrong, you correct them, and a few weeks later they work the way you want. AI is the same.

Triza Loop is our evals cycle. An eval is a review: the system builds the test out of your real work, not out of demo questions. The AI sits that test every week, the system improves it every week, and every week it fits your way of working a little better.

Put plainly, what Triza Loop does is quality control on AI output. It turns that into a system, instead of leaving it to whoever has a spare hour this week.

You do not have to take our word for it. You can watch it get better week by week. Here is what happens when you work with us. Every step hands you something you can hold, show your boss, and keep.

One jobPick the job

Start with one daily job, not the whole company.

You pick one job your team does every day, and the system writes it up on a single page: what the AI will do, who it does it for, and what it must never touch. Nothing moves until you agree with that page.

You get: a one-page agreement anyone can read.
The hard part: choosing. Most companies get this wrong the first time. They pick the job they most want to talk about, not the one that hurts most. Pick wrong and three months later you have a system nobody uses, and no idea which step broke.

An honest answerFind out what AI can really do

Get a straight answer before you spend on building.

The system tries the job with today’s AI, using your real tasks, not demo prompts. Some parts it can do, some it cannot do yet, and some it should never handle alone.

You get: a clear answer for that job — yes, no, or not yet.
The hard part: writing the test. Test with demo cases and the AI always passes, then you find out on launch day that it does not work. You have to use the cases your team genuinely could not handle last month, and picking those cases takes someone who understands your business.

Who is in controlMap the route

Where AI runs alone, where it needs help, and where a person holds control.

The system maps the route through the whole job. AI takes the parts it is good at. The rest is supported by helper tools or signed off by one of your people. Nothing risky runs unwatched.

You get: a one-page workflow anyone in the company can follow.
The hard part: drawing the line. Which step goes to AI and which step needs a human signature. Move that line one notch and it becomes a complaint or a payout. There is no universal answer; every industry and every regulator is different.

The standardWrite the test

Turn your real work into a review the AI has to pass every week.

Your team’s real emails, real forms and real cases become the test. Pass or fail, no grey area. Your team can re-run it any time, with us or without us.

You get: an AI acceptance standard that belongs to your company.
The hard part: five to ten cases is enough. The hard part is choosing which five. Choose wrong and your AI will fail on your most expensive kind of case, and your test will not tell you.

Real useRoll out small, watch, improve

A small group uses it on real work. The system watches, fixes, and rolls out again. Every week.

A few of your colleagues start really using it, with the boundaries written down. When it goes wrong, that mistake becomes the next thing the loop fixes.

You get: a working pilot, plus a short weekly report on what changed and what got closer.
The hard part: telling the difference. After launch you will get a pile of feedback, and most of it is not real failure — it is unfamiliarity. Treat “unfamiliar” as “wrong” and fix it, and you will make the system worse every week.

And then the loop continues.

Every time the loop runs, the AI gets a little closer to how you work. What it does reliably, the loop hands it more of. What it still gets wrong stays with your people. When a new AI model ships, the system runs the review again, adds what it can now do, and takes away the support you no longer need.

So your AI gets more trustworthy and better fitted to daily work over time, instead of heavier.

Working with us

One loop a week. You always know where things stand.

Every week

A short report

What worked, what failed, and what the loop changed.

Your time

A few hours a week

From the people who actually do the job, not a steering committee.

Speed

Usable in one to two weeks

And every week after that it fits your way of working a little better.

Who runs it

Our system runs the loop

The loop runs on our system, so it does not wait for someone to be free, or for your timezone to overlap with ours. The test and the reports stay with you, and your team can re-run them any time.

projects our team members worked on

companies and public bodies

industries

0–2

weeks to a usable version

Where it goes

This is not a one-off project. It is your team doing it themselves a year from now.

Starting from the first job, this is how it usually goes.

Month 1

The first job is running

One daily job already has AI doing part of it, used by a small group, with the boundaries written down. The first time you see this AI genuinely doing work inside your company.

Month 3

Your team runs the loop

That job runs steadily. Your team runs the review itself and decides what should change. The second job begins.

Month 12

AI is part of the day

Several daily jobs have AI inside them, each with its own acceptance standard. When a new model ships, you run the review yourselves, without waiting for us.

The finish line is not “the company installed AI”. The finish line is a team that can hand the next job to AI, and knows when not to.

Use cases

Your job may already be on the list

These eight are the most common daily jobs, and the easiest ones to start a first loop with. They are not client case studies. They are types of work.

Open any one and you see three things: what AI can do, what a person holds on to, and what the first loop uses to test it.

Insurance

Claims document intake and triage

Every claim means someone opening several documents to check what is missing and which type it is. Simple and complex cases sit in the same queue.

What AI does: Reads the documents, flags what is missing, sorts by type, and says what it based that on.
What a person holds: Payout amounts, declines and any grey-area case go to a person. The AI never decides whether to pay.
What the first loop tests: Your ten hardest cases from last quarter, including the badly scanned ones and the ones where the category itself was arguable.

Insurance

Agent training answers and policy lookup

An agent asks about policy wording, waits for head office, and hears back a day later. By then the customer has gone.

What AI does: Answers questions that have a wording basis, citing the source for every line so the agent can check it.
What a person holds: Anything touching underwriting or a payout for an individual customer goes to a person. The AI says what the wording states, not whether this one is approved.
What the first loop tests: The questions agents actually asked before, and got wrong answers to.

Banking

First drafts of customer enquiry replies

Dozens of enquiries a day, each written by hand. You reply slower than your competitor, and customers do not wait.

What AI does: Writes the draft and marks which policy, and which paragraph, it drew on.
What a person holds: Every reply is read by a person before it goes out. Anything that changes an account or an amount needs a signature.
What the first loop tests: Your ten hardest emails from last month, including complaints and the ones that ask three things at once.

Professional services

First drafts of quotes and proposals

Every proposal starts from nothing. Your most senior person spends half a day fixing formatting and digging through old files.

What AI does: Builds a draft from past versions, matches the format, and lists which old documents it drew on.
What a person holds: Prices, hours and commitments are filled in and signed by a person. The AI does not touch what you promise a client.
What the first loop tests: The last few you won, plus the last few you lost, to see whether the draft catches the difference.

Retail & consumer goods

Product descriptions and SEO content

Every new product needs a description. Write slowly and it misses the shelf; write fast and every product reads the same.

What AI does: Produces several angles from the specification, so someone chooses rather than accepts the one answer.
What a person holds: Brand language, efficacy claims and compliance wording stay with a person. Getting a claim wrong costs more than being slow.
What the first loop tests: Your best seller and your hardest-to-describe product, to see whether it writes a claim you are not allowed to make.

Logistics & engineering

Work order summaries and exception flags

Field work orders are each written differently, so someone reads every one before knowing which has a problem.

What AI does: Summarises each order, flags exceptions and gaps, and moves the ones needing a person to the top.
What a person holds: Downtime, safety and extra-cost calls all go to a person. The AI only gets the right order in front of them faster.
What the first loop tests: The orders where something actually went wrong and nobody caught it at the time.

Public services

Form checks and rejection explanations

A form comes back rejected, the letter reads like officialese, the resubmission is wrong again, and the same case goes round three times.

What AI does: Checks whether the form is complete and writes in plain language which field to change and how.
What a person holds: Approval and any discretion stay with a person. The AI only makes the rejection letter understandable.
What the first loop tests: Your highest-rejection form types, plus the cases that went back and forth more than twice.

General

Meeting notes into action items

Nobody writes it up, actions scatter across everyone's notes, and the next meeting discovers nobody did them.

What AI does: Pulls out actions, owners and deadlines, marking which part of the notes each came from.
What a person holds: Where ownership is disputed, the AI must leave it blank for a person to confirm, never guess.
What the first loop tests: One set of notes from a meeting where people genuinely disagreed, to see whether it guesses an owner.

Yours not on the list? Then it is worth a conversation. Send us a short email about the job and we will answer honestly — whether AI can do it, and what the first loop would look like.

What you get

Not a deck. A system your team can run on its own.

After the first few loops you own three things.

AI already inside daily work

Running in your company, with it written down what the AI handles and when it hands back to a person.

An AI acceptance standard that belongs to your company

Re-run it any time: a new model, a changed process, a new colleague. You always know whether it still passes and still fits how you work.

A team that knows how to keep iterating

You have watched the whole loop run. The next loop, you can run yourselves.

Together these three are the start of your own in-house capability to hand work to AI, not a vendor you have to depend on.

Is this you?

If this sounds like you, it is built for you

You have tried AI, but it never stuck.
One repetitive daily job eats your team’s whole week: customer emails, claims, quotes, reports, checks.
You cannot let AI near customers or money until it is proven safe.
You would rather have one thing genuinely working than a roadmap listing ten.

Who we are

A Hong Kong team that does one thing: quality control for AI.

We use AI on our own work every day, so we know what holds up in real use and what only holds up in a demo.

Our team has over a decade of enterprise systems experience in Hong Kong, across insurance, banking, public services and transport. In recent years we have gone all in on AI.

We work fully remotely, with clients in several timezones. Eval reports and handover documents are written to be read on your own time, not in ours.

Get one job
genuinely right.

You do not have to rebuild the whole company at once. Pick one job your team does every day, where mistakes are expensive, and where nobody really trusts the current way of doing it. That is enough to start the first loop.

Talk to us hello@triza.ai

Send us a short email: what the job is, and why your team is not comfortable handing it to AI yet. We will answer honestly — whether AI can do it, and what the first loop would look like.

Your team already uses AI. Reviewing it is now the work.