A fast answer you can't check is just a confident-sounding guess.

"Fast AI answer" still means "go check it by hand."

This wasn't a guess — it came from actually watching people work: 12 people interviewed across three different jobs, 5 watched at their real desks, 14 samples of their actual work, about 420 things they said or did, all written down. Every number on this page is real or clearly labeled as a guess — that rule is what shaped the whole product.

people double-check the AI's answer by hand whenever the stakes are high — true for every single person interviewed

Finding · from interviews · a strong pattern

said one made-up citation was enough to make them stop trusting the tool — permanently, for the whole team

Finding · from real incidents · a strong pattern

separate people showed us a homemade spreadsheet they'd built just to track their documents — everyone does this before they ask anything

Finding · seen firsthand · confirmed

individual things people said or did, written down during interviews and observation, sorted into 8 recurring patterns

"If I have to find the clause anyway to check it, the AI saved me nothing. It just made me do the search twice."

— P7 · compliance operations · discovery interview

The moment this helps most: inheriting someone else's mess.

Due diligence, an audit, taking over a vendor relationship, a handover from someone who just quit — that's the one moment document pain actually has a deadline, a budget, and one person responsible for it. AIDEX starts there, on purpose, instead of trying to be a tool for everything at once.

How AIDEX stacks up against the alternatives (best guess, not measured yet) What matters Best competitor AIDEX's goal Can you check the answer? Other AI tools · Handles a big pile of files well Search tools · Remembers what your team taught it Other AI tools · How fast you see value Generic AI (ChatGPT etc.) ·

Being honest about it: plain ChatGPT-style tools still win on "fastest to try" — AIDEX isn't chasing that. Nobody else combines checkable answers + handling messy piles + actually remembering corrections . That combination is the whole bet.

Not "answers given." Answers that passed the check and nobody had to correct later — because that's the only number that proves every part of the system actually worked.

And it never travels alone: it's always shown next to how often citations were actually correct . If that second number drops while the first climbs, something's wrong — you'd see it immediately, not months later.

A GOAL, NOT YET MEASURED IN THE REAL WORLD

A specific painful moment has a deadline and one person who owns fixing it. A general "AI workspace" has neither — nobody's in a hurry to adopt it. If this doesn't hold up in real pilots, that's the plan to revisit — written down now, not excused away later.

A GUESS — WRITTEN DOWN SO IT CAN BE PROVEN WRONG

◎ The one number that matters most

⌖ Why start with one moment, not a whole platform

That homemade spreadsheet everyone builds? That became the whole product brief.

Every single thing people said or did got written down and grouped into patterns — and each pattern was labeled honestly: Confirmed (I watched it happen), Likely (strong signal, not 100% proven), Guess , or Expected but untested . Four patterns ended up shaping the actual design.

12 people interviewed 5 watched at their real desks 14 real work samples reviewed ~420 things said or done, logged 8 patterns found 15 claims tracked, each labeled

PATTERN 1 · 11 OF 12 PEOPLE · STRONG

The first three days of any new pile of documents: someone quietly builds a spreadsheet just about the documents — nine people showed us theirs. The map AIDEX builds automatically replaces that habit.

PATTERN 2 · 10 OF 12 PEOPLE · STRONG

"I read the contracts. I skim the decks. I pray about the appendices." What gets attention is what's riskiest if it's wrong — so AIDEX sorts documents by risk, not by topic.

"The rate table IS the contract. The prose is decoration." Pasting a document into a plain chat box flattens tables into mush — exactly the part that mattered most.

"My review is me redoing her citations. Two hours, every memo." People double-check by hand because there's no way to trace where a claim actually came from.

These three people aren't real individuals — they're realistic composites built by blending patterns across the 12 people actually interviewed, labeled Likely, not literal so nobody mistakes them for real research subjects.

Research Lead at a consulting firm · built from 4 real interviews

"We have a process. The process is me."

Her day-to-day

Juggles 2–3 client projects at once, each one starting with a huge folder of someone else's files

Stays up till 11pm double-checking her team's sources by hand

What would win her over — or lose her instantly

Wins her: the first time checking a memo takes 15 minutes instead of 2 hours

Loses her: just one citation that turns out to be wrong. That's all it takes.

◎ What's at stake for her: client trust — one wrong number can end a relationship

Compliance Manager at a fintech company · built from 4 real interviews

"Show me where it says that." — his catchphrase became the product's core feature

His day-to-day

60–90 new documents land on his desk every month, more during an audit

Keeps his own master spreadsheet of every obligation — because a spreadsheet doesn't silently change on him

What would win him over — or lose him instantly

Wins him: an audit question answered in minutes, with proof he can download and hand over

Loses him: an answer that quietly skipped a whole folder of documents without saying so

◎ What's at stake for him: regulators — small errors turn into official findings

Knowledge Ops Lead at a product company · built from 4 real interviews

"The company forgets faster than I can write things down."

Her day-to-day

She's basically the company's memory, in human form; already gave up on two other AI tools once their quality slipped

Will patiently teach a tool a correction — exactly once

What would win her over — or lose her instantly

Wins her: a correction she made two weeks ago shows up again, applied correctly, with her name on it

Loses her: having to teach the same fix twice. "I'll train it once. I won't ask again."

◎ What's at stake for her: her own credibility, and whether the team's knowledge survives her leaving

Everyone builds a "shadow" tracker first.

People read out of fear, not relevance.

The meaning is in the layout, not just the words.

"Reviewing" someone's work usually means redoing it.

Three fictional stand-ins built from real answers

Maya

Daniel

Sofia

Every decision, plus the option that got turned down.

Sixteen decisions are written down in a public log — each one paired with the option that got rejected, and why. Five of them shaped the product the most. Most case studies only show you the choice that won; this one keeps the ones that lost, too.

Best guess, not yet proven People will accept being told "no, I can't verify this" instead of just going back to a regular AI tool that always gives an answer. The target: at least 7 in 10 people accept the block in real testing. If most people don't, this whole approach gets revisited — that's decided now, not excused away later.

TRACKED CLAIM · WATCHED CLOSELY GOING FORWARD

Decision · Keep two separate kinds of writing

Mixing the two is how case studies end up written for nobody in particular.

A bigger, harder problem makes a stronger proof of the same skills — and the honesty-first approach carries over just fine to higher-stakes work.

Two roles at every stage — one that does the work, one that checks it and can block it. That rule never got cut, no matter how much else changed.

"Built using this process, and runs on this process" is a stronger point made once than the same point made twice, separately.

The story ends exactly where you can click a button and try the real thing — not a screenshot pretending to be one.

The riskiest guess, said out loud

Docs for the product, and docs for this page — never mixed together

AIDEX became a document-intelligence tool, not a narrower case-study assistant

The "can refuse to ship" rule survived every single redesign

The same 6-step process both built this project and runs inside the finished app

The case study leads directly to the working app

Strategy just means picking what you're willing to give up.

Every real decision costs you something else. Here are the three biggest tradeoffs — each one made on purpose, with the cost written down plainly.

Making the AI pause and double-check itself before answering

The instant, always-confident feel of a regular chatbot

WHY · People already told us the real pain is redoing the AI's work by hand afterward. A confident-but-wrong answer costs more time than a slightly slower, checkable one.

Building for one specific painful moment — inheriting someone else's messy files

Being a broad, do-everything "AI workspace"

WHY · A specific painful moment has a deadline and a person who owns it. A general-purpose tool has neither, so nobody's in a rush to actually use it.

Twelve small AI specialists, each doing one job well

One general AI assistant — simpler to build and easier to explain

WHY · A specialist can be genuinely confident about one narrow thing. A generalist sounds confident about everything and is right about less of it.

The 6-step process that built this product.

I call it SIGNAL — six steps, done in order, every time. Each step pairs something AI is fast at with a decision only a human should make. Open The SIGNAL Process panel to explore the inputs, outputs, AI helpers, and quality checks for each step.

The usual way this gets built: 12 weeks → Following these 6 steps: 3.5 weeks 3.4× faster · nothing skipped, same rigor

"The process isn't the product — it's how the product got made. Every step got faster without getting shallower. AI didn't replace my judgment here; it just let me use it more often, faster."

Same steps, way less waiting around

Which AI tool did what.

No tool was used just because it's trendy — each one earned its spot for a specific job. Claude was involved at every single step: sorting research, writing the AI role specs, writing the code, and running the review passes that got this page ready to publish.

EVERY STEP

SENSE + LEARN

SENSE

GENERATE

NAVIGATE

ASSEMBLE

The AI tools gathered research, sorted it into patterns, drafted ideas, and tried to poke holes in them. Every real decision on this page — what to build first, what the AI is and isn't allowed to do, every tradeoff — stayed a human call. AI made the work faster; it didn't make the judgment calls.

After every round of building, there was a deliberate "try to find what's wrong with this" pass — checking the design, the writing, mobile screens, accessibility, and consistency. Fix what's found, then check again. This very page went through that loop too.

Claude

NotebookLM

Perplexity

v0

Figma AI

Claude Code

Who did what

Checking the work, on repeat

You don't hand over a picture of software anymore. You hand over the software.

When AI can build the real thing almost as fast as you can sketch a fake version of it, the whole job shifts. Here are seven ways design work changed on this project.

Every layout drawn by hand, one at a time

A small rulebook that generates the screens itself

AIDEX doesn't have a fixed set of screens to keep in sync. It has one simple structure (document → claim → source), six actions a person can take, and one consistent shape for every answer. Design that rulebook once, and thousands of screen states come out consistent automatically — even the ones nobody thought to draw.

A person clicks through steps a designer planned out

People and AI helpers hand work back and forth

AIDEX's flow map shows twelve AI helpers passing work between each other: one reads the files, another double-checks the read, a third writes the answer, a fourth can send it all back. What actually needed designing was the handoff — who's allowed to act on their own, who has to check with a person first, and who's allowed to say no.

The exact same interface for every user, every time

The interface changes based on how sure the AI actually is

An AIDEX answer isn't a fixed template — it reshapes itself around the evidence. A confident answer looks different from a shaky one; a refusal uses that same familiar shape, just with the first line flipped to "here's what I couldn't confirm." The screen reflects the evidence, not a fixed mockup.

Rough sketches promising what the software might look like later

The "sketch" is already the real, running app

The design phase of this project didn't produce a single clickable mockup — it produced the actual app you can open from this page right now. Real data, real memory, real refusals. When the sketch already works, usability problems just get fixed on the spot instead of turning into a to-do list for later.

Specs thrown over a wall for engineers to build from

A live link, not a set of instructions to build one

There's still a handoff folder in the project files — but it documents something that's already running, not something someone still has to build. Load speed is measured on the real thing, not requested as a target. The deliverable is a working link; the history of how it was built lives in version control, not in sticky notes.

Guidelines that go stale the moment nobody enforces them

Rules built into the software so they can't be quietly ignored

AIDEX's rulebook for its AI helpers isn't a suggestion someone can skip under deadline pressure — the app itself throws out any answer that breaks the rules before it ever reaches the screen. A rule that's actually enforced by the software never goes stale, because skipping it isn't an option.

Designs the surface, then waits for a team to build it

Researches, decides, designs, builds, ships, and improves it — solo

One person, directing a set of AI helpers: about 420 things learned from research, 16 decisions logged, 12 AI roles specified, 76 planning documents, a real working app, actually deployed. The job isn't "a designer who also codes" — it's closer to someone who designs the whole system that then produces the product. This page is the evidence.

Twelve AI helpers. Six simple actions. One of them can say no.

Nobody using the app needs to know or care that there are twelve separate AI helpers behind it — research showed people just want the outcome, not an org chart. So you never see "agents." You see six simple actions, easy to learn in one sitting, and every single response follows the same predictable shape.

Six actions — this is the entire toolkit a person uses

jump to the exact source of any answer

correct it so it remembers next time

Every single answer follows this same 6-part shape

1 THE ANSWER what it's claiming, in plain words 2 THE PROOF the exact quote it's based on — or a clear note that it's the AI's own reasoning, not a quote 3 HOW SURE a plain-language confidence level, tap to see why 4 WHAT IT CHECKED searched · not searched · couldn't be read 5 WHAT IT DOESN'T PROVE the limits, stated honestly 6 WHAT'S NEXT the actions you can take from here

The house rules every AI helper has to follow

Every AI helper follows the same rulebook: always use that 6-part answer shape · quote the source or clearly say it didn't · describe confidence in words, never a fake-precise number like "92%" · never change its memory without a person confirming first · never leave a blocked answer as a dead end — always suggest the smallest fix · and keep one consistent tone: exact, calm, never full of itself.

These aren't just guidelines someone can ignore under a deadline — the app itself throws out any response that breaks a rule before it's ever shown to you. Even a refusal follows the same 6-part shape, just flipped: "here's what I can't confirm" → what it did find → exactly where it falls short → the easiest fixes, ranked.

Meet the twelve — one does the work, one double-checks it, at every stage

"There's no button anywhere that lets a broken answer through. Not one. That rule is the whole product."

— the one hard rule the checking AI never breaks, no exceptions

Built like a real company would — by a team of one.

Behind this page is a public folder of work: research, strategy, AI role specs, and product plans, each one checked before anything got built on top of it. This page is just the easiest-to-read summary of that folder.

◆ No pre-built toolkits, built by hand — plain HTML (the structure of a web page), CSS (the styling — colors, spacing, layout), and JavaScript (the code that makes it interactive), so the page loads fast with nothing extra weighing it down.

◆ Runs entirely in your browser — the sample documents, the six actions, and the seven-check review all run right on your device. It even works offline.

◆ Light and dark mode, both designed on purpose — dark mode isn't just light mode with the colors flipped; both were designed properly, and your choice is remembered everywhere.

◆ Built to work for everyone — usable with a keyboard alone, readable by screen readers, respects reduced-motion settings, and text stays readable in both themes.

aidex-v2/ — 76 documents · nothing skipped a check ├── 00-foundation/ what this is, and how success is measured ├── 01-research/ findings + every claim tracked ├── 02-strategy/ where to start · the one metric that matters · principles ├── 03-information-architecture/ the four screens, mapped out ├── 04-ux/ the six actions · flows · accessibility ├── 05-ai-system/ the rulebook · a job description for each of the 12 AI helpers ├── 06-design-system/ reusable colors, buttons, spacing rules · 07-visual-language/ the overall look and feel ├── 08-engineering-handoff/ notes for whoever builds it next · 09-storytelling/ early drafts of this page's story ├── 10-case-study/ the blueprint for this very page ├── 11-product-docs/ requirements · risks · release notes └── _ops/ decision log · glossary · progress tracker

A few technical choices that mattered

A few claims from the tracked list, word for word

Not "faster answers." Answers you can actually check.

The product, in one breath

Teams inherit piles of hundreds of files, then have to answer important questions using nothing but keyword search and memory. Regular AI made this worse, not better — it sounds confident, but you can't verify it exactly when it matters most. AIDEX flips that trade: it's fine being a little slower, as long as every answer can be checked.

The app never just says "it's in one of the documents." It points to the exact sentence — click it, and you jump straight to the highlighted line, in under a second.

Almost everyone builds a messy spreadsheet by hand just to understand what they're working with. AIDEX builds that map for you automatically, before it answers a single question.

FOUND IN RESEARCH · 11 OF 12 PEOPLE DID THIS

Before anything gets exported, a built-in checker double-checks every answer seven different ways. If the proof doesn't hold up, it blocks the export and explains how to fix it. Not a warning you can ignore — an actual stop.

Built using the same 6-step process the finished app runs on internally

Every answer, one click from its proof

A map of the pile, before any question

It's allowed to say "I won't ship this"

Four screens. One simple rule underneath.

Map, Sessions, Artifacts, Memory — four screens that walk you through the same 6-step process without ever naming it. The one rule that makes all of it work: every answer breaks down into small claims, and every claim points to its exact source. That's what makes one-click checking possible.

How it'll be measured, before what the numbers say.

AIDEX is still an early-stage concept. So instead of showing off a cherry-picked number, here's the actual plan for measuring it — including exactly what result would prove each claim wrong. One claim below is already proven. It's listed first, on purpose.

planning documents in the public project folder, each one checked

decisions logged — including the option that got turned down each time

claims tracked, each one labeled honestly

faster than the usual way this gets built — without cutting corners

Every claim below is labeled — and every guess says exactly what would prove it wrong Claim Label How it's checked / what would prove it wrong This page was made using the exact process it describes. Verified Go look for yourself: the public project folder has 76 checked documents, a 16-entry decision log, and 15 tracked claims. See claim C7. AIDEX cuts the time to understand a document pile from days to one sitting. Expected, not tested yet Will be tested with real pilot teams on their own documents. Proven wrong if people still build their own spreadsheet anyway because it's faster. Being able to check answers actually stops people from re-checking everything by hand. Expected, not tested yet Will be measured directly with real usage. Proven wrong if people still redo important answers by hand anyway. People accept a blocked answer instead of just going around it. Best guess, unproven Target: at least 7 in 10 people accept it, in real testing. Proven wrong if most people give up and go around it instead — then this whole approach gets revised. Honestly labeling claims (like this page does) makes people trust early-stage work more. Best guess, unproven Will be tested by comparing a labeled version against an unlabeled one. Proven wrong if the labeled version actually scores lower. This page is the labeled version.