Each scenario is a complete recipe: the pain point, the tool stack, roughly how long it takes, and what you end up with. Follow the links from a recipe to the underlying courses and measured articles when you want the mechanics.
A scenario can serve several tracks, so the total exceeds the scenario count.
Sensitive data cannot go to a cloud API, yet you still need AI to process it — a seeming dead end.
A local OpenAI-compatible endpoint your existing code can switch to by changing one base URL, with model and quantisation guidance.
An hour-long call still needs write-up afterwards, and shifts in management wording are easy to miss.
A structured note covering guidance changes, shifts in management tone, key Q&A exchanges, and the delta versus last quarter.
After the research is done you still spend a full day on layout — the same content handled twice.
A clean deck with one argument per slide, automatic chart placement, and sources plus speaker notes.
Digging numbers out of an annual report by hand takes two hours per company and still produces transcription errors.
Structured JSON for revenue, gross margin, cash flow, and segment data — each field carrying its source page number.
A pitch book means template-wrangling, number-filling, and alignment — two or three all-nighters spent mostly on layout.
A structurally complete draft: market overview, comparables, valuation range, and proposed deal structure.
Manually refreshing HKEX filings and broker notes daily — 90% of the time just confirming nothing changed.
A push summary only on material change, carrying what changed, impact assessment, and a link to the source.
Reading contracts page by page for key clauses is slow, and non-lawyers miss risk points.
A clause table covering term, payment, breach, IP, and liability — unusual clauses flagged with their location in the source text.
Validating an idea waits on engineering capacity, and by the time it ships the market has moved.
A shareable prototype with basic auth, a database, and the core flow — usable by real testers.
A DCF model takes half a day to build, every assumption change forces a rebuild, and the sensitivity table is manual hell.
A working DCF model with three-stage cash flows, WACC, terminal value, plus a sensitivity matrix and an explicit assumption list.
Human translation is slow and stylistically inconsistent; machine translation breaks structure and leaves source-language residue.
Batch translation plus automated quality checks: aligned section structure, intact code blocks, consistent numbers, and source-language residue detection.
Data moves between systems by manual copy-paste — things get missed, get wrong, and cannot be traced.
A schedulable, retryable automation that alerts on failure, with LLM nodes inserted for judgement and rewriting.
Every AI tool needs its own integration rebuilt, and internal systems stay unusable by agents.
One implementation usable across tools, with permission controls and optimised tool descriptions.
A single agent loses focus on long tasks, blows its context window, and delivers unstable quality.
A lead-dispatches, sub-agents-execute-in-isolation, results-merge architecture, with write-scope partitioning and conflict avoidance.
Every prompt revision is declared an improvement on vibes, with no data to prove real progress.
Your own question set and scoring script, run identically on every change, reporting accuracy, coverage, and cost deltas.
Nobody writes up the meeting, action items go untracked, and the next meeting rehashes the same points.
Automatic post-meeting notes with extracted action items and owners, pushed to the relevant people and tracked.
Two hours every week stitching data from several systems into a report, in a format that keeps drifting.
Scheduled data pulls into a fixed template, producing a report with anomalies flagged and pushing it out.
Outreach emails are generic and get low reply rates, yet researching each prospect properly takes too long.
A pre-meeting brief per lead covering company news, inferred pain points, angles to open with, and a personalised opener.
Output cannot keep up with channel demand, every piece starts from scratch, and the style drifts.
A topic-to-draft pipeline with consistent style, automatic per-platform rewriting, and a human review gate built in.
Hundreds of resumes per role — screening is slow, standards drift, and bias creeps in.
A candidate list ranked against explicit criteria, each with a match rationale and suggested interview questions — the decision stays human.
Competitors quietly change prices or ship features, and by the time you notice a batch of customers is gone.
Periodic comparison of competitor pricing pages and changelogs, pushing a diff summary only on material change.
You want to judge quickly whether a company deserves deeper research, but do not know where to start or when you have enough.
A one-page summary covering business, ownership, related-party transactions, risk signals, and what to check next.
Collecting multiples for a dozen peers by hand gives inconsistent formats, and every refresh starts over.
A consistently formatted comparables table with PE, PB, and EV/EBITDA, flagging definitional differences between companies.
Staff cannot find policy and process documents, so they keep asking colleagues the same questions dozens of times.
Q&A directly inside your chat app, answers citing document and paragraph, with unmatched questions tracked to close documentation gaps.
With cross-holdings inside a group, the same shares get counted twice and the effective stake comes out wrong.
A shareholding topology with layer-by-layer dilution math, double-count checkpoints, and traceable evidence for every step.
Reported profit looks great while cash flow disagrees — nearly impossible to spot by hand across dozens of pages.
A red-flag list covering receivables-turnover anomalies, inventory build-up, and operating-cash-flow versus net-profit divergence, each with supporting numbers.
Support answers the same questions daily, new hires ramp slowly, and knowledge is scattered across documents.
A support bot that answers from internal documents, escalates to a human when unsure, and reports hit-rate statistics.
You inherit undocumented legacy code where one change breaks three things, so nobody dares touch the core logic.
An architecture map and risk list first, then small refactoring steps, each covered by tests and a revertable commit.
AI conclusions look professional, but the numbers may be invented — and the cost of being wrong is severe.
Layered verification: coverage and accuracy reported separately, automatic abstention at low confidence, and every conclusion traceable to its source.
You used a framework but do not understand the mechanics, so when retrieval quality drops you do not know what to tune.
You implement chunking, embedding, retrieval, reranking, and generation yourself, plus a measurable retrieval-quality evaluation.
It works locally and fails in production: cost runs away, errors go unnoticed, and nothing is traceable.
A deployment with health checks, usage and cost monitoring, failure alerts, traceable logs, and a degradation strategy.
We do investment research, corporate due diligence, and AI adoption consulting. This site is where we publish our methodology — free of charge, because a method that can be inspected is the only kind worth trusting.
Start learning