AI Coding Agents · Live Demonstration

AI coding agents discover, organize, and analyze enterprise data — on the fly

From a chaotic Google Drive to board-ready people-operations insights in one conversation — no pre-built pipelines, schemas, or warehouses.

Presenter Anand S Agent Codex · GPT-5.4 Date 16 July 2026

Anonymized

The story in one line

Four plain-English prompts turned 45,547 scattered files into board-ready insights — by discovering, organizing, then analyzing on the fly

45,547 files

1,011 auto-clusters

25 Q1 datasets

49 files · 1 chosen

3 insight stories

1

Discover

Wrote its own crawler, inventoried every Drive file, self-clustered them — nothing pre-catalogued.

2

Organize

Read the workbooks and built the right data model for the question. No tables pre-created.

3

Analyze

Proposed the questions worth asking, then delivered non-obvious, actionable findings with evidence.

Every step is code the agent wrote during the conversation — so the whole path is repeatable, auditable, and re-runnable in under a minute.

02 / 08

Prompt 1 · Discover

A single prompt produced a self-organizing catalog of every accessible Drive file

What I prompted

"Using gws drive explore datasets accessible to me, with metadata. Write a repeatable script that updates incrementally — only new / modified files — and stores the output in a structured datasets.json."

Codex wrote inventory_drive_datasets.py, then ran it — resumable, checkpointing into .datasets-state.json after each page.
It chose what metadata mattered — ownership, recency, type, access path — and grouped files into datasets itself.

First run: Drive crawl, API-bound (one-time) Cached re-run: ~7.2s

What Codex produced — one file record in datasets.json

{ "name": "demo-sites-mgn", "mime_type": "…google-apps.spreadsheet", "size": 87892, "modified_time": "2026-03-24T08:37:26Z", "owners": [ { "email": "…@domain.com" } ], "accessible_via": [ "shared_with_me" ], "web_view_link": "https://docs.google.com/…" } // + a "datasets" array grouping files into clusters

03 / 08

Prompt 2 · Score & rank

"Which are promising?" made the agent invent its own relevance formula and rank the top datasets

What I prompted

"Add a relevance score using these rules of thumb — recency, size, transactional vs. static, project richness. A dataset is a logical collection, not one row. Re-run without re-fetching."

The weighting Codex chose itself

Recency

0.30

Transactional

0.25

Size proxy

0.20

Project richness

0.15

Data density

0.10

Recomputed from cache — --skip-refresh, no second Drive call. A later filter (domain · 2026) narrowed 1,011 clusters to 25.

One scored dataset in datasets_detailed.json

{ "label": "employee + exit mastersheet", "relevance_score": 82.7, "score_breakdown": { "recency": 100.0, "transactional_signal": 41.0, "data_density": 100.0, "size_proxy": 61.7 }, "suggested_analyses": [ { "analysis": "Build an offer→joining funnel…", "why_useful": "Reveals where attrition happens" } ] }

04 / 08

Organize · Schema on the fly

The agent read 49 messy workbooks and built the right data model itself — choosing the structure that fit the question

49 files · 5 clusters · no common schema

KYC Janemp id · status
doj · leave

Payroll Febname · PF
bank · ctc

F&F inputlwd · notice
shortfall

Exit masteremp id
resignation

Onboardingdoj · location
recruiter

SPI trackeremp id · dept
cost code

Different headers, different grain, different layouts in every file.

Codex weighed candidate structures …

One flat tableloses the process steps

Time-serieswrong grain — these are people, not metrics

Employee-grain process graph ✓joins files on emp id across intake → admin → exit

Join keys it inferred: emp id 97 doj 94 location 88 dept 46 Profiled all 49 books in ~33s

The key point: nobody handed it a schema. It picked emp id as the spine and modelled an employee journey — a different question would have led it to a different structure.

05 / 08

The model it built · a knowledge graph

Modelled as a knowledge graph, each employee is one node that links all their records across every workbook

Employee — join key emp id Process record (a file) Shared attribute = a row that links on emp id IDs & names anonymized

No table was pre-built. The agent stitched 97 sheets into this graph on emp id — so one query can walk an employee from onboarding through payroll to exit, and shared nodes (Client 1, Hyderabad) reveal where delays cluster.

06 / 08

The insights · in plain English

Three findings a leader can act on this week — the real problems were hidden behind duplicate, out-of-date spreadsheets

Hiring

The hold-up isn't recruiting — it's paperwork after the offer.

Two teams keep the same hiring list twice (132 identical records). Most delay comes after the hire is decided — laptops, IDs, setup — and just 4 recruiters and the Client 1 & Client 2 accounts drive most of it.

Fix one tracker; run a 2-week sprint on those pockets.

Payroll

KYC isn't blocking pay — it's just a stale dashboard.

People were paid and set up while the KYC sheet still showed them missing — like a departures board that hasn't refreshed. The real risk is joiners who start too close to the payroll cut-off.

Stop trusting KYC as the readiness view; flag late joiners.

Exit

The work is happening — the tracker just went dark.

The official exit sheet stopped updating in October, yet 21 settlements kept moving in side files. Final settlements take a median 41 days, and most cases are special exceptions, not standard exits.

One live queue; separate exception cases with their own owners.

Each focused analysis ran in seconds once built — hiring ~2s · payroll ~3s · exit ~11s — and every claim traces back to the source workbooks.

07 / 08

The takeaway for CIOs

AI coding agents collapse data discovery, modelling, and analysis from weeks to minutes — repeatably

No pipeline tax

No warehouse or schema built in advance. The agent built exactly the structure each question needed.

Asks the right questions

It surfaced the analyses worth doing and found non-obvious, actionable findings.

Idea to insight, today

Four prompts. Under an hour the first time, under a minute every time after.

The question is no longer "can we analyze this data?" — it's "which conversation do we want to have first?"

08 / 08