What it can do, how we organise ourselves around it, and what this team becomes. No coding required.
Barely anyone here writes software for a living. That is exactly the point — Codex is for the other work.
None of this is a coding problem. It is a "figure it out and do it" problem — and that is what an agent is for.
You say what you want in plain language. Codex looks at the real files, does the work step by step, checks its own result, and shows you everything before anything is saved.
It reads spreadsheets, PDFs, logs, exports and folders — not just code. You never write a line of anything.
Excel, CSV, PDF, Word, logs, config, screenshots. Hundreds at once, not one paste at a time.
"Which of these 200 devices is out of warranty?" — it goes and looks instead of guessing.
Hands back a finished spreadsheet, document, report or checklist saved to your folder.
Renames, sorts, converts, merges, exports, cleans up — the fiddly multi-step work, done in order.
Runs the result, spots its own mistake, fixes it, and tells you what it changed.
Our naming rules, formats and standards written down once — it follows them every time after.
Reads vendor docs, release notes and the web, then tells you what applies to us specifically.
Hand off a long job, close the laptop, come back to a finished result and a summary.
Same conversation. Completely different amount of work left for you.
Real requests, exactly as you would say them to a colleague:
Excel and CSV exports, PDFs, Word docs, emails saved to a folder, screenshots, log files, config files.
Runs on your own machine in a normal window, or in the browser. Points at any folder you choose.
Long jobs run without you sitting there. Come back to a finished result and a summary of what changed.
Can be connected to approved tools and data sources — one at a time, with IT deciding what is in scope.
Repositories are optional. For most of this team, the "project" is a folder on a shared drive.
A technician asks for the monthly ticket summary they normally spend half a day on. No coding, no template, no formula.
Before: hours of sorting and formatting, once a month, one person.
After: one sentence, then a review.
Illustrative — we confirm real timings on our own data.
The thing people worry about — "will it break something?" — is a setting, not a gamble.
Everyone starts on look, don't touch. We open it up only where it earns trust.
It can look at files and answer. It cannot change or delete anything until you say so.
You see exactly what it wants to write, before it writes it. You approve or reject.
Work happens in a sandboxed folder. Anything outside it needs your explicit permission.
A simple instructions file holds our standards and limits — no secrets, no production, ever.
| OpenAI Codex | Claude Code | |
|---|---|---|
| Who it suits here | Everyone, including non-technical roles | Comfortable terminal users |
| Getting started | App, browser or editor — pick one and talk to it | Terminal-first; more setup before value |
| Non-code work | Strong — documents, data, reports, research | Strong, but framed around codebases |
| Hand-off jobs | Yes — runs in the background, collect the result | Mostly runs with you present |
| What we already pay for | Included with ChatGPT plans under discussion | A separate subscription decision |
Both are capable. Codex simply has the shorter path for a team where 10 of 20 people will never open a terminal. Plans and features change quickly — verify before purchase.
A capable tool with nothing
written down is just
a faster guess.
Everything so far was what Codex can do. The rest is the part that is on us: what we write down, where we put it, and who keeps it true.
Right now most of how this team works lives in three places Codex cannot reach.
Alignment is the real project. The tool is the easy half.
"Ask Minh, he knows the trick." Perfect knowledge, zero availability, walks out the door on holiday.
The answer exists — in a message from March, in a channel nobody searches.
Written once, never updated, quietly wrong. Worse than nothing, because people trust it.
Short, current, plain files next to the work — readable by a person and by the agent.
Rule of thumb: write it where it applies. A rule for everyone goes on shelf 1. A rule for one report goes in that report's folder.
This is the habit that separates a team that gets value from one that gives up in week three.
When it gets something wrong, you don't just fix the output — you say "add that to the rules" and it writes the correction down itself.
Documentation stops being a project nobody has time for, and becomes a by-product of doing the work.
Roughly a day to set up, then an hour a month to keep honest. If we won't do that hour, we should not start.
Twenty people each finding their own way works for a month. Then it quietly becomes twenty different versions of the same job.
Same tool, same people. The difference is entirely in how we organise it.
Rule that keeps it honest: nothing shared without a named owner and a date.
Not a help desk for AI questions. A small group whose job is that everyone else needs them less — measured that way, deliberately.
If Codex can do all that —
what does this team become?
Nobody gets a straight answer by avoiding the question, so here is ours — in three parts.
Nobody's job disappears. The bottom rung gets handed over, and the top rung — the one we never have time for — finally gets done.
Execute the steps by hand. Sort, format, copy, check, repeat. Accurate only if you are fresh.
Judge whether the output is right, sane and safe to send. Requires knowing our business.
Decide how the job should be done, write it down once, make it repeatable by anyone.
Choose which problems matter, what to stop doing, what to fix at the root. Never delegable.
Codex takes level 1 and assists with 2. Levels 3 and 4 are the job — and they are the part that grows careers.
| Role | Handed to Codex the level-1 work |
Stays yours, always judgement & accountability |
What you gain room for level 3 & 4 |
|---|---|---|---|
| Technician · 10 | Retyping notes, repeat fixes, manual checklists | Talking to the person, deciding what's urgent, the final call | Fixing root causes so the ticket never returns |
| Infrastructure · 4 | Diffing configs, scanning logs, chasing what changed | Approving every change, owning uptime and rollback | Designing the standard, then proving we meet it |
| Business Analyst · 3 | Gathering, cleaning and formatting data; waiting on pulls | Knowing what the business actually needs, and why | Asking harder questions, of far more data |
| Managed App · 3 | Reading vendor docs, reconciling lists by hand | The vendor relationship, the risk call, the commitment | Judging vendor claims against our own reality |
| DevOps & AI · 3 | Repetitive build and glue work | Guardrails, access, what touches live systems | Making everyone else need them less |
Same people, same names. The higher-value half of the same job.
The team becomes a team that owns procedures instead of performing them — with the knowledge written down, not carried in four people's heads.
That is a change in what a good day looks like. It is not a change in who is needed.
Treat it like a fast, tireless junior colleague who never minds being asked again — and who always shows their work.
No committees, no waiting list — the whole team gets it, starting in read-only.
Pick the task you dread most this month.
Describe it in one sentence.
Let Codex work it out.
Then write down how it did it — that part is ours.