By Eric X.D. Tan · 29 October 2026

AI fluency is now an operating skill.

Build fluency around a real workflow

A producer knows where a handover breaks. An account director knows why a weekly report takes half the afternoon. A finance manager knows which missing details keep an invoice moving between inboxes.

That knowledge gives a creative team a useful starting point for working with tools such as Codex or Claude Code. The person closest to the work can explain the inputs, exceptions and standard an output needs to meet.

For a creative business, I would define fluency through three abilities: brief the job clearly, check the result against its source, and document a process a colleague can repeat.

OpenAI’s Codex workflow guidance describes defining outcomes, success conditions and boundaries. Anthropic’s Claude Code guidance emphasises giving the tool a way to verify its output. These are useful working habits to bring into an operational pilot. OpenAI workflow guidance · Claude Code verification guidance

Start with a small reporting tool, a checker for missing campaign details or a draft that follows an approved template. Give the workflow an owner who understands the work and can judge the result.

Finance can check the numbers. Operations can examine the handover. The team can then decide whether the result is useful enough to repeat.

Calculate the workload behind the cost of waiting

Consider a hypothetical team of 20 people. Each spends 30 minutes a week on the same kind of recurring task. Across 48 working weeks, that amounts to 480 staff hours a year.

The calculation is 20 × 0.5 hours × 48 weeks. It measures workload to investigate, rather than savings already achieved. A pilot must establish what can actually be removed after setup, review, corrections and ongoing maintenance.

Waiting also delays learning. A team that has never tested a real workflow has little evidence about where the tool helps, how much checking it requires or which exceptions need human judgement.

A small test can leave behind useful instructions, example inputs and a record of what went wrong. A colleague can use those lessons in the next attempt.

For a CFO, the commercial question follows the evidence. Does the improvement reduce paid rework or overtime, create capacity for additional billable work or change a hiring decision? What benefit remains after tool and operating costs?

Recovered time creates capacity. Its effect on EBITDA depends on what changes in the business and the costs of achieving that change. A faster first draft alone cannot settle that question.

Use one weekly report as the first test

Book 30 minutes with the person who prepares a weekly campaign report. Bring a sample export and an example of an accepted report. The timebox is for the experiment; it is not a promise that a finished workflow will be ready in 30 minutes.

Ask Codex or Claude Code to prepare a draft in the same structure, trace every figure to its source, flag missing or contradictory data, and work from copies while leaving the originals unchanged.

Compare the result with the accepted report. Check totals, omissions and any added interpretation. Record the time needed to set up the task, inspect the output and make corrections.

The result should tell you whether this particular workflow deserves another test. A useful trial produces both an output and evidence about the effort needed to trust it.

If it helps, save the instructions and ask a colleague to repeat the process. If checking takes longer than doing the task, narrow the scope or choose another task.

Use the test to decide what to improve next. One workflow, one owner and one result you can examine is a practical starting point.

For a practical starting framework, read SWAY’s guide to AI workflow automation for creative agencies. For a live operating problem, explore SWAY’s AI workflow optimisation service.

Which recurring task deserves less of your team’s talent?

Sources and scope

The 480-hour example is a hypothetical workload calculation, not a client result or an estimate of recoverable savings. The reporting task is a proposed experiment.