In a customer case study dated September 1, 2026, OpenAI says Australian law firm Gilbert + Tobin used Codex to check and rename 1,100 files for upload into another system. For office managers and small-business owners facing occasional batches of paperwork, it is an example of AI carrying out several steps with tools, rather than only drafting a reply.
The same account describes preparing audit reports for 300 entities, with people responsible for reviewing and approving the work. These are vendor-published customer claims, not independently verified results. The account does not disclose enough about errors, checking time or usage charges to predict what another business would save.
A practical way to decide whether an agent is worth trying is to require four things: defined inputs, permitted data, a checkable output and someone responsible for accepting it. The deciding factor is how you will establish correctness, not whether the job sounds clerical.
Start with the job, not the product
Use a simpler tool when the task follows one fixed rule and your existing software already handles it. A bulk-renaming command or spreadsheet operation may be enough. A short rewrite may need only ordinary chat.
Consider an agent when a substantial one-off job combines several steps or varied source materials, but the output can still be checked. Preparing a batch for import or assembling a referenced draft report are candidates.
Keep the work human-led when nobody can assess the result, the data is not approved for the tool, or the task requires a professional judgment rather than document preparation. A prepared report is not permission to sign, file or act on its conclusions.
OpenAI’s current product guidance distinguishes Chat for questions and short drafts, Work for completed deliverables, and Codex for developer-oriented tools and detail. Office work does not require choosing the coding interface. Feature access depends on the plan, platform, region, rollout and workspace settings.
If your files already live in Dropbox, our guide to its AI file-task access and approval options can help you assess that existing workspace before adding another tool.
Decide what would count as correct
OpenAI’s Work getting-started guide recommends specifying the outcome, sources, constraints and review point. For a file-preparation job, turn that advice into an acceptance check before execution.
For example, work on a copy of an approved folder. Specify the naming rule and destination, ask for a list pairing every old filename with its proposed new name, and require missing identifiers, duplicate names and unreadable files to be listed as exceptions. Do not let the agent guess how to resolve them. This is a suggested workflow, not a reconstruction of Gilbert + Tobin’s setup.
Account for every input with either an output or an exception. Compare the new names with authoritative identifiers; matching file counts alone cannot prove that names belong to the right documents. If only filenames should change, use a file-content comparison to confirm that the contents stayed unchanged. Review ambiguous cases individually before importing anything.
For report preparation, require source references for important statements, dates and totals, plus a list of unresolved gaps. A reviewer who understands the subject should check the conclusions against those sources. A polished document without traceable support is not a finished result.
Check both the data access and the proposed actions
Use an organisation-approved workspace and limit access to the material needed. Keeping a file on your laptop does not mean its contents stay there: OpenAI’s local-security documentation says relevant excerpts, screenshots and tool results may be sent to its services. Confirm that the intended data use is permitted before starting.
Action controls answer a different question. OpenAI’s permissions guide separates the boundary on accessible resources from approval of actions. Neither establishes that an output is accurate. For an initial job, keep preparation separate from sending, deleting, uploading or changing a live business system, and preserve the originals. Check which controls your chosen tool actually provides.
Count the checking time and actual usage
Compare staff effort for the same accepted result, including preparation, supervision, review and corrections. As a hypothetical example, suppose manual work takes 120 staff-minutes. If the agent-assisted route needs 15 minutes of preparation, five of supervision, 30 of review and 10 of corrections, the time released is 120 − (15 + 5 + 30 + 10) = 60 minutes. Those are illustrative assumptions, not measured savings.
Track machine elapsed time separately, especially if a deadline matters. An hour released for other work is not automatically an hour removed from payroll.
As checked on November 11, 2026, OpenAI’s pricing documentation says Work and Codex share usage, which varies with the model, task complexity, context and tools. Neither a prompt count nor API token prices establish how many file batches a subscription will cover.
Record usage alongside any additional charges. For eligible credit-based agreements, OpenAI distinguishes consuming committed credits from a new invoice charge; overage depends on the agreement and settings. Agree a stopping point before an exploratory task expands.
Start with a small, representative batch and a named reviewer. Expand only if the checked output is useful and the total effort justifies it. If verification takes as much work as doing the job directly, choose the simpler method.