Every credible AI rollout keeps a person reviewing the output before it goes to a customer, a regulator or a client's file. That is the right call for almost every business use case that matters. What gets missed in the planning stage is that the review time itself has a real, ongoing cost, and if the business does not budget for it honestly upfront, the project's return on investment case quietly deteriorates over the following year without anyone noticing until someone finally adds up the hours.
The budgeting mistake that shows up six months in
The typical pattern looks like this: a business models the time saving from AI-drafted content or decisions based on a fully automated workflow, then bolts on human review as an afterthought, assuming it will take a few minutes per item. In practice, reviewing AI output properly, actually reading it, checking it against source data, catching the occasional error, takes real cognitive effort, and for anything client-facing or compliance-relevant it should take real effort. A business that models five minutes of review time per item and actually needs fifteen has quietly cut its projected return on investment by a meaningful margin, and that gap compounds across every item processed.
We have seen this directly with a Sydney professional services client automating first-draft client correspondence. The original business case assumed 3 minutes of partner review time per drafted letter, projecting a payback period of about five months against a $14,000 setup cost. Actual review time, once partners were doing it properly rather than rubber-stamping, ran closer to 9 minutes for anything non-routine. The project still paid for itself, but the honest payback period was closer to eleven months, and if the business had budgeted for that from the outset, nobody would have felt misled when the early results looked slower than promised.
What honest budgeting for review time looks like
Time-test the actual review process with real staff on real output before finalising the business case, not a guess
Separate 'routine, low-risk' items that need a light check from 'complex or client-facing' items that need a proper read
Build review time into the ongoing cost model as a permanent line, not a temporary ramp-up cost that disappears
Revisit the review time assumption after three months of live use, since it typically drops as reviewers build pattern recognition
The good news in that last point is real: review time does fall over the first few months as staff develop a feel for where AI output tends to go wrong and where it is reliably fine. A reviewer who has seen two hundred drafted client emails knows within the first sentence whether this one needs a close read or a quick skim, which a reviewer on day one cannot yet judge. Budgeting for the higher early-stage review time, rather than the steady-state figure a vendor might quote, is what keeps the business case honest without becoming needlessly conservative.
The AUD conversation worth having upfront
There is a second, less obvious cost that shows up when review time is underbudgeted: reviewer fatigue and the resulting drop in review quality. When staff are given far less time than the task genuinely needs, they do not do a worse job on purpose, but the review inevitably becomes shallower, catching the obvious errors and missing the subtle ones. That is a worse outcome than not automating at all in some cases, because it creates a false sense of security that a human has checked the work properly when in practice they were skimming under time pressure. Budgeting review time honestly is as much a quality control measure as a cost measure.
None of this is an argument against human-in-the-loop review, which remains the right design for almost any AI workflow touching a customer, a client file or a regulatory obligation. It is an argument for budgeting it as what it actually is: a real, ongoing cost that deserves the same scrutiny as the software subscription line, not a footnote assumed away to make a business case look better than it will turn out to be.
For a business evaluating an AI rollout in the $10,000 to $40,000 setup range, the practical move is to run a two to four week pilot on a small, real sample, timing actual review effort rather than estimating it, before committing to the full build. That pilot cost, typically a few thousand dollars of setup and a couple of weeks of staff time, is cheap insurance against building a business case on an assumption that turns out to be wrong by a factor of three, which is common enough in our experience that we now insist on it before quoting a fixed price on anything involving ongoing human review at scale.
