Workforce Economics & ROI
Budget for the Work AI Creates
Measure review capacity and completed work before turning AI time savings into an investment claim. Connect AI investment decisions to the capacity needed to review, accept and deliver useful work.
PsychFlo Editorial
AI coding tools are increasing the supply of work faster than organisations can complete its review. A working paper updated on 14 November 2026 by Fiona Chen and James Stratton, covering 655 software firms, documents the gap: coding activity rises substantially, but finished software grows much less. Code spends longer waiting for review and longer being reviewed.
That finding changes the investment question. Executives need to ask how much useful work the organisation can accept, not simply how much material a tool can produce.
This distinction matters as executives set their next AI budgets. In McKinsey’s August 2026 survey, 80% of respondents reported improved individual productivity, while 37% attributed some enterprise earnings impact to AI. These are self-reported measures of different outcomes, not a calculation of how much productivity disappears. They nevertheless make a useful distinction: a better experience for an individual does not establish a financial result for the firm.
The software research is a working paper, not a universal estimate of AI's effects. Its practical importance is the mechanism it examines: improvements in one stage of production can expose a constraint elsewhere. That gives managers a specific place to look when local productivity gains fail to reach the business.
For leaders, the practical question is whether an AI investment has funded the work required to turn its output into something the organisation can accept.
Give the receiving team a place in the business case
Most AI pilots have a sponsor and a group of users. They also need an explicit receiving team: the people who verify, approve, implement or explain what the system produces.
For a sales proposal, that might include a commercial reviewer and a pricing owner. For an internal policy summary, it might include the policy specialist who checks whether an apparently helpful simplification changes the meaning. For an agent-generated software change, it includes the people who must decide whether the change belongs in production.
Ask those people to describe an acceptable handoff before the pilot starts. What source material must accompany the output? Which claims need checking? What makes an item too incomplete to review? Who handles a disagreement?
This conversation can reveal a cheaper intervention than recruiting additional reviewers. A shorter output, a narrower task or a mandatory source reference may remove work that a larger model would merely produce faster.
Keep an effort account and a delivery account
A pilot should maintain two separate views of performance.
The first is total effort: preparation, generation, verification, correction and implementation. Include the work done outside the adopting team. Treat time estimates as estimates and validate a sample against actual work.
The second is delivery: how many outputs meet the agreed quality standard and reach their intended destination, and how long that takes. Waiting time belongs here even when nobody is actively working on an item.
Separating the two prevents a common misunderstanding. A workflow can use fewer labour hours while delivering more slowly because items are waiting for a scarce approver. It can also deliver faster while consuming more labour, which may be a sensible choice for a time-sensitive service. Leaders need to see which trade-off they are making.
Chen and Stratton report a 56.1% increase in the time from submission for review to approval after firms adopt AI agents. Their analysis distinguishes waiting in a backlog from time spent in active review. Both increase. These are estimates from the studied software firms; they should not be presented as a forecast for another organisation.
The distinction gives a business case a more useful structure. Record the time people actively spend checking and correcting work, and separately record how long work waits for someone authorised to accept it. Add preparation, implementation and operating expenditure. Compare equivalent outputs against a baseline with the same quality requirements.
An improvement in drafting time is one input to that account. Treating it as an immediate reduction in labour cost skips the decisions that would turn available capacity into a financial outcome.
Set a review budget before increasing volume
A review budget specifies who will do the receiving work, how much capacity they can commit and what happens when demand exceeds it. It should identify both routine checks and the specialist review needed for unusual cases.
The budget need not be a permanent allocation. It can begin as a small trial agreement. For example: one defined category of report, an agreed number of submissions, a named reviewer, source-linked drafts and a weekly comparison of accepted output, rework and waiting time.
Avoid treating capacity as a single pool of interchangeable hours. The person who can verify a financial claim may not be able to assess a security exception. A nominally available reviewer may have no authority to approve the final action. The constraint is often a specific kind of judgement.
Where possible, make the burden visible through short, aggregate samples rather than continuous individual surveillance. The purpose is to redesign the workflow, not to create another system that employees must spend time feeding.
Use quality to govern the queue
When a backlog appears, the wrong response is to preserve the headline saving by making checks more superficial.
Instead, inspect what enters review. Are requests sufficiently specified? Are repeated failure types being returned to the generating team? Can an item with missing evidence be stopped before it reaches a specialist? Does everyone know which decisions the AI is not authorised to make?
The receiving team should be able to reject an incomplete handoff. Otherwise, the adopting team can improve its own numbers by moving unfinished work across an organisational boundary.
There is also a limit to adding reviewers. If the system repeatedly creates material that nobody needs, more approval capacity only increases the cost of processing it. Reduce unnecessary output before expanding the queue around it.
Make the next investment conditional
Before expanding a pilot, require a joint decision from the team producing the output and the team responsible for accepting it. Agree what evidence would support expansion, revision or a pause. Include a fallback if the specialist reviewer is unavailable.
None of this establishes that more review always creates more value. Low-consequence, reversible tasks may need little checking. The aim is proportionate review with a visible owner, not an elaborate approval process for every sentence.
An AI budget buys more than generation. It changes the demand placed on the people and systems around generation. A credible business case funds that change and measures the result at the point where useful work is actually delivered.