Can AI Reduce Delivery Costs?

The problem, in numbers

Professional services sell time and judgement, and time is getting dearer. In 2024-25, Australia's professional, scientific and technical services industry grew its sales and service income by 3.7%, but wages and salaries rose 4.6%. Operating profit before tax fell 11.4%, which the Australian Bureau of Statistics attributes to labour and operating costs outpacing revenue.^{Source 1} The industry employs about 1.34 million people. When labour is the main input and its cost rises faster than fees, any tool that cuts the hours in a deliverable gets attention.

AI is that tool on paper. The Productivity Commission expects AI to lift multifactor productivity by more than 2.3% across the economy over the next decade, while stressing considerable uncertainty about the size of the effect.^{Source 5} The question for a firm is narrower: does that productivity show up as a lower cost to deliver an engagement?

What improves, and why

The work AI does best is the work professional firms have traditionally pushed down to junior staff: first drafts, research summaries, slide narratives, document comparison and data extraction. Those are the hours that make up the base of the leverage pyramid.

The vendor estimates point the same way. Clio's analysis of legal billing data suggests up to 74% of hourly billable tasks could be automated, but the exposure is uneven: 57% of lawyers' tasks against 81% for legal secretaries and administrative assistants.^{Source 6} That is a model of exposure, not an observed saving, yet it shows where the hours sit.

The saving does not fall to zero, because someone still has to check the output. The cost moves from producing the draft to reviewing it, and review is done by more senior, more expensive people. A firm that removes junior hours but adds partner review time may cut total hours and still lift cost per hour.

What the evidence actually says

The best evidence comes from two randomised experiments. In a Science study of 453 college-educated professionals doing occupation-specific writing tasks, those given ChatGPT took 40% less time and produced work rated 18% higher in quality.^{Source 2} In a field experiment with 758 Boston Consulting Group consultants, those using GPT-4 on tasks inside the AI frontier completed 12.2% more tasks, 25.1% faster, with results 40% higher in quality.^{Source 3}

The same BCG study holds the warning. On a business problem-solving task the tool handled badly, consultants using GPT-4 performed 23% worse than those working without it.^{Source 4} Gains were also uneven by person: on the creative task, lower-baseline consultants improved 43% while top performers improved 17%. AI compresses the gap between junior and senior output, which is good for cost, but only where the task sits inside what the model does well.

What the experiments do not measure is the cost of an engagement end to end. They time individual tasks, not the scoping, client meetings, review cycles and rework that make up a real job. Firm-level evidence is mostly survey data. Thomson Reuters' 2025 survey of 2,275 professionals found respondents expect AI to save 5 hours a week within a year, but only 22% said their organisation had a visible, defined AI strategy.^{Source 7} Expectations are not savings.

Billing decides where any saving lands. Under hourly billing, fewer hours means less revenue unless rates rise, so the efficiency can be absorbed rather than passed on. Clio reports that law firms now use flat fees for 34% more cases than in 2016, and that 71% of clients prefer a flat fee for an entire case.^{Source 6} Under a fixed fee, a faster delivery becomes margin for the firm; competition then decides whether clients see lower prices.

Where it falls short

• Review is not free. Every hour of drafting saved needs some senior checking. If review is skipped, the saving is borrowed against future rework or liability.

• The frontier is jagged. On tasks outside the model's competence, AI users did worse, not better. Firms cannot assume a saving on one task carries to the next.

• Task gains are not engagement gains. The strongest studies time discrete tasks. There is little controlled evidence on whole-engagement cost or on client prices.

• The training pipeline thins. Junior hours are where professionals learn. Cutting them saves money now and may cost review capacity later.

Where governance decides the outcome

The Australian case that made this concrete was a review for the Department of Employment and Workplace Relations, worth just under $440,000. It contained references to reports that do not exist and a made-up quote attributed to a court judgment, and Deloitte agreed to repay the final contract instalment.^{Source 8} The department said the substance and recommendations stood, but the lesson is about cost as much as safety: an unreviewed AI draft turns a delivery saving into a refund and a reputational loss. Firms need a named reviewer accountable for every source and figure, disclosure to clients of where AI was used, and engagement terms that say who carries the risk.

How to think about adoption

Start with the tasks the evidence supports: drafting, summarising and structured writing, where controlled studies show large time savings. Budget the review hours explicitly and assign them to people senior enough to catch what the model gets wrong. Measure cost per engagement, not hours saved per task, because that is the number that tells you whether delivery actually got cheaper. Decide early how the saving will be shared: under hourly billing it can quietly disappear, and under fixed fees it becomes margin or a price advantage. Finally, protect some junior work for training. The firms that save most on delivery will still need people who can tell a good draft from a plausible one.