When AI makes requests cheaper, who pays to answer them?

Scope

AI-assisted requests, with Australian public services as a focus and explicitly bounded international evidence

Confidence

Moderate on the accounting framework; low on the size and balance of real-world effects

This is a provisional assessment of an open question. It distinguishes the available evidence from Alex’s established submitted positions. It does not attribute a settled new policy position to him.

When AI makes requests cheaper, who pays to answer them, and how much of the extra demand is a benefit we previously failed to deliver?

AI can make asking easier without making answering easier. A person saves time drafting a complaint, application or appeal. Someone else still has to establish the facts, find the records, make a decision and explain it. Other people may wait longer.

But some of that extra work is work we should have been doing already. A valid entitlement left unclaimed, a mistake left unchallenged or information left inaccessible can make an organisation look efficient while leaving people worse off.

The answer to “who pays?” is therefore broader than the organisation receiving the request. The harder question is how much useful access the extra work buys. Current evidence cannot give us a reliable percentage.

Follow the work beyond the send button

Consider a hypothetical complaint. AI saves the sender twenty minutes. It adds ten minutes of checking for the recipient and creates no other effects. There is still a net time saving, although the employer and employee capturing it may differ.

Now suppose it adds thirty minutes of checking, generates a follow-up and delays another customer. The sender’s productivity gain has become a cost elsewhere. If it corrects a serious error, that cost might nevertheless be worthwhile.

Four groups can pay: the sender through subscriptions and verification; the recipient through processing and rework; other users through queues; and customers, taxpayers or owners through the eventual funding bill. Staff can also absorb pressure through unpaid effort or displaced work.

Counting generated documents or minutes saved at the drafting stage misses most of this account.

Extra demand contains different things

A useful assessment separates at least four possibilities:

• A valid request that previously never happened because the person lacked time, confidence, language skills or knowledge.
• An existing request made clearer and more complete, reducing clarification and improving the decision.
• Extra words, duplicates or unsupported assertions that add checking without improving the outcome.
• Fraud, harassment or deliberate attempts to overwhelm a service.

These can coexist, including within one submission. Poor wording does not prove a weak claim. Polished wording does not prove a strong one.

A successful claim also needs careful accounting. Paying an eligible person an entitlement raises government expenditure, but the payment itself is a transfer to that person. It is not all consumed as administrative waste. Processing uses resources; the payment has distributional and budget consequences; its value depends on what the policy is intended to achieve. Counting the payment as a social loss while ignoring the recipient would prejudge the answer.

What the evidence actually establishes

Characterizing Agentic Flooding of Government Services collects 84 selected cases across 11 jurisdictions where officials or other sources attributed pressure to AI. Most involved people submitting AI-generated text. The authors describe observed impacts as moderate and explicitly say their method cannot establish prevalence or AI’s causal contribution. Their risk framework and proposed responses are mainly analysis, rather than measured intervention effects. It is a useful warning about a mechanism, not an estimate of society’s net loss.

I have not independently audited the linked case dataset. I also do not rely on the paper’s Australian fee example: its descriptions conflict, and the parliamentary bill record’s indexed status raises a further discrepancy that I could not resolve from the full page.

Australian evidence also cautions against blaming every backlog on AI. The ANAO’s May 2026 audit found administration only partly effective in three departments, with shortcomings in assistance, recordkeeping and disclosure. It reported 43,456 Commonwealth FOI requests in 2024–25, up 25%. Those figures do not identify how much, if any, of the growth AI caused.

There is evidence pointing towards possible benefits too. A study by Ruiyu Zhang and Lin Nie, using vignettes with 220 citizens and 214 civil servants in China, found improvements in perceived communication quality from AI modification. That supports testing clearer communication. It does not establish faster real-world case resolution or transfer automatically to Australian services.

The OAIC’s guidance also distinguishes a person using AI to make an FOI request from a request made without a person or corporate entity. Treating all AI-assisted requests as equivalent would miss that distinction.

The strongest objection is capacity

Even valuable new demand can arrive faster than a service can expand. A person exercising a legitimate right can unintentionally delay someone with an equally legitimate, more urgent need. Better access at the front door does not guarantee better access to a completed service.

That is a serious reason to plan capacity and prioritisation. It is also a reason to test responses carefully. Blanket fees or additional paperwork might reduce volume while screening out people with the least money or time. A shorter queue alone cannot tell us whether that was an improvement.

Possible alternatives include clearer eligibility information, structured intake with room for unusual cases, help gathering evidence, duplicate handling, better published answers and assistance for receiving staff. Each needs testing for errors, exclusion and total effort. Automating both sides could otherwise produce a faster exchange of messages with little improvement in resolution.

Measure useful resolutions, including the missing ones

A practical trial would compare AI assistance with the existing process, preferably through randomised offers of assistance or a phased rollout with a credible comparison group. Existing access should remain available.

Measure:

• Valid issues resolved, including cases that would otherwise never have been submitted.
• Total applicant and staff time, including checking, follow-ups and appeals.
• Errors, reversals, unresolved cases and delays imposed on other users.
• Who gains access and who drops out, particularly people facing existing barriers.
• Technology and implementation costs, alongside any capacity added.

Submission counts are insufficient. Following eligible people who never submit helps reveal the missing benefit. Following the whole queue helps reveal displaced costs. Avoid counting the same time saving twice, or treating an appeal as both a new success and an unqualified first-stage success.

For SMEs and students, the practical lesson is immediate: use AI to make the recipient’s task easier. Check the facts, state the decision needed, include relevant evidence and remove repetition. Assess accuracy and resolution, rather than word volume.

For Australia, overseas ownership of the tool does not settle the balance either. Imported AI can still create domestic value. The assessment needs domestic gains, resource costs and payments flowing overseas, with their distribution made explicit.

Cheaper requests are worth pursuing when they help resolve more worthwhile problems at an acceptable whole-system cost. How much extra demand meets that test remains an empirical question. The right target is useful access delivered, with the work of delivering it counted honestly.

What would change this answer

Credible comparisons measuring useful resolutions, previously unmet need, total applicant and staff effort, errors and queue effects would change this assessment. Evidence of better access at a lower total cost would support wider use. Evidence of displacement or exclusion without worthwhile gains would support changing the design. The hypothetical time examples and accounting above are analysis, not study results.

Related questions

The question connects to Australia’s net benefits, productivity and affordability and what students should learn. Each requires us to look beyond the volume of assisted output.