[Government](https://bfcs.ai/industry/government/) · [Cheaper](https://bfcs.ai/cheaper/)

# Can AI Reduce Administrative Costs?

Governments handle claims and transactions in the hundreds of millions a year, so even a small saving per document adds up. The case for lower unit cost is strongest where the documents are high in volume and stable in format. The evidence that agencies have actually banked those savings is much thinner, because the agencies we found had not measured it.

Updated **October 2026** · 7 min read · Evidence-graded · 8 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Moderate confidence | Helps on routine drafting and sorting; editing still needed · In the Australian trial 61% of matched survey respondents (n=330) agreed Copilot improved the quality of their work, but editing was almost always needed, and UK departmental staff reported hallucinations. |
| Faster | Partly · Moderate confidence | Self-reported time savings on a narrow set of tasks · Summarising, first drafts and searching show the largest reported savings. These are participant estimates, and one UK department found some tasks took longer with the tool. |
| Cheaper | Partly · Low confidence | Plausible on stable, high-volume documents; rarely measured · Treasury estimates a Copilot licence pays for itself if an APS6 redirects about 13 minutes a week to higher-value work, but did not measure time saved. The national auditor found IP Australia had not substantiated a link between its AI tools and examination efficiency. |
| Safer | Conditional · Governance-gated | Only with registers, impact assessments and evaluation · Commonwealth policy now requires use case registers, impact assessments and ongoing monitoring. Audits of the ATO and IP Australia found evaluation and benefits tracking lagging behind deployment. |

## The problem, in numbers

Public administration is paperwork at industrial scale. Services Australia processed 468.5 million claims in 2023-24 and handled 1.1 billion online interactions. (Source 4) In the UK, researchers at the Alan Turing Institute estimate, in a preprint we could not confirm was peer reviewed, that central government conducts about one billion citizen-facing transactions a year, of which about 143 million are complex but repetitive, the kind that involve exchanging data and making decisions. (Source 7)

At that scale, the arithmetic is seductive. The same researchers calculate that saving one minute on each of those complex transactions would free about 1,200 person-years of work every year. (Source 7) That is a projection, not a result. Our separate review of [whether AI makes government faster](https://bfcs.ai/articles/can-ai-make-government-faster) deals with backlogs and speed. This one asks a narrower question: does AI lower what each document, form or piece of correspondence costs to handle?

### The evidence, in four numbers

468.5m · claims processed by Services Australia in 2023-24 · about 95% of its 600-plus automated processes are rules-based, not AI (agency strategy)

1.1 hours · a day saved summarising information, as estimated by Copilot trial participants · self-reported, 330 respondents; the evaluators say it likely reflects the upper bound (DTA, 2024)

13 minutes · a week an APS6 must redirect for a Copilot licence to pay for itself · Treasury break-even estimate; the trial did not measure time saved

~300 · extra full-time staff DWP estimates it would have needed to screen its post by hand · department estimate of an avoided cost, not a measured saving; the tool supplements existing processes (NAO, 2026)

## What improves, and why

The cost case rests on stable inputs. The UK National Audit Office describes computer vision as suited to extracting information from documents "that have a specific format, such as an invoice", and found agencies piloting or planning tools to lift text from forms into case systems and to compare documents for discrepancies. (Source 7) In our reading, where the same form arrives in the same layout thousands of times a day, a model can be trained once and checked against a known answer.

Services Australia shows what this looks like in practice. It uses intelligent document processing to check uploaded documents for missing or rotated pages, classify scanned images and route them to the right team, and optical character recognition to cut manual data entry. (Source 4) Its strategy is candid about where most of the automation sits, though: about 95% of its more than 600 automated processes are rules-based, not AI, and its transparency statement says it does not use AI to process claims. (Source 4) In our reading, the bulk of today's administrative automation in Australia's largest service agency is conventional automation on predictable work, with AI handling the intake around it.

> A saving per document is only a saving once someone measures it. In the audits we read, agencies had not.

## What the evidence actually says

The evidence falls into three kinds, and they should not be added together.

**Self-reported time savings.** In the Australian Government's six-month Copilot trial, over 5,765 licences were issued to staff nominated by their agencies. Participants estimated the tool saved them 1.1 hours a day on summarising, 1.0 hour on first drafts and 0.8 hours on searching for information. (Source 1) The evaluators caution that these figures likely represent an upper bound, that editing was almost always needed, and that up to 7% of participants said Copilot added time. (Source 1) A UK cross-government experiment with 20,000 licences produced an average self-reported saving of 26 minutes a day, but could not establish how the time saved was spent. (Source 3)

**Estimates and break-even calculations.** Treasury's own evaluation of 218 staff estimated that an APS6 would need to redirect about 13 minutes a week to higher-value work for the licence alone to pay for itself, not counting onboarding and training. It also states plainly that the trial did not quantify time savings. (Source 2) In the UK, the Department for Work and Pensions scans around 25,000 items of customer post a day with an AI tool, and estimates that manually processing that volume would have required almost 300 more full-time staff. (Source 6) The scanner supplements rather than replaces existing processes, and DWP told the auditor its impact is hard to quantify. (Source 6) In our reading, it is an avoided cost on paper, not a reduced one.

**Measured outcomes.** These are scarce, and the closest Australian example is cautionary. IP Australia put its first AI tool into patent examination in 2018 and now runs four. Between 2018-19 and 2024-25, applications across all IP rights rose 13.5% while average staffing rose 5.2%, but patents took 45.9% longer on average to process. (Source 5) The national auditor rated the agency's use of AI largely effective, yet found it "has not substantiated a link" between its AI tools and the efficiency of patent examination, that draft benefits profiles were not finalised or costed, and that there was no evidence of cost estimates before any of the four tools was built. (Source 5) A UK department that added observed tasks to its self-reported diaries found small time savings, some tasks taking longer, and no evidence that the savings improved productivity. (Source 3)

## Where it falls short

-   **Hours are not dollars.** A saving becomes cheaper administration only if fewer hours are bought or more work is done for the same cost. IP Australia said its efficiencies were meant to be reinvested in quality, not harvested. (Source 5)
    
-   **Projections have no price tag.** The UK estimate that almost a third of civil service tasks could be automated did not examine feasibility or cost. (Source 7)
    
-   **Legacy systems hold it back.** The UK auditor found that updating legacy systems and improving data access is fundamental to exploiting AI. (Source 7)
    
-   **Unstable inputs cost more to check.** In our reading, free-form correspondence and complex judgements need human review, and the Copilot trials found that checking and editing eroded the time saved. (Source 1), (Source 3)
    

## Where governance decides the outcome

Cost claims and safety claims rest on the same foundation: knowing what each tool does and what it costs. Under policy in effect since December 2025, Commonwealth agencies must keep internal registers of AI use cases, complete impact assessments for those in scope (existing uses by April 2027), and regularly monitor and evaluate them, with high-risk uses reviewed at least every 12 months. (Source 8) The audits show why. The ANAO found the ATO had 43 AI models in production but no evaluation approach for its automation and AI strategy, and a model list without cost or performance information. (Source 8) In our reading, an agency that cannot say what a tool costs or whether it works cannot say it is safe either. For decisions that affect people's entitlements, see our review of [benefits decisions](https://bfcs.ai/articles/can-ai-make-benefits-decisions-fairer).

## How to think about adoption

Start where the inputs are stable: standard forms, fixed layouts, high daily volumes and a known correct answer to check against. Measure the cost per item before the tool goes in, using staff time and error correction, not just licence fees. Cost the build and the running bill alongside the expected benefit; DWP's child maintenance model (a predictive model, not document processing) is one of the few published examples we found, at £625,000 to develop and about £460,000 a year to run, against a departmental estimate that it saved the equivalent time of 49 full-time employees, freed for other work; its full evaluation is not due until late 2027. (Source 6) Decide in advance whether freed hours will reduce spending, absorb growth or improve quality, because each is a different result. Treat self-reported time savings as a reason to evaluate, not as the evaluation. And keep free-form, high-judgement work under human review, where the unit-cost case is weakest.

## How we read this evidence

BFCS.ai does not fabricate figures. Trial time savings here are self-reported by participants, break-even and staffing figures are agency estimates, and the UK automation figures are projections. We found no independent audit that measured a fall in the unit cost of government document processing after AI was introduced, and we say so where it matters.

## Sources

1.  Digital Transformation Agency, "Evaluation of the whole-of-government trial of Microsoft 365 Copilot," summary and full report (evaluation by Nous Group), 2024. [digital.gov.au](https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full/executive-summary-glossary)
    
    Productivity: [digital.gov.au](https://www.digital.gov.au/initiatives/copilot-trial/microsoft-365-copilot-evaluation-report-full/productivity)
    
    Findings: [digital.gov.au](https://www.digital.gov.au/initiatives/copilot-trial/summary-evaluation-findings/cts-evaluation-findings)
    
    **What we took from it:** Over 5,765 licences were distributed from January to June 2024; the trial was non-randomised, with agencies nominating staff. Over 2,000 participants from more than 50 agencies contributed (1,556 pre-use, 1,159 pulse and 831 post-use survey respondents). Averaged post-use estimates of hours saved per day from respondents who completed both pre- and post-use surveys (n=330): summarising existing information 1.1, preparing a first draft 1.0, searching for information 0.8; the evaluators state these "likely quote the upper bound" of savings, assuming staff do the tasks every day. 69% of post-use respondents agreed Copilot improved the speed of their work and 61% the quality; around 40% reported reallocating time; up to 7% said Copilot added time; editing was almost always needed; agencies "will need to carefully weigh the potential benefits of efficiency and quality improvements against the costs, risks and suitability of generative AI". The 69% and 61% agreement figures are from the 330 matched respondents (Figure 11); "up to 7%" is the highest share for any single activity (preparing slides). Executive summary:
    
2.  Australian Centre for Evaluation (The Treasury), "Evaluation of a trial of generative AI (Copilot) in The Treasury," February 2025. [evaluation.treasury.gov.au](https://evaluation.treasury.gov.au/publications/evaluation-generative-artificial-intelligence)
    
    **What we took from it:** 218 participants over 14 weeks, 20 May to 23 August 2024. 49% of post-trial respondents agreed Copilot improved the speed at which they completed tasks, against 85% who expected it to before the trial; 61% said it reduced time spent on low-value tasks. ACE estimates an APS6 would need to redirect approximately 13 minutes a week to higher-value tasks to offset the licence cost (the licence alone, not onboarding or training) and concludes the benefits are likely to offset licence costs; "the data collected during this trial did not quantify time savings." The report does not state the licence price. It gives 218 participants, while its survey response rates use a base of 232.
    
3.  UK evidence on Microsoft 365 Copilot. [gov.uk](https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html)
    
    Department for Business and Trade, "The Evaluation of the M365 Copilot Pilot in the Department for Business and Trade," August 2025: [assets.publishing.service.gov.uk](https://assets.publishing.service.gov.uk/media/68adbe409e1cebdd2c96a19d/dbt-microsoft-365-copilot-evaluation.pdf)
    
    **What we took from it:** Government Digital Service, "Microsoft 365 Copilot Experiment: Cross-Government Findings Report," 2 June 2025: 20,000 licences across 12 organisations, 30 September to 31 December 2024; 7,115 survey responses; average self-reported saving of 26 minutes a day; 17% reported no clear time saving; "due to experimental constraints it was not possible to identify how time saved was spent"; no cost figures.
    
    1,000 licences, October to December 2024, about 30% randomly allocated; diary study (32% response) plus observed tasks to address self-report bias; respondents reported observing hallucinations. Small time savings were observed across most use cases, some tasks such as scheduling took longer, and the evaluation "did not find evidence that time savings have led to improved productivity"; value for money is still to be assessed.
    
4.  Services Australia, "Automation and AI Strategy 2025-27" (Word version, file dated May 2025): [servicesaustralia.gov.au](https://www.servicesaustralia.gov.au/sites/default/files/2025-05/automation-and-ai-strategy-2025-27.docx)
    
    Services Australia, "How we use Artificial Intelligence": [servicesaustralia.gov.au](https://www.servicesaustralia.gov.au/how-we-use-artificial-intelligence)
    
    Services Australia, "Automation and Artificial Intelligence Transparency Statement" (last updated 10 October 2025): [servicesaustralia.gov.au](https://www.servicesaustralia.gov.au/automation-and-artificial-intelligence-transparency-statement)
    
    **What we took from it:** 468.5 million claims processed in 2023-24 and 1.1 billion online interactions; over 600 automated processes at 30 October 2024; approximately 95% of the agency's automations are rules-based; OCR is used to extract customer responses from images and forms, for example for disaster payments.
    
    intelligent document processing checks uploads for missing or rotated pages and classifies scanned images to the right area; OCR reduces manual data entry.
    
    "We do not use AI to process claims."
    
5.  Australian National Audit Office, "Artificial Intelligence Use in IP Australia," Auditor-General Report No. 43 of 2025-26, 29 June 2026. [anao.gov.au](https://www.anao.gov.au/work/performance-audit/artificial-intelligence-use-ip-australia)
    
    **What we took from it:** First AI tool in patent examination deployed in 2018; four tools now deployed. Applications for all IP rights rose 13.5% between 2018-19 and 2024-25 against a 5.2% rise in average staffing; patent processing took 45.9% longer on average. IP Australia "has not substantiated a link between the effect of AI tools and changes in the efficiency of patent examination"; three of four tools had draft benefits profiles, none finalised and none including project costs; there was no evidence of cost estimates or approvals before development for any of the four tools; there is no evidence that a benefits realisation plan requested by the Executive Board was commenced. The ANAO rated IP Australia's use of AI in the patent rights process as largely effective overall. One tool cost $256,315 to develop and deploy (2019-20 to 2024-25) and $49,713 a year to maintain, as advised by IP Australia.
    
6.  UK National Audit Office, "DWP's approach to innovation," value for money report, HC 616, 21 October 2026. [nao.org.uk](https://www.nao.org.uk/reports/dwps-approach-to-innovation/)
    
    **What we took from it:** Whitemail Insights and Vulnerability Scanner: reviews around 25,000 documents a day of customer post that is not an administrative form, classifies it into themes and flags potentially vulnerable customers; it "supplements rather than replaces existing processes"; DWP estimated that manually processing the volume screened would have required almost 300 additional full-time staff; DWP said impact is hard to quantify; funding not disclosed. Child Maintenance Service predictive model: £625,000 development (September 2023 to March 2024) and around £460,000 a year to run; DWP estimates the early intervention it prompts saved the equivalent time of 49 full-time employees, "freeing up staff to work on other priorities", and plans a full evaluation in late 2027.
    
7.  UK National Audit Office, "Use of artificial intelligence in government," HC 612, 15 March 2024. [nao.org.uk](https://www.nao.org.uk/wp-content/uploads/2024/03/use-of-artificial-intelligence-in-government.pdf)
    
    Straub, V. J., Hashem, Y., Bright, J. et al., "AI for bureaucratic productivity: Measuring the potential of AI to help automate 143 million UK government transactions," Alan Turing Institute working paper, March 2024 (preprint; we could not confirm peer review). [turing.ac.uk](https://www.turing.ac.uk/sites/default/files/2024-03/ai_for_bureaucratic_productivity.pdf)
    
    **What we took from it:** Survey of 89 bodies in autumn 2023 (87 responded, 98%): 37% had deployed AI, with 74 use cases; computer vision can extract information "from documents that have a specific format, such as an invoice"; Cabinet Office (CDDO) indicative analysis estimated almost a third of civil service tasks could be automated but "did not examine the feasibility" or cost; baseline measures and KPIs had not been put in place; legacy systems and data are fundamental barriers.
    
    About one billion citizen-facing transactions a year across around 400 services, of which about 143 million are complex repetitive transactions; 84% of these are estimated to be highly automatable; saving one minute per transaction would equal about 1,200 person-years a year. Desk-based, assumes equal time per task, and covers only transactions that start access to a service.
    
8.  Digital Transformation Agency, "Policy for the responsible use of AI in government," v2.0 in effect from 15 December 2025, minor updates (v2.1) 21 October 2026: [digital.gov.au](https://www.digital.gov.au/ai/ai-in-government-policy)
    
    and [digital.gov.au](https://www.digital.gov.au/ai/ai-in-government-policy/ai-use-case-impact-assessment)
    
    Australian National Audit Office, "Governance of Artificial Intelligence at the Australian Taxation Office," Auditor-General Report No. 26 of 2024-25, 24 February 2025: [anao.gov.au](https://www.anao.gov.au/work/performance-audit/australian-taxation-offices-adoption-automation-and-artificial-intelligence)
    
    **What we took from it:** mandatory accountable officials, transparency statements, internal use case registers and AI use case impact assessments; in-scope use cases must be regularly monitored and evaluated; agencies must begin use case assessments within 12 months of the policy taking effect and apply it to existing use cases by 30 April 2027; high-risk use cases must be reviewed at least every 12 months.
    
    43 ATO-built AI models in production at 14 May 2024; 56 entities reported adopting AI in 2023-24; the ATO had not developed an evaluation approach for its automation and AI strategy, and its model list lacked ongoing cost and performance information.
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Make Government Faster?](https://bfcs.ai/articles/can-ai-make-government-faster/): More in Government · GOVERNMENT · FASTER · 3 min · Read →
- [Can AI Make Benefits Decisions Fairer?](https://bfcs.ai/articles/can-ai-make-benefits-decisions-fairer/): More in Government · GOVERNMENT · SAFER · 7 min · Read →
- [Can AI Cut Prototyping Costs?](https://bfcs.ai/articles/can-ai-cut-prototyping-costs/): Also Cheaper · ENGINEERING · CHEAPER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-reduce-administrative-costs/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml