Australia's $93 Billion AI Forecast Has a Roster Problem

AI-generated editorial illustration.

On October 22, the Tech Council of Australia put a large number on small-business AI: up to A$93 billion in additional annual economic value if adoption spreads over roughly five years. It also reported that only about one in four Australian small and medium enterprises used AI in the 2025 baseline behind the estimate, and just 5% used it broadly. Those figures describe a model and a survey period. They do not describe A$93 billion already earned in 2026.

A café, retailer or construction subcontractor does not receive a share of a national forecast. An owner buys software, trains a manager, checks a roster or quote, and pays the people who did the work. A faster schedule might free up an hour of administration. It might also move that hour into checking award rates, handling a missed shift and explaining changed hours to a worker. The difference has to appear in an operating account, not only a product demonstration.

The report prints the assumptions behind its headline. Employment Hero and Deputy, which sell small-business software, also contributed examples. Their cases identify tasks worth testing, but their reported savings are not an independent estimate for every customer. The narrower question for a buyer is whether a tool reduces paid effort or improves service after the setup, review and correction work is included.

A 2025 baseline behind a 2026 headline

The Tech Council published the full report on October 22, 2026. Delwood Advisory conducted the research with support from Employment Hero, Google and Deputy; the report says Deputy, Xero and LinkedIn contributed data. The headline rests on adoption data from the National AI Centre’s 2025 survey and on a Delwood model. A sponsor may know where a product helps. It also benefits when more companies buy software. The forecast should be read with both facts in view.

The report’s 2025 second-half chart puts AI use among firms with 0 to 199 employees at 26%: 21% limited use and 5% broad use. For firms with 200 or more employees, the chart shows 46% using AI, with 16% broad use. Smaller firms are different within the SME category too. It puts 0-to-4-person businesses at 23% use, firms with 5 to 19 employees at 31%, and firms with 20 to 199 employees at 37%. A buyer considering software for a three-person shop should not assume the experience or capacity of a 150-person company.

But the report is not the latest reading of the underlying adoption tracker. In a June 2026 National AI Centre update, the government said 43% of surveyed SMEs reported some AI adoption across December 2025 to February 2026, with February reaching 44%. The earlier Tech Council chart and that later quarterly result cover different waves. The underlying questions, respondent mix and cutoffs need checking before someone treats the gap as a precise rise from 26% to 43%. The later reading does establish that the 2025 baseline is a dated input to an October 2026 economic scenario.

The difference is too large to hide in a footnote. The Tech Council report labels its 2025 input as ‘broad’ and ‘limited’ use and says its National AI Centre data came from roughly 400 respondents per month. The National AI Centre’s later update calls 43% ‘some level of AI adoption.’ The two releases do not publish a reconciled, like-for-like bridge. A reader cannot calculate a reliable growth rate by subtracting the two headline percentages. The sensible conclusion is narrower: Australian small-business AI use is being counted under changing survey windows, while the economic forecast fixes one older window as its starting point. A future update should publish the exact question, weighting and cohort definition alongside the modeled gain.

A dated baseline changes the budget debate. If adoption has accelerated, a five-year upside calculated from the older starting point can overstate the number of firms still to be reached. If the surveys define use differently, neither percentage can settle how many businesses have redesigned work across payroll, scheduling and customer service. A manager with a chatbot subscription counts differently from a manager who moved the weekly roster, wage-costing and exception review into a tested workflow.

The Australian Bureau of Statistics reported 689,600 employing businesses with one to four workers, 232,912 with five to 19 and 68,325 with 20 to 199 at June 2026. It also counted a much larger non-employing population. The firm-count composition makes an average per-SME benefit a dangerous purchasing shortcut: size, industry, revenue, labor mix and software readiness vary before a single forecast is run. The Tech Council’s estimate is an aggregate scenario, not a quoted return for a particular shop.

The 53 percent adoption assumption

The report does not simply multiply today’s users by a savings figure. Its base case assumes AI adoption reaches 53% of SMEs in roughly five years, split into 27% limited use and 26% broad use. The model says that could produce A$93 billion in extra annual value added, equivalent to roughly 9% of the current contribution it assigns to SMEs. A lower case, in which SME adoption approaches the reported enterprise rate, yields A$61 billion. These are two modeled outcomes, not observed year-end accounts.

Five assumptions carry the base case. The model assumes all SMEs already in the process of implementing AI become limited users; all that say they intend to adopt do the same; 25% of firms unaware of AI start limited use; 10% of firms that do not intend to implement it change their minds; and all current limited users move to broad use. The final step moves an entire existing cohort across a much harder operational boundary. That is where an owner must connect data, change a process, train staff and decide what to do when the output is wrong.

The composition of the A$93 billion case makes that dependency visible. The report assigns A$75 billion of additional value to broad adoption and A$18 billion to limited use. Broad use is roughly half of the projected adopter share but about four-fifths of the modeled value. A training course that helps a firm open a chatbot account could raise an adoption statistic without moving it to the high-value part of the model. That does not mean the course has no value. It means the public program needs a separate measure of connected work and realized value.

There is a respectable case for starting with limited use. A three-person operator may learn more from automating invoice drafts or customer messages before connecting sensitive pay data. A staged rollout can expose data-quality problems at lower cost. The model’s broad-use assumption therefore should not become a reason to rush payroll automation; it is a statement about where the upside is concentrated if the implementation works.

For an average SME that goes from no AI to broad use, Delwood estimates A$136,000 of additional annual value added, a 36% productivity uplift over its assumed A$381,000 baseline. Limited use carries A$54,000 in modeled value. The report’s productivity table converts A$71 of hourly value added without AI into A$81 for limited use and A$96 for broad use. It says these figures are indicative and that the hourly ratios assume stable employment. Stable employment is an important condition when a firm evaluates a scheduling tool that can change weekly hours.

Value added is also not a cash saving. A business could serve more customers with the same staff hours, cut avoidable overtime, or improve margins. It could instead pass the gain to customers, raise wages, or invest in training. The model does not tell a particular owner which path occurred. Nor does it show that a generated roster causes a 36% uplift in a hospitality venue. The estimate combines adoption transitions with productivity assumptions drawn partly from broader external research. The report’s vendor cases illustrate possible applications rather than testing the national causal claim.

Here is the report’s adoption path beside the records a business would need to call it an operating result.

Step in the report’s five-year scenario	Model status	What an owner could verify
26% of SMEs used AI in the 2025 baseline	Survey reading, 21% limited and 5% broad	Which tasks use AI now, by employee and shift
53% use AI in the base case	Assumed future adoption mix, 27% limited and 26% broad	Paid rollout, active users and completed workflows
All current limited users become broad users	Assumption, not measured migration	Whether payroll, roster or sales data actually connect
A$136,000 per average firm moving from none to broad use	Modeled annual value added	Before-and-after contribution, hours, error costs and service quality
A$93 billion additional annual value	National modeled scenario	Audited outcomes across firms, not a sum of vendor anecdotes

The table identifies what a public projection leaves for a shop to prove. A founder can pay for a pilot now. A minister can fund training now. Neither should book five years of projected value as current revenue.

Rosters turn productivity into a labor decision

AI-generated editorial illustration. The three paths represent separate operating checks, not measured results.

A weekly roster puts the forecast on someone’s Tuesday shift. Deputy’s October 12 product guide describes separate capabilities: forecasts use sales, transactions or foot traffic to estimate demand; Smart Scheduling compares scheduled wages with expected sales; Deputy AI can assist an authorized manager with schedule tasks. The company says managers still review the final roster. That distinction between forecasting, labor modeling and an AI assistant matters because a buyer could otherwise attribute an ordinary rules-based scheduling gain to a generative model.

Imagine a venue with a strong Saturday forecast. That is an illustrative operating example, not a reported customer case. The manager might add a shift to protect service, or decide that the forecast overstates demand because a local event was canceled. A product that saves twenty minutes building the initial schedule could still increase the time spent checking availability, split shifts and the expected wage bill. If the manager cuts one shift, software savings and reduced labor spend become mixed. The worker may experience fewer paid hours while the vendor reports faster planning.

The wage bill leaves little room for a sloppy pilot. Employment Hero’s September 2026 jobs report said headcount among businesses in its platform data rose 0.8% from August, while annual median wage growth slowed to 2.2%. It described the headcount rise as the strongest single month since May and said wages had weakened in 11 of 13 industries. Platform customer data are not the entire Australian labor market, and a monthly employment gain does not prove AI was used. The data do show that a small-business owner can be hiring and restraining pay at the same time.

Deputy’s separate analysis of shift records covers tens of millions of shifts from September 2025 through August 2026 and ranks 85 regions by hours, shift length, night and weekend work, and roster notice after adjusting for industry mix. It measures what happened in Deputy’s customer base, not the effect of adopting AI. Its relevance is the denominator: an optimization tool should be judged against actual hours, notice and coverage, not just minutes saved in a manager’s calendar.

The Tech Council says regional SME AI adoption was 16% in its 2025 chart, compared with 30% in metropolitan areas. A regional operator may have fewer staff to swap into a shift, less help integrating software, and weaker training access. That does not mean adoption will be less valuable. It means the vendor’s onboarding and the manager’s exception capacity are part of the economics. A tool bought by a small team can turn its only supervisor into a permanent data cleaner if the inputs are unreliable.

Deputy offers one concrete, limited example. In its October guide, Funk Drinks Co. general manager Mari Bornelli says a pay cycle that previously took about two to three hours now takes 45 minutes with Deputy. She also reports greater confidence in compliance. This is a named customer’s account on the vendor’s site, not an independently audited comparison; the page does not isolate the contribution of a generative AI feature from digitized timesheets, configured rules or process cleanup. It is still a useful pilot question: can the new pay cycle be reproduced in another venue after setup time and corrections are counted?

A worker’s objection may be different from the manager’s. A forecast can fit expected sales and still distribute weekend shifts unevenly, offer less notice, or put a new worker on a difficult service period without support. A manager can choose to override the suggestion. The override is not necessarily a product failure. It may be the point at which local knowledge prevents a tidy forecast from becoming an expensive shift. The pilot should record that judgment instead of counting every override as friction to be eliminated.

Payroll savings have a compliance floor

Payroll makes the value claim more complicated. The Tech Council’s report points to hospitality cases where vendors report substantial administration savings. One example describes Experience Gold Coast after a merger of five entities and a unified Employment Hero system. That is a particular implementation of HR, payroll, rostering and related tools in a merged organization. It is not evidence that a café with five employees will save the same number of hours, or that the gain came from AI rather than consolidated records and process change.

Fair Work Ombudsman guidance says awards and registered agreements can set rules on hours, breaks, roster changes and notice. Its record-keeping guidance requires employers to keep information on pay rates, gross and net pay, deductions, loadings and penalty rates, among other details. Software may calculate or surface these items. The employer still needs the correct award configuration, inputs and records. A wrong classification can make an apparently faster pay run expensive to repair.

The Fair Work Ombudsman’s consultation guide says awards and enterprise agreements can require consultation on significant technological or organizational changes, and some awards require consent for roster changes. Those duties depend on the applicable instrument and change, not on whether the scheduling suggestion came from AI. A buyer should check the rules that govern its staff before testing automated publication. A fast roster issued under the wrong notice or consent rule is not a productivity gain.

Deputy’s guide separately tells managers that its forecasting tools and AI assistant support review; employers remain responsible for applying current award rules. This is a more useful procurement statement than a blanket automation promise. A system may be fast at generating a proposed roster. The test is whether the proposed roster survives an award-rule check, a worker’s availability correction, the actual sales week and the final pay run. Those steps occur on different days and often involve different people.

A cost calculation needs subscription fees, setup, data cleaning, manager review, payroll correction, training, help-desk time and any extra labor needed to cover a service gap. It should also record the positive side: avoided overtime, fewer pay corrections, better roster notice, more sales per paid hour, or less owner time spent on administration. A single percentage of “time saved” hides which hours were paid, whose time moved, and whether customers received the same service.

An illustrative calculation shows how narrow the margin can be. Suppose a manager saves two paid hours each week and those hours are valued at A$35 each. Over 52 weeks the gross labor value is A$3,640. A subscription costing A$300 a month consumes A$3,600, leaving A$40 before setup, training, review and errors. Those prices and hours are assumptions for arithmetic, not Australian averages or a forecast of any product. If the manager is an owner who works unpaid extra time, the cash saving can be lower still. If the same hours instead allow the shop to take more orders without another hire, the value must be demonstrated in contribution from those orders.

The arithmetic forces the unit of benefit into view. A pay-cycle estimate should specify who saved time, how many cycles occur each year and what the replacement use of that time was. A roster estimate should say whether fewer hours were actually paid or whether the work moved to service and training. A revenue claim needs the incremental margin after the labor that produced it. Without those denominators, a 20-minute improvement can be both real and financially immaterial.

Employees experience a scheduling change in paid hours and notice, not in annual value added. The Tech Council cites an older Deputy survey in which 54% of SME employees who used AI reported improved job satisfaction. That is a self-reported experience among users, not a finding that a particular scheduling deployment improved every worker’s hours or pay. Employee feedback about roster predictability and correction burden belongs beside the manager’s savings. An operator may discover a useful tool and a bad implementation at the same time.

Britain offers a second model, not a validation

The Australian report arrived within days of another small-business AI forecast. The UK’s Social Market Foundation published research on October 19 based on a survey of 1,015 senior decision-makers in SMEs, fielded June 27 to July 9, 2026. It modeled possible revenue gains under wider adoption, using self-reported outcomes and external data. Intuit sponsored the work; the foundation says it retained editorial independence. That is a different sample, economy and outcome from Australia’s value-added model.

The UK paper makes a useful methodological point. Surveying adopters about whether AI increased revenue can reveal where firms believe benefits exist. It cannot by itself prove that the same firms would not have grown without AI. Businesses with stronger managers, cleaner records or more cash may be more likely to adopt AI and more likely to grow. A model can control for some observed differences; it cannot turn a cross-sectional self-report into a randomized deployment. The Australian A$93 billion forecast and the UK revenue scenario should be read as separate possibilities, not mutual confirmation.

Both reports describe a practical expense that adoption rates omit: helping staff use the software. The Social Market Foundation identifies data security, quality and skills concerns and proposes one-to-one advice and an assurance system. The Tech Council calls for practical training and sector-specific guidance, especially for regional firms. The software sellers backing both studies stand to benefit from adoption. Their presence does not make every observation false. It raises the value of an independent operating test before a public grant or private budget is renewed.

The Australian report’s sponsorship and data contributions are disclosed in its opening pages. The UK foundation’s release describes a separate survey and modeling exercise sponsored by Intuit. Neither has published a controlled estimate of what happens to a randomly selected small employer that buys a specific product. Buyers should not borrow the A$136,000 modeled gain as a sales target or the UK revenue estimate as corroboration for it. A careful vendor could win trust by publishing task-level costs, error rates and worker effects for customers who did and did not expand beyond basic use.

Australia’s National AI Centre adoption tracker asks businesses about tools, applications, benefits, challenges and responsible practices. That kind of repeated survey can show diffusion. It cannot directly show what happened to one firm’s payroll cost, customer wait time or hours offered to staff. A public program evaluating AI training could pair adoption survey data with operational follow-up. Without that second part, “adoption” can become the final outcome even though it is only the start of the work.

An operator’s four-week test

A small firm does not need to settle national productivity economics before running a controlled purchase. It needs a baseline that is short enough to collect and hard enough to challenge a sales claim. The following four-week scorecard is an editorial proposal, not a published trial or promised return. It suits a roster-and-payroll pilot; a construction or professional-services firm should substitute the relevant job and service measures.

Measure	Baseline before the tool	Pilot observation	Decision rule
Manager time	Minutes to plan, publish and revise a roster	Include data entry, checking and employee messages	A faster first draft alone does not count
Payroll effort	Time and cost to process and correct a pay cycle	Count corrections after the pay run	A gain must survive the full cycle
Labor deployment	Paid hours, overtime and roster notice	Compare with demand and worker availability	Investigate savings caused by fewer offered hours
Service	Sales per paid hour plus wait or missed-service incidents	Watch both revenue and customer experience	Reject a lower wage bill that breaks service
Worker outcome	Hours offered, late changes and reported predictability	Ask affected workers and track actual schedules	Keep change visible by role and shift
Cash cost	Software, integration, training and support	Include the manager’s checking time	Renew only on net contribution, not demo speed

Week one should capture how the existing process behaves, including its mistakes. Week two should run the new system in parallel without letting it silently change pay or hours. In week three, the manager can publish a reviewed roster while recording every override and why it happened. The fourth week should close the payroll and service loop. One calendar month is not enough to infer long-run productivity, seasonal performance or causal labor-market effects. It can reveal whether the promised saving survives a complete operating cycle.

A vendor might object that this test is too cautious: pilots add parallel work, and some benefits appear only after staff learn the tool. That is a reasonable concern. The buyer can record training time as an investment and set a longer follow-up. It should still avoid recognizing a benefit before pay corrections, customer service and worker hours have been observed. The model estimates an economy-wide possibility. The pilot can show whether the tool worked in one firm.

The Australian Cyber Security Centre’s small-business AI guidance adds another necessary input: data handling. Rosters carry availability and personal information; payroll carries identifiers and wage records. An owner should know what information enters a model, who can retrieve it, and what happens if a vendor or integration fails. Security work is part of deployment cost, not an optional footnote after the savings estimate.

The missing denominator is paid work

The Tech Council’s 53% adoption scenario could prove conservative, ambitious or simply hard to compare with later surveys. More recent Australian surveys show experimentation, but the national value-added estimate cannot say who will receive the gain. The shift manager can say whether next week’s roster is quicker to make. The worker can say whether hours and notice improved. The payroll record can say whether the pay run was correct. None of these is interchangeable with a vendor’s adoption count.

Government and software companies can use the A$93 billion figure to specify what must happen to adoption, task design and productivity. A purchase order needs closer evidence. At the first renewal meeting, the owner should lay out the roster, paid hours, correction log and service result. The number on the invoice is real; the modeled national gain still waits for those records.