Google’s AI Economy Map Goes Beyond Office Jobs: What ATLAS Reveals About Hands-On Work

Google’s AI & Economy ATLAS suggests workplace AI is spreading beyond office roles into manual and technical work, particularly diagnostics, troubleshooting and on-the-job learning. For Sydney and NSW operators, the practical opportunity is not replacing tradespeople. It is reducing the information friction surrounding physical work, while keeping site judgement, licensing, safety controls, approvals and responsibility with qualified people.

The most important finding in Google’s latest map of the AI economy may be what it does not show.

It does not show artificial intelligence simply moving down a list of occupations and automating each job in turn.

Instead, Google’s AI & Economy ATLAS research describes a much more fragmented transition. AI use is appearing inside hundreds of occupations, but workers are applying it selectively to particular tasks. In a typical occupation represented in the research, AI is used for only about 21 per cent of tasks, while fewer than 10 per cent of work-related interactions fully automate a task.

That distinction becomes especially important when the work involves a vehicle, machine, electrical system, concrete slab, air-conditioning plant, building defect or active construction site.

The physical task may still belong firmly to a technician, tradesperson, supervisor or specialist contractor. AI increasingly sits beside that task, helping the worker interpret information before acting.

The AI Economy Is Reaching the Workshop and the Job Site

Google launched the first ATLAS dataset in July 2026 using approximately 15 million aggregated and de-identified interactions across the Gemini App, AI Mode and Gemini API. The research mapped activity across more than 150 countries, 140 languages, 800 occupations and 4,000 tasks.

Google then expanded the project in September with an open-access ATLAS exploration tool that allows users to examine AI adoption by occupation and geography, including occupations such as electricians.

One finding challenges the assumption that generative AI is principally an office technology. Google says workers in predominantly physical and manual occupations are already using conversational AI for adjacent tasks such as real-time diagnostics, troubleshooting and learning.

Automotive technicians and industrial mechanics are examples in the research. Google reports that when workers in these types of occupations use its AI tools, they are twice as likely to use multimodal capabilities involving visual information. Examples include interpreting test results, investigating electrical wiring problems and inspecting equipment for signs of wear.

The September ATLAS update adds a geographic dimension. Google reported that manual-task uses such as equipment diagnostics and troubleshooting accounted for 7 per cent of work-related AI usage in Brazil and Germany, compared with 4 per cent in Japan.

These are not Australian labour-market measurements, and they should not be presented as Sydney adoption rates. They do, however, reveal an operating pattern relevant to Australian property, construction, maintenance and infrastructure businesses.

AI can enter a physical occupation without performing the physical work.

Hands-On Jobs Have an Information Layer

A construction worker, flooring specialist, HVAC technician, mechanic or maintenance contractor does not spend every minute physically manipulating material or equipment.

The job also contains an information layer.

Workers inspect conditions, compare measurements, retrieve specifications, interpret fault information, decide what tool or product is appropriate, communicate with supervisors, record evidence, order materials and explain unexpected conditions to customers or project managers.

That is where conversational and multimodal AI can begin to matter.

• Concrete grinding and floor preparation
• AI-adjacent information task: Organising site photographs, comparing substrate observations, retrieving product information and preparing variation notes.
• What still requires human control: Assessing the slab, selecting the work method, controlling machinery, managing silica risk and verifying the finished surface.
• Air-conditioning diagnosis
• AI-adjacent information task: Interpreting fault codes, comparing readings, retrieving technical information and structuring troubleshooting steps.
• What still requires human control: Testing the actual system, electrical and refrigeration work, confirming the fault and carrying out regulated work.
• Electrical maintenance
• AI-adjacent information task: Finding technical references, explaining test results and helping organise diagnostic possibilities.
• What still requires human control: Electrical isolation, testing, wiring, compliance certification and decisions made by an appropriately licensed electrician.
• Property maintenance
• AI-adjacent information task: Turning photographs and notes into an inspection summary, categorising defects and retrieving maintenance information.
• What still requires human control: Determining the actual condition, establishing causation and deciding whether specialist investigation is required.
• Renovation coordination
• AI-adjacent information task: Summarising site information, preparing scope drafts, identifying missing documentation and organising programme dependencies.
• What still requires human control: Approving scope, resolving site conflicts, sequencing contractors and accepting commercial or safety consequences.

The distinction matters because it changes the question businesses should ask.

Instead of asking, “Can AI do this trade?”, the more useful question is, “Which information bottlenecks are slowing the qualified person who already does this trade?”

Multimodal AI Is Particularly Relevant to Physical Operations

Text-only AI naturally suited office environments because much professional work already arrived in text: emails, reports, contracts, spreadsheets and presentations.

Physical environments are different.

A site condition may first appear as a photograph. A fault may be represented by a control panel, test result, wiring configuration, unusual wear pattern or surface defect. A renovation variation may begin with a supervisor sending three photographs and a short voice note from a basement.

That makes multimodal systems more relevant than a conventional chatbot for many hands-on roles.

Consider a Sydney floor-preparation project where an unexpected second adhesive layer appears after carpet or timber removal. AI could potentially help the project team organise photographs, retrieve previously supplied scope information, compare product documentation and draft an explanation of the variation.

It cannot establish from a photograph alone that the substrate is safe to grind, determine every contaminant present, physically measure the slab or guarantee compatibility with the next flooring system.

The useful role is information acceleration around the field decision, not substitution for field verification.

That distinction complements Elyment’s analysis of AI agents moving into engineering and construction workflows. Faster digital reasoning becomes commercially useful only when its outputs can survive contact with the physical project.

ATLAS Shows Breadth, Not Full Automation

Some of the strongest numbers in ATLAS are easy to misread.

Google found AI activity across occupations representing the great majority of US employment. That does not mean the corresponding jobs have been automated, or that every worker in those occupations is using AI.

The research instead describes broad but shallow penetration.

In practical terms, a worker may use AI during one troubleshooting task, one documentation step or one piece of research while completing the rest of the job conventionally.

This is especially plausible in physical operations because a single role may contain radically different task types.

• A technician may diagnose from test data, then physically dismantle and repair equipment.
• A project supervisor may summarise a defect report, then inspect the condition personally.
• A flooring contractor may retrieve a technical data sheet, then mechanically prepare the substrate.
• A building manager may use AI to organise maintenance requests, then engage a licensed specialist for regulated work.
• A site coordinator may prepare a revised sequence digitally, while the actual programme remains constrained by access, labour, drying time, deliveries and other trades.

This is less dramatic than the idea of entire occupations disappearing overnight, but operationally it may be more important.

Small improvements across dozens of decisions can change labour utilisation, response times and project administration even when the core physical work remains unchanged.

Sydney Projects Make the Adjacent Work Expensive

The information surrounding physical work carries a particularly high cost in Sydney property and renovation environments.

Contractors may be working around strata access windows, loading zones, lift protection requirements, noise restrictions, tenant occupancy, waste movements, parking constraints, delivery timing and multiple subcontractors using the same space.

A five-minute technical question can become an hour-long disruption when the right document, photograph, person or approval cannot be found.

This means the economic value of AI on physical projects may not come from asking a model to operate a grinder, install flooring or service an air-conditioning unit.

It may come from reducing the friction immediately before and after those activities.

Elyment has previously examined why removal, grinding, levelling and installation must be sequenced around real site constraints. AI can help organise the information surrounding that sequence, but the programme still depends on physical readiness and accountable coordination.

The Most Valuable Use Case May Be Faster Diagnosis

Google’s emphasis on troubleshooting is significant because diagnosis is often one of the most expensive uncertain stages of physical service work.

A technician can lose time searching manuals, reviewing fault histories, contacting another specialist or reconstructing information from incomplete records.

A multimodal assistant could shorten part of that process by combining photographs, measurements, codes, technical documentation and historical notes into a more structured starting point.

The benefit is not necessarily fewer technicians.

It may be more productive technician time.

A business testing AI-supported field diagnosis should therefore measure operational results rather than chatbot activity. Useful measures include:

• time from reported problem to initial diagnosis;
• percentage of jobs resolved on the first visit;
• repeat call-outs caused by incorrect diagnosis;
• time spent searching technical information;
• number of escalations to senior technicians;
• parts ordered incorrectly;
• rework attributed to poor information;
• time from site discovery to approved variation; and
• percentage of AI suggestions materially corrected by a qualified worker.

The last measure is especially important. Faster information is useful only if the information is sufficiently reliable for the decision being made.

NSW Regulation Does Not Disappear Because the Advice Came From AI

For NSW businesses, AI adoption inside physical work also intersects with a changing regulatory environment.

SafeWork NSW’s 2026 work on Digital Work Systems Guidelines is particularly relevant. NSW legislation now expressly recognises digital work systems including algorithms, artificial intelligence, automation and online platforms, and clarifies that existing work health and safety duties apply to risks arising from their use.

That matters if software starts influencing how work is allocated, prioritised or carried out.

A digital system recommending a faster sequence does not eliminate the obligation to identify the site's real hazards. An automated job allocation does not make fatigue, travel time, competency or equipment availability irrelevant. A generated procedure does not become safe simply because it appears detailed.

SafeWork NSW also requires site-specific planning for high-risk construction work. A safe work method statement for high-risk construction work must address the actual hazards, risks and controls affecting that site.

AI may help organise information for a safety workflow. It should not turn site-specific risk assessment into generic document production.

Licensing Creates Another Boundary

The ability to generate technical guidance is not the same as legal authority to carry out regulated work.

NSW Fair Trading states that a person must hold an electrical licence before performing electrical wiring work in NSW. Specialist residential work such as electrical wiring, plumbing, gasfitting and air-conditioning or refrigeration work can also require appropriate licensing regardless of job value.

That boundary becomes increasingly important as AI becomes better at explaining technical procedures.

A model may be able to describe a fault, identify possibilities or retrieve technical information. That does not convert an unlicensed person into a licensed tradesperson, transfer statutory responsibility to software or establish that a particular recommendation is safe on a particular site.

For operators, the governance rule should be straightforward: AI can expand access to information without expanding a worker's legal authority.

The Workflow Should Separate Assistance From Authority

Businesses introducing AI into field operations need a clear point where machine assistance stops and accountable human authority begins.

1. Capture. Collect photographs, measurements, equipment data, client information and relevant project records.
2. Interpret. Use AI to organise information, retrieve references, compare possibilities or prepare an initial diagnostic pathway.
3. Verify. A competent worker checks the physical condition, the source material and the assumptions made by the system.
4. Authorise. The appropriate person approves the technical action, commercial variation, safety control or regulated work.
5. Record. The actual result, not merely the AI suggestion, becomes part of the project or maintenance record.

This prevents one of the more subtle automation failures: an AI-generated possibility gradually becoming treated as an established site fact because it has been copied through several systems.

The same principle applies to autonomous business processes. Elyment’s analysis of AI services operating as background workers noted that automated work still needs limits, approval rules, audit records and clear ownership of exceptions.

Faster Analysis Can Simply Move the Bottleneck

Google’s September ATLAS update contains another useful lesson from scientific work.

In separate research highlighted alongside ATLAS, surveyed scientists reported saving just under seven hours per week through AI. Google also noted that faster intellectual work could create downstream bottlenecks, including more hypotheses waiting for physical experimentation and validation.

Physical businesses can experience the same effect.

If AI allows a maintenance company to diagnose faults faster, the bottleneck may move to parts availability.

If renovation variations are documented faster, the constraint may become customer approval.

If estimators produce scopes faster, the constraint may become inspection capacity.

If supervisors identify defects faster, the programme may still depend on the correct trade becoming available.

AI therefore does not automatically increase end-to-end throughput. It can expose the next limiting step in the process.

This connects with another part of Elyment’s coverage: AI infrastructure itself is increasing demand for skilled trades and project-delivery capacity. Digital intelligence can expand quickly. Physical capacity remains constrained by people, equipment, access, materials and time.

What ATLAS Does Not Prove

ATLAS is unusually large, but businesses should resist turning usage data into conclusions it was not designed to establish.

It is primarily an observational picture of how Google AI tools are being used. It is not a controlled productivity trial, a forecast of Australian employment or proof that a particular occupation can be automated safely or economically.

The occupational findings also reflect the people choosing to use Gemini and related Google products. Adoption may therefore look different across businesses, countries, age groups, technology ecosystems and industries.

Most importantly, frequency of AI use does not establish quality of outcome.

A technician asking AI ten questions has adopted AI more intensively than someone asking one. That does not tell a project manager whether the first technician reached the correct diagnosis, avoided rework or completed the job faster.

For property and construction operators, outcome measurement still has to happen inside the business.

The Competitive Advantage May Sit Between Digital and Physical Work

ATLAS makes the future of hands-on work look less like a contest between humans and machines and more like a redesign of the boundary between information and execution.

The worker still sees, touches, measures, installs, removes, grinds, tests, repairs and verifies.

AI increasingly helps retrieve, compare, interpret, document and coordinate.

The businesses likely to gain the most operational value are therefore not necessarily those that automate the largest number of activities. They are the ones that identify where information delays physical work and then redesign that specific handoff without weakening technical judgement, safety or accountability.

In Sydney renovation and property operations, that may mean better site-to-office communication, faster access to technical information, clearer variations, improved documentation, more structured defect analysis and fewer hours lost reconstructing what happened.

None of those improvements removes the need for skilled people.

They make skilled time more valuable.

PHYSICAL WORK. BETTER INFORMATION.

Review Where AI Can Support Your Operations Without Losing Human Control

Elyment works across physical project delivery, renovation planning and operational systems. Review where diagnostics, documentation, approvals, compliance and contractor coordination can be improved before automation becomes another layer of complexity.

Discuss Your Project With Elyment

The Real ATLAS Lesson

Google’s AI economy map does not suggest that hands-on work is becoming software.

It suggests that software is moving closer to hands-on work.

For Sydney businesses, trades and project operators, the important opportunity sits in that gap: using AI to make the information surrounding physical execution faster and more useful, while preserving the professional judgement, site verification, safety controls and accountability that the real-world job still requires.

Sources and References

• Google: Understanding the AI Economy
• Google: AI & Economy ATLAS
• Elyment: AI Agents Moving Into Engineering and Construction Workflows
• Elyment: Sequencing Removal, Grinding, Levelling and Installation Around Real Site Constraints
• SafeWork NSW: Digital Work Systems Guidelines
• SafeWork NSW: Prepare a Safe Work Method Statement
• Elyment: AI Services Operating as Background Workers
• Elyment: AI Infrastructure Increasing Demand for Skilled Trades and Project-Delivery Capacity
• Elyment: Discuss Your Project