EY Is Expanding Microsoft AI to 400,000 Staff After a 15% Productivity Gain: What Should Businesses Expect From AI Consulting Firms?

EY's move to expand Microsoft's enterprise AI environment across more than 400,000 people signals a shift in what businesses should expect from AI consulting. For Sydney and NSW organisations, buying licences or running isolated pilots is no longer enough. Consultants increasingly need to prove adoption, redesign workflows, establish governance, measure operational outcomes and show how time saved by AI is converted into better delivery rather than simply reporting theoretical productivity.

The Important Number Is Not 400,000. It Is What Happened Before 400,000.

Microsoft says EY initially deployed Microsoft 365 Copilot to about 150,000 people and reported a 15 per cent productivity gain. Microsoft has since said EY is expanding its Microsoft AI environment across a global workforce of more than 400,000 people. In Microsoft's FY2026 fourth-quarter earnings call, the company described the EY deployment as a 400,000-employee rollout and its largest E7 win at that point.

That distinction matters.

The reported 15 per cent improvement did not emerge from handing 400,000 employees a new AI interface on the same morning. EY's published account describes an earlier deployment to approximately 150,000 Copilot users, with the productivity improvement being reinvested into client delivery and learning.

Microsoft has also reported strong usage indicators within the deployment, including 94 per cent monthly adoption, 85 per cent weekly use and 81 per cent of enabled employees reporting time savings. Those figures are vendor-reported outcomes, rather than an independent productivity audit, but they illustrate the scale of operational adoption required before an AI program can reasonably be called enterprise transformation.

For Sydney businesses, this reframes the AI consulting conversation.

The emerging benchmark is not whether a consultant can demonstrate Copilot, ChatGPT, Gemini or an AI agent. It is whether they can build the operating conditions that make a useful system repeatable across departments, employees, records, approvals and real customer work.

Elyment's Existing AI Coverage Makes a Different Angle Necessary

Elyment's existing AI coverage has already examined why workflow-embedded tools can outperform standalone chatbots, why AI strategy should precede automation, how automation economics should be measured around completed outcomes and why agents need real business context. Its published guidance on free AI consulting agents also warns that generated strategy should not be mistaken for an implementation plan.

The EY development raises a different operational question:

What should an organisation expect an AI consulting firm to deliver after the technology has already demonstrated that it can save time?

That is increasingly the more difficult stage.

A pilot can prove that AI drafts emails faster, summarises meetings, searches documents or accelerates analysis. Scaling requires the organisation to decide which work should change, who owns the changed process, how adoption will be managed, how information is controlled and where the resulting capacity should be redeployed.

The Consulting Deliverable Should Be an Operating Model, Not a Tool Demonstration

The traditional enterprise technology engagement often moved through familiar stages: requirements, procurement, configuration, deployment and training.

AI makes that sequence less reliable because the software may be available before the organisation has decided how work should change.

A credible AI consulting engagement should therefore produce an operating model covering at least six areas:

• Workflow ownership: Who is responsible for each AI-assisted process.
• Access design: Which systems, documents and records AI can reach.
• Decision authority: What AI may prepare, recommend or execute.
• Exception handling: What happens when the system is uncertain, incorrect or outside scope.
• Measurement: How time, quality, cost and error rates will be compared before and after deployment.
• Change control: How prompts, models, integrations, permissions and workflows are updated without silently changing business risk.

This is materially different from installing software.

A structured AI consulting review for Sydney businesses should connect technology choices to actual workflow design, implementation responsibilities and measurable business outcomes rather than treating model selection as the project itself.

Productivity Has to Be Traced Through the Workflow

A reported productivity gain sounds straightforward until management asks where the gain appears financially.

Consider a hypothetical Sydney property services company where ten operations staff each recover 45 minutes per day through AI-assisted inbox triage, meeting summaries, quote preparation and project reporting.

The company has not automatically created additional profit.

The recovered time may become:

• More completed quotes.
• Faster project scheduling.
• Better client follow-up.
• Additional quality checks.
• Less overtime.
• Higher capacity without immediate recruitment.
• Additional training.
• Or simply unused fragments of time distributed across the week.

These outcomes are commercially different.

EY's own public description is notable because it says time saved was reinvested in areas including client delivery and continuous learning. That is an organisational decision about capacity allocation, not an automatic property of Copilot.

AI consultants should therefore be expected to help management answer two questions separately:

1. Did the AI reduce the effort required to perform the task?
2. Did the organisation successfully convert that reduction into a business outcome?

The first is a technology and process measurement.

The second is an operating-management question.

A 15% Improvement in One Task Is Not a 15% Improvement in the Business

One of the easiest ways to overstate AI economics is to generalise a local productivity result across an organisation.

A workflow that moves from 20 minutes to 17 minutes has improved by 15 per cent on that activity. That does not mean revenue, profit, employee capacity or enterprise productivity has increased by the same percentage.

The workflow may represent only a small portion of the employee's day. It may also push work into another queue.

• Faster document summarisationWhat must be checked next: Does review time fall, or do staff spend the saving checking AI output?
• Faster quote preparationWhat must be checked next: Are more accurate quotes actually issued and converted?
• Faster customer responsesWhat must be checked next: Are response quality, promises and escalation accuracy maintained?
• Automated CRM updatesWhat must be checked next: Are records more complete, or are incorrect classifications multiplying?
• Meeting-note automationWhat must be checked next: Are decisions and actions reaching the right owners?
• AI-assisted project reportingWhat must be checked next: Is management receiving better information or simply more information?

This is where an experienced AI consulting firm should become more rigorous, not more promotional.

Businesses Should Expect Baseline Measurement Before Deployment

A consultant cannot credibly claim an improvement if nobody measured the original process.

Before a significant AI workflow is introduced, the operating baseline should record items such as:

• Average handling time.
• Total monthly transaction volume.
• Human review minutes.
• Error and rework rates.
• Queue or response time.
• Handoffs between systems.
• Number of exceptions.
• Customer or project escalation rate.
• Cost per completed outcome.
• Staff adoption of the current process.

Without this baseline, a business may know that employees like the AI tool but remain unable to establish whether operations improved.

This is particularly important for small and mid-sized Sydney businesses because they often do not have dedicated transformation teams collecting process data before a technology project begins.

An AI readiness assessment can therefore be more valuable when it documents workflows, data access, operational constraints and measurable starting conditions before implementation rather than merely producing a list of attractive use cases.

The First Scale Problem Is Usually Adoption, Not Model Intelligence

When an AI pilot involves 20 motivated employees, informal support can compensate for a weak implementation.

Someone explains how to phrase the request. Another employee knows where the source document is stored. A project lead notices incorrect output. Users share successful prompts with each other.

That informal system becomes less reliable when deployment spreads through hundreds or thousands of employees.

The organisation then needs repeatable mechanisms for:

• Role-specific training.
• Approved use cases.
• Data handling rules.
• Prompt or agent templates.
• Internal support.
• Usage monitoring.
• Quality review.
• Feedback collection.
• Retirement of unsuccessful workflows.

Microsoft's published EY figures are significant partly because they include adoption metrics, not just licence numbers. High monthly and weekly usage indicates that deployment was being evaluated through actual use as well as software availability.

Businesses assessing AI consulting firms should demand the same distinction.

Provisioning is not adoption.

Licences Should Not Be Confused With Transformation

This becomes especially relevant as enterprise vendors package more AI capability into broader productivity, identity, security and agent platforms.

A business may purchase access to multiple AI functions through software it already uses.

The temptation is to treat wider licence coverage as evidence that the AI strategy is advancing.

A better maturity sequence is:

1. Access: Employees can use the technology.
2. Adoption: Employees actually use it for defined work.
3. Consistency: Successful methods are repeatable across teams.
4. Integration: AI interacts appropriately with business systems and information.
5. Redesign: The surrounding workflow changes because the AI has removed unnecessary work.
6. Measurement: Operational performance demonstrably improves.
7. Governance: The system remains controlled as usage, models and workflows change.

The fifth stage is where consulting value becomes particularly important.

If staff still copy information manually between four systems after an AI assistant produces the first draft, much of the original process remains intact.

Sydney Service Businesses Need Workflow Redesign More Than Enterprise Theatre

The scale of EY's deployment is obviously different from a 40-person strata business, 70-person construction company, 25-person conveyancing practice or growing property-services operator.

The operating principles, however, can still transfer.

Take a renovation enquiry.

A customer may submit an address, photographs, property type, approximate floor area and requested work. An AI system could classify the enquiry and prepare a summary in seconds.

But the commercial workflow may still require:

• Checking whether the project is an apartment or freestanding property.
• Identifying strata access constraints.
• Separating removal, preparation and installation scopes.
• Checking whether a site inspection is required.
• Routing the lead to the correct estimator.
• Recording photographs and correspondence against the same job.
• Confirming scheduling dependencies.
• Escalating unusual substrate or access conditions.

An AI consultant who automates only the first summary has improved one task.

A consultant who understands the entire intake-to-delivery pathway can potentially redesign the queue.

The same distinction applies in professional services. Summarising an email may save two minutes. Correctly routing a new matter, identifying missing information, preserving the source record, applying access controls and creating the right review task can affect the entire service-delivery process.

What Businesses Should Demand From an AI Consulting Firm

The enterprise AI market is making tool access easier. That should raise expectations for consultants rather than lower them.

• Workflow discoveryEvidence the business should expect: Current-state maps based on real work, systems, handoffs and exceptions.
• PrioritisationEvidence the business should expect: Clear reasoning for why one workflow should be automated before another.
• Baseline measurementEvidence the business should expect: Pre-AI time, cost, volume, error and review metrics.
• ArchitectureEvidence the business should expect: Defined applications, data sources, permissions, integrations and boundaries.
• GovernanceEvidence the business should expect: Owners, approval gates, escalation paths and audit requirements.
• AdoptionEvidence the business should expect: Role-based rollout, training, feedback and usage monitoring.
• Quality assuranceEvidence the business should expect: Testing against known examples, exceptions and failure cases.
• Business measurementEvidence the business should expect: Post-deployment comparison against the original baseline.
• Lifecycle managementEvidence the business should expect: Processes for model changes, permissions, incident review and workflow updates.

If the engagement ends when users receive access to the software, the difficult part of AI transformation may not yet have started.

Governance Needs to Follow the Workflow Into Production

AI governance is also moving beyond policy documents.

The Office of the Australian Information Commissioner advises organisations to consider privacy risks when using commercially available AI products, noting that generative AI can create particular risks because of factors including its probabilistic operation and use of large datasets.

Australian Government guidance published for AI adoption now emphasises fit-for-purpose risk management, defined responsibilities and ongoing management of AI throughout its lifecycle.

The NSW AI Assessment Framework is specifically directed at NSW Government agencies rather than being a mandatory framework for every private Sydney company. It is nevertheless a useful reference point for commercial teams because it concentrates attention on matters including fairness, privacy, security, transparency and accountability.

For a private business, this can translate into practical questions:

• Can the AI access customer personal information?
• Does it need access to the entire database?
• Who can change the workflow?
• Can it send external communications?
• Can it create or edit a business record?
• Can it initiate a financial or contractual action?
• What happens when confidence is low?
• Are outputs and actions logged?
• Who investigates an incorrect result?
• How can the AI function be disabled without stopping the underlying operation?

These are implementation questions. They should sit inside the project plan, not in a governance document that nobody revisits after launch.

Agentic AI Makes the Consulting Standard Higher Again

The EY and Microsoft story is also moving beyond conversational assistants.

Microsoft says EY has been applying agentic systems in core operations, including finance, assurance and tax workflows. Microsoft reported substantially faster finance lead times, lower operational costs in the cited finance implementation and major reductions in manual effort for certain document processes. These are Microsoft-reported examples and should be read as case-study outcomes rather than universal performance benchmarks.

Agents change the risk model because software can move from answering a question to performing a sequence.

An agent might:

1. Receive an enquiry.
2. Identify the customer.
3. Retrieve relevant records.
4. Classify the issue.
5. Prepare a response.
6. Update the CRM.
7. Create a task.
8. Trigger a follow-up.

Each step may appear minor.

Together, they form an operational process.

This is why businesses deciding between packaged platforms and custom systems should evaluate the underlying workflow before committing to a technology path. Elyment's build-versus-buy AI framework for Sydney organisations focuses on that choice between licensed capability and systems built around more specific operational requirements.

The Consultant Should Also Tell You What Not to Automate

A mature AI adviser should be willing to leave some processes alone.

Good candidates generally have:

• Repeatable inputs.
• High transaction volume.
• Clear business rules.
• Observable outcomes.
• Manageable consequences when mistakes occur.
• Straightforward human escalation.

Poor early candidates often involve:

• Ambiguous authority.
• Rare but highly consequential decisions.
• Unclear source records.
• Unstructured exceptions.
• Sensitive personal or commercial information without appropriate controls.
• Processes that are already poorly designed before AI is introduced.

Automating a disorganised workflow can make its defects move faster.

The Next Consulting Battleground Is Capacity Reinvestment

The most interesting element of EY's reported productivity story may ultimately be what happened to the saved time.

When AI removes administrative effort, management has a choice.

It can reduce cost. It can increase throughput. It can shorten customer response times. It can increase review depth. It can give technical staff more time for judgement-intensive work. It can accelerate staff development. It can absorb growth without proportional headcount expansion.

Those are strategic choices, but they need to be designed into operations.

A Sydney project-delivery business that recovers 100 staff hours each month could deliberately redirect 40 hours into faster estimating, 30 into project-quality reviews and 30 into client follow-up.

Without that decision, the organisation may still be more efficient at individual tasks while struggling to identify where the benefit went.

That should become part of the AI consulting brief:

When this workflow becomes faster, what will the organisation do with the capacity?

A Practical Enterprise-to-SME Rollout Model

Most NSW businesses will never require a 400,000-person deployment. They can still use the same discipline at a smaller scale.

1. Select one operational queue.
2. Choose an end-to-end process rather than a collection of disconnected AI features.
3. Measure the current state.
4. Capture volume, handling time, delay, rework, errors and human review.
5. Map data and permissions.
6. Identify exactly what the AI must access and deliberately exclude what it does not need.
7. Design the future workflow.
8. Define what disappears, what changes and what remains human-led.
9. Run controlled production tests.
10. Test real examples, difficult cases and known exceptions.
11. Measure adoption separately from performance.
12. High usage does not automatically mean high value.
13. Measure business outcomes.
14. Compare the new process with the baseline.
15. Decide where saved capacity goes.
16. Make productivity conversion an explicit management decision.
17. Scale only after the process is stable.
18. Extend patterns that have evidence behind them rather than multiplying experiments.

What EY's Expansion Changes for the AI Consulting Market

The EY deployment does not prove that every organisation will achieve a 15 per cent productivity improvement by adopting Microsoft AI.

It demonstrates something more useful.

Large organisations are now moving beyond the phase in which AI success is defined by experimentation alone. Microsoft's account of EY combines deployment scale, adoption measures, reported productivity outcomes and workflow-level automation.

That raises the standard for the consulting market.

Businesses should expect consultants to understand process design as well as prompts, information governance as well as integrations, employee adoption as well as software configuration, and operational measurement as well as demonstrations.

The organisations most likely to capture lasting value will not necessarily be those with the most AI licences.

They will be those that can show, workflow by workflow, what changed, what improved, what remained under human control and where the resulting capacity was reinvested.

AI, WORKFLOW AND PROJECT DELIVERY REVIEW

Review the Operating Model Before Scaling the Technology

Elyment works with Sydney and NSW organisations to review operational workflows, AI readiness, data access, approval controls, implementation sequencing and measurable project outcomes before automation is expanded across the business.

Request a Project and AI Operations Review

The Operational Conclusion

EY's expansion across more than 400,000 people is significant because it illustrates what happens after an AI pilot begins producing measurable value.

The next problem is scale.

Scale requires more than technology procurement. It requires workflow ownership, staff adoption, permissions, measurement, quality controls, exception handling and a management decision about where recovered capacity should go.

For Sydney and NSW businesses choosing an AI consulting partner, this creates a clearer standard.

Do not ask only what the consultant can build.

Ask what operating process will exist when the build is finished, how the improvement will be measured and who will remain accountable when the AI becomes part of everyday work.

Sources and Further Reading

• EY: EY and Microsoft global AI initiative
• Microsoft: From AI pilots to enterprise impact
• Microsoft: FY26 AI transformation review
• Office of the Australian Information Commissioner: Privacy and commercially available AI products
• Australian Government: Guidance for AI adoption
• Digital NSW: NSW AI Assessment Framework
• Elyment: AI Consulting Sydney
• Elyment: AI Readiness Assessment Sydney
• Elyment: Build vs Buy AI Sydney

Editorial note: The 15 per cent productivity figure discussed above is the outcome reported by EY and Microsoft in relation to EY's earlier Copilot deployment. It should not be interpreted as a guaranteed productivity outcome for other organisations. AI performance depends on the workflow, implementation, data, governance, adoption and method used to measure results.