Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsEY Global Chief Innovation Officer Joe Depa’s warning is less about robots suddenly replacing software agents than about organizations facing several changes at once: generative AI, agentic AI, physical AI, quantum computing, legacy-system replacement and workforce retraining. For finance leaders, his most useful advice is practical: choose a business use case, check its data and infrastructure, test it safely, and define the outcome or action it must deliver. The technology matters, but so do process design, employee adoption and controls.
What Depa said—and what the warning means
In a January 15, 2026, Computerworld interview, Agam Shah spoke with Joe Depa, EY’s Global Chief Innovation Officer. Depa described a shift from generative AI toward agentic AI, physical AI and, eventually, quantum computing. He also argued that companies can struggle to turn AI experiments into business value when they lack a clear outcome, usable data, suitable infrastructure, a safe place to test, or a plan for changing how employees work.
Depa’s sequence—use case, data, simulation, outcome or action—is a useful way to evaluate a project. It is not proof that every organization needs an autonomous agent, robot or quantum-computing initiative now. The interview is an executive’s perspective, not a quantified study: it does not supply deployment figures, measured returns, error rates or timelines for broad adoption.
There is also a commercial perspective to keep in view. EY sells consulting and transformation services, and its biography of Depa describes his work leading AI, data and innovation strategies. His comments are relevant executive insight, but not neutral market analysis.
Recommended Free Tools
#1 Best Overall
How agentic AI differs from a chatbot
“Agentic AI” is used broadly, rather than as a single standardized technical category. It generally describes software that can pursue a goal over multiple steps, retrieve information, use tools or enterprise systems, make decisions within delegated limits, and take actions. The degree of autonomy varies: a product described as an agent may be little more than workflow automation, while another may plan and act across several systems.
- Generative AI produces text, images, code or analysis in response to a prompt.
- An assistant or copilot helps a person complete a task, with the person generally retaining control of the work.
- An agent may plan and carry out parts of a workflow, including tool use, under permissions set by its organization.
- Physical AI connects AI capabilities to machines or other systems that perceive and act in the physical world.
The key change is not merely that an agent can answer a question. It may be able to act on that answer. If a system can update a record, send a supplier request or initiate a transaction, an incorrect decision can have operational consequences. Multi-step errors can compound, especially when data is stale or conflicting, or when the agent’s access is broader than its task requires. More autonomy does not, by itself, make a system more reliable.
Before granting execution rights, an organization needs to decide who owns the agent, which data and tools it can access, which actions require approval, how its decisions are logged, and how staff can stop or reverse an action. Monitoring should also account for repeated tool calls, unexpected retries and changes in model or connector behavior. Those details affect both risk and cost: a workflow that repeatedly calls models, databases or APIs may cost more than a simple demonstration suggests.
Where an early enterprise agent may fit
Depa named finance, procurement, human resources and software development as areas with agentic-AI opportunities. These functions often include recurring workflows and defined records or rules, which can make a task easier to evaluate than an open-ended strategic decision. That does not make every process in those departments a good candidate.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
- Finance: An agent might gather information for invoice exceptions or prepare a reconciliation item for review. A first deployment should not be assumed to have authority to approve payments or change financial records without controls.
- Procurement: It could check purchase-order status or draft a supplier follow-up, while a buyer reviews messages or commitments that could affect a relationship or contract.
- Human resources: It might route employee-service requests or locate policy information. Decisions affecting an individual’s employment require particular care and should not be treated as routine administrative actions.
- Software development: It could help with test generation, documentation or issue triage. A developer still needs to review changes and assess their effects before they enter production.
A promising first workflow has a clear start and finish, repeatable steps, data that can be accessed and checked, a measurable baseline, and a responsible business owner. Its permissions should be narrow; mistakes should be reviewable or reversible; and staff need a clear way to escalate exceptions. A process that is constantly changing, lacks an accountable owner or depends on sparse and contradictory information is a poor first choice.
Use four checks to move beyond a demo
Depa’s framework puts a business need ahead of a technology experiment. Apply it as four questions before expanding a pilot.
1. What specific use case is being improved?
Name the task, its owner and the bottleneck. Define a baseline—such as completion time, error rate or cost per completed task—and decide what improvement would justify further investment. A broad goal such as “use AI in finance” is not specific enough to guide design or evaluation.
2. Can the system rely on the right data?
Map the data sources and check ownership, access permissions, freshness, metadata, duplicates and conflicting records. Finance workflows may touch sensitive personal or financial information, so access should be limited to what the task requires. Consider how the agent will connect to systems such as ERP, CRM, HR, ticketing and document repositories, and how changes to those connections will be monitored. Better data cannot rescue a workflow that has little business value.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
3. Can it be tested in a controlled environment?
Use a sandbox or simulation with representative cases, including unusual and failure scenarios, before permitting live actions. Compare outputs against an evaluation set and test whether the agent handles missing information, contradictory records and denied access safely. Start with read-only access or recommendations for human approval; grant limited execution authority only after the behavior is understood.
4. What outcome or action will count as success?
Specify the result the process should produce, who can approve it, and what the agent must do when it is uncertain or wrong. Track quality and exception rates as well as speed. Include the cost of integration, model use, monitoring and human review; a faster workflow is not necessarily a better one if it shifts work to reviewers or increases compliance and customer-service problems.
Why adoption and process redesign matter
Depa’s point about change management is especially relevant where employees must rely on, review or work alongside an AI system. A technically capable tool may be ignored if staff do not trust it, have incentives to keep using the old process, fear losing professional judgment or cannot tell who is accountable for an AI-assisted decision. Training that covers only the normal case leaves people unprepared for exceptions; a new tool can also add review work rather than remove it.
Depa used robotic surgery as an illustration of technology whose value depends on training and adoption as well as capability. That example should not be read as evidence that robotic surgery is universally safer or better; the interview does not establish such a medical conclusion. The broader organizational lesson is that deployment requires people to learn new roles and workflows.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Rank #4
- Bring process owners and affected employees into design before choosing a product.
- Set out which decisions remain human and who is accountable for them.
- Train for normal operation, uncertainty, failure and escalation.
- Give users a formal way to report errors, override recommendations and pause unsafe behavior.
- Measure whether the workflow is adopted and how often it produces exceptions, not just how many people log in.
Physical AI raises different risks
Physical AI applies AI to systems that perceive and act in the physical environment. It can include industrial robots, warehouse automation, autonomous vehicles, drones, medical robotics and smart manufacturing equipment. These systems are not all alike, and many commercial robots remain specialized, constrained or supervised rather than broadly autonomous.
A software agent can produce a faulty document or update the wrong record; a physical system can also damage equipment or injure someone. Deployment therefore has to account for safety, supervision, maintenance, latency, environmental conditions and liability alongside data and software controls. Depa’s forecast that physical AI is part of what organizations should prepare for is an executive outlook, not a timetable for mass adoption.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Quantum computing is a longer-horizon consideration
Quantum computing is not a general-purpose replacement for classical cloud computing. Its potential is associated with selected problem classes, such as some optimization, simulation, chemistry, finance and security applications; most organizations should not assume that a quantum system will improve an ordinary business workload.
For a company with a plausible use case, preparation can mean identifying the problem, developing technical expertise, monitoring hardware and software progress, and comparing any quantum approach with classical alternatives. Access through a cloud or partner ecosystem is a more realistic route for many organizations than building a quantum computer internally. Depa recommends partnering rather than trying to build one in-house, and IBM’s quantum product material describes access and services rather than ordinary enterprise hardware procurement. Neither fact establishes that quantum computing is ready for general business use.
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
What changes for consulting—and what it does not prove
Depa argues that consulting work will shift toward people who can identify worthwhile problems, deploy AI, connect systems and vendors, manage regulatory and compliance risk, and redesign processes so employees can adopt new tools. That is a claim about changing demand for skills, not evidence that consultants will disappear or that every company needs outside advisers.
The interview also leaves important implementation questions open. It provides no quantified return on investment, customer case study, production error rate, cost model or specific accountability framework. It does not compare agents with simpler options such as rules-based automation, search or an existing process redesign. For each proposed project, buyers should make those comparisons themselves rather than treat a successful demonstration as proof of operational value.
A practical readiness plan for finance and operations teams
- Choose one measurable bottleneck. Select a bounded task with an owner, a clear baseline and a meaningful target.
- Map the workflow. Record the inputs, decisions, systems, handoffs and failure points; identify which steps must remain human.
- Check data and access. Verify that the required information is current and authorized, and keep the agent’s permissions limited to its task.
- Test before execution. Use representative cases in a sandbox, including exceptions, and establish evaluation criteria before reviewing results.
- Introduce it gradually. Begin read-only, then recommendation mode; consider bounded execution only for low-risk actions that can be reviewed or reversed.
- Assign controls and ownership. Name the accountable business owner, define approvals and escalation, retain usable logs, and plan how to stop or roll back an action.
- Compare full costs and results. Include integration, model and tool usage, monitoring and human review, then assess whether the process improved the chosen business measure.
- Expand only on evidence. If the pilot cannot show a meaningful improvement without unacceptable risk or extra review burden, revise the workflow or stop it rather than scaling a demo.
The immediate task is not to buy every emerging technology. It is to make one valuable workflow measurable, testable and governable. Physical AI and quantum computing may matter to some organizations over different time horizons, but neither removes the need to execute well on today’s processes, data and workforce transition.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
