When AI takes over foundational work, organizations need a new way to develop professional judgment. That is the most important management lesson from PwC’s announced certification program for more than 85,000 employees. This is not another course on prompting. It addresses a deeper question: how do people develop expertise when some of the experience that built it is being automated?
For enterprises, this is not simply an HR issue. It affects service quality, risk management, and the ability to reduce workloads without eroding the knowledge that keeps the business running.
- 85,000+ employees are expected to participate.
- Approximately six months is the planned duration of the certification.
- June 2027 is the target for participants to complete certification.
According to the program details, participants will come from PwC’s offices in the United States and Mexico and its acceleration centers. The launch is planned for December, with training that combines an initial assessment, learning, simulations, and a personal AI coach. The scale matters, but the choice of what to teach matters more.
Automation saves work. It can also remove learning.
A junior employee learning to prepare a financial analysis was not just learning to operate a spreadsheet. They encountered a number that did not reconcile, tracked down a missing document, discovered that an accounting definition changed the picture, and received corrections from a manager. The repetition could be tedious, but it also built understanding.
A model can now produce a draft analysis quickly. That is genuine progress. There is no reason to preserve manual work simply because that is how previous generations learned. The problem starts when an organization eliminates a task without identifying the learning embedded in it.
In an interview with Business Insider, Yolanda Seals-Coffield, PwC’s US chief people and inclusion officer, explained that some of the repetition that developed critical thinking and judgment would no longer be there. The program is intended to provide a structured response.
We do not need to preserve unnecessary work. We need to preserve, and improve, the learning mechanism hidden inside it.
At a services company, for example, an agent that summarizes customer inquiries could save hours. But if employees stop engaging with the underlying material, they may lose the ability to spot an emerging pattern of complaints. A local efficiency gain can come at the expense of the organization’s diagnostic capability.
What should enterprises teach beyond prompting?
Communicating effectively with models is a foundational skill. Employees need to define tasks, provide context, and ask for useful outputs. But that is only one layer. Someone can write an excellent prompt and still fail to recognize that the resulting recommendation is dangerous for the business.
Critical thinking with AI should focus on behaviors that people can observe and practice:
- Distinguish facts, assumptions, and inferences, even when all three are expressed with equal confidence.
- Check whether a source is current and relevant to the client, market, and regulatory context.
- Identify what information is missing before making a decision.
- Compare alternatives and explain the trade-offs of each.
- Recognize when to stop a process and escalate to someone with the authority to decide.
PwC plans to develop its courses with professors and experienced employees. That is the right combination: academic expertise helps establish a framework for reasoning and assessment, while professional experience connects it to situations without a neat answer at the end of the exercise. AI is multidisciplinary. Knowing a tool is no substitute for understanding the profession and the business process.
An AI coach can challenge employees or validate their mistakes
A personal coach can turn one-off training into ongoing practice. It can simulate a skeptical client, critique a proposal, or ask an employee to justify a recommendation. AI’s availability between training sessions offers a real advantage here.
But there is a trap. If the same type of model produces the analysis and then confirms that it is sound, the result can be a loop of confidence without verification. Persuasive feedback is not necessarily correct feedback.
Watch out: An AI coach is not the final professional authority. Its role is to challenge reasoning and expose missing information, not automatically certify that a decision is correct. Coaching quality must be tested against real cases, professional standards, and human oversight.
Practice should therefore use cases approved by subject-matter experts, with clear criteria for a good answer. Using real client information also requires an approved environment, minimization of sensitive data, and checks on access permissions and data retention. An excellent learning system is not necessarily approved to handle every business document.
Human in the loop does not mean signing off on every output
The usual response to risk is to require human review of every action. In a sensitive process, that may be justified. As a blanket operating model, however, it can eliminate much of the economic benefit.
The goal is to let a professional oversee a larger volume of work without losing control of consequential decisions. That requires defining what an agent may do, which exceptions go to a person, and how to detect errors that were missed in real time.
- Human review: Burdensome oversight treats every output the same; scalable oversight allocates review according to risk and exceptions.
- Operating boundaries: Burdensome oversight relies on general instructions; scalable oversight defines permissions and decision authority.
- Quality control: Burdensome oversight relies on individual approvals; scalable oversight uses sampling and failure analysis.
- Learning: Burdensome oversight fixes the case; scalable oversight improves the process and the training.
This is not an argument for fully automating sensitive processes. It distinguishes risk-based controls from a review workload that does not necessarily improve quality. A material financial decision, for example, may require human approval, while routine document classification may be checked through sampling and additional controls.
Turn the lesson into an operating plan
Enterprises do not need to copy a global firm’s program. They need to choose a process where AI is already changing the work and connect employee development to automation design.
- Map the learning being removed: Identify which tasks are being automated and what professional knowledge employees previously gained through them.
- Build practice cases: Have subject-matter experts and learning specialists develop scenarios involving missing information, flawed assumptions, and plausible alternatives.
- Set agent boundaries: Define permissions, stop conditions, escalation paths, and decision logging.
- Measure changes in the work: Assess decision quality, handling time, and oversight workload against the baseline.
The connection between these steps is critical. If training is detached from the process, employees will practice well-crafted exercises that bear little relation to their work. If an agent is developed without the professionals who understand the domain, it may accelerate a process whose business logic was never properly defined.
Organizations need to advance on two tracks at once: AI literacy for employees and an internal capability to build and manage agents. A personal tool often requires people to change their working habits. An agent integrated into an existing system may require less change from the user, but more infrastructure responsibility from IT. Training cannot replace infrastructure, and infrastructure does not create judgment.
Certificates are not the measure of success
PwC plans to incorporate skills assessment into performance reviews rather than rely solely on a pass-or-fail test. That is a sound direction, provided the assessment measures professional behavior rather than general impressions.
An enterprise can assess whether employees identify flawed assumptions, support recommendations with evidence, and escalate appropriate cases. Alongside those measures, it should track total handling time, rework, and the cost of oversight. A faster first draft is not a success if a manager then spends hours correcting it.
When selecting a training provider, organizations should also ask for more than an impressive demonstration: relevant education, experience with business processes, an assessment methodology, and evidence of implementation capability. Familiarity with the latest model is not enough to design professional training.
The lesson from PwC’s move is that automation and the development of expertise cannot be separated. Leaders who count only the work saved overlook the knowledge that must be rebuilt. Those who connect training, agents, and oversight can improve output without making the organization dependent on answers that nobody knows how to evaluate.