Main Point
A successful conversation with AI does not, by itself, show that someone is ready to work with a real client. A well-designed simulation can, however, create opportunities to practice asking questions, responding to objections and explaining decisions before a mistake consumes someone else’s time, money or trust. The company experiences reported by The Wall Street Journal illustrate that possibility, while leaving important questions about the strength and reach of the evidence. The central choice is whether AI completes the learner’s work or creates a setting in which the learner must do the work. [1]
Consider an illustrative meeting. A junior marketer presents a campaign to a small service business. The presentation is polished, the audience is defined and the expected results sound persuasive. Then the client asks why the company should spend more on advertising when its employees failed to respond promptly to inquiries from the previous campaign. The marketer continues explaining the benefits of the new advertisements. The proposal may be well presented, but the conversation has moved from advertising to service operations, and the marketer has not moved with it.
Now suppose the client is played by AI in a learning exercise. The conversation can be paused, the missed issue examined and the meeting attempted again. This time, the marketer starts by asking what happens after an inquiry arrives. The value of the exercise is not that it supplies a model answer. It creates an opportunity to connect a decision with its consequences without making a real client’s interests the testing ground for a beginner’s first attempt.
When work becomes easier, learning does not happen automatically
In the September 26–27, 2026 print edition of The Wall Street Journal, Allison Pohle describes a problem companies call “experience starvation.” Automation and organizational restructuring are removing some of the repetitive assignments through which junior employees used to recognize professional patterns. At the same time, beginners may be asked to review AI-generated work before they have accumulated the experience that makes such review meaningful. The issue is not merely whether an entry-level task still exists; it is where the judgment once developed through that task will now come from. [1]
This does not mean universities or employers should preserve every routine assignment unchanged. Manually transferring figures into a report and noticing that those figures contradict one another are different learning activities. The useful question is whether the redesigned workflow still gives the beginner opportunities to request information, identify mistakes and compare alternatives. Saving time can be valuable without automatically developing the ability to make a sound decision.
That distinction changes the assignment itself. Instead of asking a student to prepare an excellent proposal, an instructor might ask the student to determine whether enough information exists to prepare any proposal responsibly. AI could complete the first assignment on the surface by producing a persuasive document. The second requires the learner to choose questions, decide what needs verification and recognize commitments that would be premature. A practice partner becomes useful precisely because it can preserve the difficulty that the learner needs to work through.
What the company examples actually show
According to the Journal, EY uses simulations to let employees practice requesting information from clients and addressing missed deadlines with colleagues. In one exercise, AI plays a difficult client who challenges a recommendation; the employee responds and receives feedback. The article also describes Cisco setting aside time for customer-pitch role-play. It does not establish that this particular Cisco example uses AI. That distinction matters: role-play is the broader training method, while an AI counterpart is one possible way of organizing it. [1]
A more specific episode involves a Stanford Law School laboratory and participants in Dechert’s summer associate program. The Journal reports that one group had access to a negotiation simulator and spent two hours a week interacting with AI counterparts, while another group did not have access. Three of the four finalists in an end-of-summer mock negotiation came from the simulation group. In the same report, researcher Megan Ma cautions against drawing large conclusions from a small group. [1]
The result should not be made to carry more weight than it can support. The distribution of finalists does not tell us how much every participant improved, how the groups differed at the outset or how the participants later performed with actual clients. It is an informative training case, not a universal effectiveness measure. The important boundary in the source is that simulated practice may enrich professional development without replacing observation of experienced practitioners and participation in real work. [1]
A supplementary public source provides a related company example. In its April 2026 announcement, Cognizant describes Skillspring as AI-enabled learning aligned with roles and projects. The announcement documents the company’s stated approach; it is not an independent evaluation of learning gains. [4]
A practice partner is not an answer service
Ethan and Lilach Mollick’s conceptual paper offers a useful distinction here: AI as a simulator is one educational role among several, with active human participation remaining important. Their framework is a pedagogical proposal, not an experiment establishing the success of every application. [2]
The practical approach proposed in this article starts with a particular skill. “Client communication” is too broad to tell an instructor what to look for. Clarifying an ambiguous request, responding to criticism without becoming defensive, or defining responsibilities after a meeting are more usable objectives. Once the target is clear, both instructor and learner can distinguish an enjoyable conversation from behavior that demonstrates the intended skill.
A University of Melbourne teaching guide illustrates this distinction with AI playing a client while a student asks questions to gather project information. It also warns that a model may invent details or lose the thread of the scenario. A convincing conversation is therefore not automatically a dependable learning environment. [3]
In the illustrative marketing exercise, the instructor could establish the business’s service, the previous campaign’s experience and the operational obstacle in advance. The student would receive only an initial brief. The AI client would reveal additional information in response to relevant questions, within the agreed scenario. A learner who never asks why earlier inquiries went unanswered should not be handed a complete diagnosis anyway. Otherwise, the system quietly performs the very task it was supposed to help the student practice.
Difficulty can then vary for a reason. One client might understand the problem clearly; another might confuse a business objective with a preferred solution; a third might lack agreement within the team. These variations should serve a learning purpose. An abusive or permanently irrational counterpart does not automatically make an exercise more realistic. The aim is to examine listening, inquiry and judgment, rather than simply measure how long someone tolerates an artificial conflict.
Turning a conversation into useful practice
The following design is this article’s practical proposal, not a procedure claimed to be used unchanged by the companies above. Before the conversation, the learner briefly identifies what needs to be established and which commitments cannot responsibly be made without more information. The learner then speaks with the AI client. The system does not write the learner’s replies during the exercise. Assistance can be requested separately, but its use should be visible when the performance is assessed.
Afterward, an instructor might begin with a question: What did you learn about the client’s problem that you did not know at the start? The student then identifies one effective choice and one moment worth revisiting. AI feedback and the instructor’s observations can follow. This sequence is intended to give the learner practice in interpreting their own experience, rather than turning them into a passive recipient of an apparently authoritative assessment.
Feedback needs to refer to observable decisions. “Be more persuasive” offers less guidance than pointing out that the learner promised campaign results before checking the service team’s capacity to handle inquiries. Effective restraint also deserves recognition. A student may correctly explain that further information is needed before a final recommendation can be made. Finishing the meeting without a signed agreement is not necessarily failure; avoiding an unsupported commitment can be the professionally sound outcome.
A second attempt should have a defined purpose. The student can revisit the episode in which a question was missed or a conclusion reached too quickly. A different case should then require the same underlying skill in another setting. Without that change, repetition may produce a memorized conversation rather than adaptable competence. The important question is whether the learner can carry the logic of inquiry into an unfamiliar problem, not whether the familiar script now sounds smoother.
The instructor remains responsible for checking the scenario, judging the feedback and deciding what the learner is ready to attempt next. AI can play a client in one phase and support feedback in another, but that role change should be explicit. Using several agents does not remove the need for scrutiny. Agreement between their evaluations is not independent confirmation when they may be repeating the same mistaken assumption.
A high simulation score is not client readiness
The most misleading shortcut would be to equate the number of completed exercises with professional readiness. A completed conversation shows participation. An automated score shows what a particular scoring system concluded. To establish what the student has learned, an instructor needs additional evidence: can the learner identify important information in a new case, explain a choice and recognize when assistance is necessary?
One possible check is a short meeting with a person playing an unfamiliar but carefully prepared client. The evaluator should know which behaviors to observe, and the student should know what assistance is allowed. Supervised real work can follow: preparing questions, attending a meeting and documenting agreed next steps, for example. Independent responsibility should expand gradually rather than arrive automatically after a single simulator score.
There are also boundaries to what a conversation can represent. A text-based exercise cannot assess every nonverbal signal, a team’s accumulated history or the actual delivery of a promise. An AI client’s agreement is not a forecast of customer choice. It does not establish that Georgian customers will buy a product, accept a price or respond successfully to a campaign. The exercise examines a person’s behavior within a defined scenario; it does not measure market demand.
This article also proposes clear safeguards around the practice environment. Cases should be fictional or appropriately de-identified, the training space should be separated from live work systems, and only necessary records should be retained. An agent playing a client does not need permission to send actual emails, confirm orders or make commitments on a company’s behalf. Learners should know who can see their conversations and how those records affect assessment, with a way to stop an exercise when sensitive issues arise.
BTU Researchers’ Assessment
The central proposition of this analysis prepared for BTU is that AI’s educational value may increase when universities use it to organize better practice, rather than merely produce better answers. An answer service makes an assignment easier to finish. A practice environment gives a person opportunities to perform the professional actions they need to learn. This is an analytical design proposition developed for the article, not a report of learning gains already measured at BTU.
A useful principle for evaluating such a system is targeted practice: how often does the learner make a decision, receive specific feedback and test an adjustment in a new case? Repetition alone does not meet that standard. The evidence needs to show what changes in the person’s behavior and whether the change helps in more demanding, less predictable circumstances. That is what would distinguish a learning environment from an archive of interesting conversations.
Universities could embed progressively more demanding exercises within subject-specific courses. A marketing student clarifies a client’s request; a management student addresses a delayed project; a technology student turns an ambiguous user need into a workable assignment. The overall pattern may be shared, but the professional criteria belong to the discipline. A single simulator that rewards politeness and fluent expression across every field cannot substitute for those differences.
Key Findings
- AI can act as a practice counterpart while the learner remains responsible for gathering information, making choices and explaining decisions. [1–3]
- The Journal’s cases illustrate a practical opportunity, but a small group’s results do not establish effectiveness for all students or professions. [1]
- The unit of the proposed learning design is a decision, specific feedback and a further attempt—not merely a completed conversation.
- Success in simulation should be checked in an unfamiliar task and then in supervised real work.
- A sound professional outcome may involve postponing a promise, requesting more information or asking for help at the appropriate moment.
- Georgian adaptation should address the substance and assessment of work; an AI client’s responses cannot substitute for research with real customers.
Why This Matters for Georgia
For Georgian universities and employers, the practical opportunity is a better connection between study and the first professional assignment. Students could encounter important beginner mistakes in a training setting, while mentors would have specific episodes to discuss rather than only a general impression of preparedness. None of this requires a claim that AI has replaced internships. A good simulation should instead prepare students to participate more productively in an internship.
Local adaptation is more than translation into Georgian. The case needs a defined service, clear decision authority, a record of what the fictional team has promised and an account of the information still missing. Practitioners from different sectors could help create these cases without disclosing their real clients’ confidential information. This article does not use representative Georgian outcome data. It proposes a direction for adaptation rather than reporting a result already established locally.
A university pilot could assess performance against the same professional criteria before and after training, account for mentor time and include a new task to test whether learning transfers. A meaningful comparison should distinguish the contribution of AI practice from additional practice time and human feedback. Otherwise, it will remain unclear whether a benefit came from the technology, from more deliberate study or from combining them appropriately.
Conclusion
The purpose of practicing before meeting a real client is not to manufacture an employee who never makes a mistake. It is to develop someone who listens more carefully, asks consequential questions, recognizes limits and does not conceal uncertainty behind confidence. AI can be a useful partner when it creates the difficulty a learner needs to engage with, rather than removing that difficulty on the learner’s behalf. The final evidence of learning is what the person can do at the next meeting, when the scenario changes and responsibility becomes real.
Data and Main Sources
[1] Allison Pohle – “We All Hate Busy Work. Now It’s Disappearing-And We Really Miss It.” The Wall Street Journal, 26–27 September 2026, pp. B1 and B5.
[3] University of Melbourne, Melbourne School of Design – GenAI and Self-Directed Learning.
[4] Cognizant – Skillspring launch announcement. 21 April 2026.
The main source is the supplied print issue. Supplementary sources provide conceptual and institutional context; they do not replace the Journal’s reporting or validate the effectiveness of the model proposed here.
This material is analytical and educational in nature. It does not constitute specific technology, legal, financial or organization-specific professional advice.
Prepared by the academic team of Business and Technology University and the BTUAI Research Team, Tbilisi, Georgia.