AI and the Future of Conversation Skills

Generative AI is very good at removing small pieces of friction from knowledge work. It can summarise a document, restructure a proposal, draft an email, turn notes into a presentation, compare options, clean up language and help someone get from blank page to acceptable first version much faster.

That is real productivity.

But a new Microsoft Research study raises a second question that deserves more attention: as AI changes the balance of how people work, what happens to the human capabilities that are built through interaction rather than documentation?

The researchers analysed digital trace data from Microsoft 365 across multiple large international companies. Among heavy AI users, defined in the study as people who used the AI system more than 100 times during a 20-week post-adoption period, productivity-oriented actions increased by 21.2 percent while communication-oriented actions increased by 7.1 percent. Both increased, but not at the same rate. The overall balance shifted toward more individual, documentation-focused work.

The researchers do not claim that AI is making people worse at communicating. Neither should we. The study measures activity patterns, not communication skill.

The interesting question is what follows if this pattern continues.

AI removes work from the easy conversations first

A great deal of workplace communication is not really conversation in the demanding sense. It is transmission.

“Can you summarise this?”

“Can you turn these notes into an update?”

“Can you draft a response?”

“Can you pull the main points from this meeting?”

AI is extremely good at absorbing some of that work. That can be a welcome improvement. Fewer unnecessary status exchanges and less time spent formatting information is not a loss.

But the conversations that remain are often the ones that were difficult to automate in the first place.

The supplier who wants 12 percent more.

The stakeholder who refuses the specification change.

The manager who rejects the proposal.

The colleague who feels bypassed.

The customer who is escalating.

The employee who needs difficult feedback.

The negotiation where both sides understand the facts but disagree about what should happen next.

These are not primarily information-transfer problems. They are live coordination problems under uncertainty, pressure, competing interests and social consequence.

If AI removes more routine written work, the average human conversation may actually become more demanding, not less.

There is a difference between communicating more efficiently and becoming better at conversation

A manager who uses AI to draft a difficult email may communicate more clearly. A buyer who uses AI to prepare negotiation arguments may arrive better prepared. A salesperson who asks AI to anticipate objections may have better material.

All of those things can improve performance.

But none of them automatically builds the ability to handle the moment when the other person responds in a way the preparation did not predict.

Conversation is a production skill. It requires retrieval, judgment and adaptation in real time.

You have to listen while deciding what matters. You have to notice when the frame changes. You have to formulate a response before you know exactly how it will land. You have to regulate your own reaction while reading the other person's. You sometimes have to say no without destroying cooperation. You sometimes have to challenge someone senior. You sometimes have to remain silent when every instinct tells you to fill the gap.

Reading a strong AI-generated response is not the same mental act as producing one under pressure.

That distinction becomes more important as the generated response gets better.

The risk is not that people stop talking

The simplistic version of this concern would be that AI makes people communicate less and therefore social skills decline.

The Microsoft data does not support that claim. Communication actions still increased among the heavy users in the study. What changed was the balance. Productivity-oriented activity increased much faster.

The more useful concern is about practice distribution.

Skills become reliable through use. If more analytical, drafting and coordination work is delegated to AI, people may get fewer repetitions of some lower-stakes interpersonal tasks while still being expected to perform well in the high-stakes ones.

That creates an unusual capability problem.

We may become more efficient at preparing for conversations at exactly the same time that we get fewer opportunities to practise the live behaviour required inside them.

For experienced professionals, this may be manageable because they already have a large library of patterns to draw from. For people earlier in their careers, the effect could be more significant. Many professional instincts were historically built through hundreds of small interactions: asking the awkward question, explaining a decision, resolving a misunderstanding, pushing back on a request, watching a senior colleague handle resistance.

If part of that interaction layer becomes asynchronous or AI-mediated, organisations may need to become more deliberate about where the practice comes from instead.

Procurement is a useful example

Procurement is becoming increasingly AI-assisted.

AI can help research suppliers, analyse spend, compare contracts, identify cost drivers, structure negotiation preparation, draft RFQs, summarise meetings and generate counterarguments.

Much of that is valuable, and there is no good reason to preserve manual work simply because it used to be done manually.

But the commercial result still often depends on a spoken interaction.

A supplier rejects the cost model.

The buyer asks for evidence.

The supplier escalates to continuity risk.

The buyer has to decide whether the risk is real, whether to challenge it and how strongly.

The supplier offers movement in exchange for volume.

The buyer has to recognise whether the trade is actually valuable.

There is no perfect prompt that can remove the need for judgment in that sequence while the conversation is happening.

AI can make the buyer better informed. It can support the decision. But if the goal is to improve the buyer's capability, someone still has to practise being the person in the conversation.

The conversations left behind are often the ones where relationships matter most

There is another reason this shift deserves attention.

Routine communication can often be made more efficient without much relational cost. Nobody needs a deeply human experience when requesting a meeting summary.

Difficult conversations are different.

Negotiation, feedback, conflict, escalation and persuasion all involve questions of trust, status, fairness, intent and identity. People do not react only to the information being exchanged. They react to how they are treated while the disagreement is happening.

That is why replacing these interactions with perfectly written AI text is not necessarily equivalent to handling them well.

A message can be technically excellent and still avoid the conversation that needed to happen.

This matters especially in organisations that already default to email or chat when the topic becomes uncomfortable. AI can make avoidance more polished. It can produce a diplomatic paragraph that postpones a disagreement rather than resolving it.

Efficiency and avoidance can look surprisingly similar in the output.

Saved time does not automatically become development time

The capability problem becomes sharper when we ask what organisations actually do with the capacity AI creates.

Culture Amp's first AI at Work benchmark, published on 17 September 2026, draws on surveys covering roughly 112,000 employees. Seventy-one percent said AI made them feel more productive. Yet reported workload barely changed with intensity of AI use: 72% of power users said their workload was reasonable, compared with 69% of non-users. At the same time, 85% said their organisation encouraged AI experimentation, while only 58% said leaders had clearly explained how AI would help achieve company goals.

Those are employee perceptions, not measured hours or output, and Culture Amp notes that the participating organisations are likely more AI-engaged than the wider market. Still, the pattern makes a useful point. Capacity created by AI does not arrive with instructions telling an organisation to spend some of it on learning.

American Management Association data from 1,088 professionals makes the allocation problem more concrete. Forty-one percent of managers said AI had increased their expectations for how much work employees should complete in a given period. Managers also reported more time reviewing AI-assisted work and using AI-generated information in decisions. Employees, meanwhile, continued to judge managerial engagement through visible behaviours such as regular communication, clear direction, listening and support for growth.

That creates a very ordinary route from productivity gain to capability loss. Twenty minutes disappear from drafting, and the twenty minutes do not become coaching, observation or practice. They become another task.

Talogy has separately reported that 78% of HR and talent leaders in its research are concerned about the long-term loss of critical leadership skills as AI absorbs work that historically helped people develop. That is evidence of employer concern, not evidence that the loss has already happened. But the concern is coherent with the mechanism: if lower-stakes repetitions disappear and the saved capacity is immediately converted into throughput, development does not happen automatically in the space left behind.

Procurement now has a specific deskilling warning

On 9 September 2026, Gartner published research aimed specifically at CPOs under the title Prevent AI-Related Skill Loss to Maintain Procurement Cost Goals. Its public abstract makes a narrow but notable claim: overreliance on AI can unintentionally hinder the development of critical human skills needed for strategic value.

The public research abstract does not provide enough methodological detail to treat that as measured proof that procurement skills are already declining. It should not be cited that way.

What has changed is that skill loss is no longer only a general workplace hypothesis being inferred from changing patterns of work. It is now being framed explicitly as a procurement management risk by a major procurement research provider.

That is particularly relevant because procurement has a strong incentive to automate preparation work. Supplier research, spend analysis, contract review, meeting summaries and negotiation preparation are all obvious candidates. The question is not whether those tasks should stay manual. It is whether organisations are deliberately replacing the practice that used to sit inside them when the human capability still matters.

A buyer who no longer spends an hour building a first draft of a negotiation brief has gained an hour. If that hour becomes another task, the productivity gain is real and the development opportunity is gone. If part of it becomes coached review, live rehearsal or reflection on a difficult supplier conversation, the organisation has converted some of the efficiency into capability.

Gartner's warning therefore sharpens the issue rather than proving the outcome. Procurement leaders need to decide which repetitions are obsolete work and which repetitions were quietly doing developmental work at the same time.

The AI adoption gap is becoming a capability-design problem

A report published by London Business School and Sifted on 21 September 2026 makes the organisational side of this issue unusually explicit. Solving Europe's AI Adoption Puzzle argues that Europe's challenge is less a shortage of AI technology than a problem of organisation, including how companies redesign work, build skills and create enough trust for adoption.

That shifts the question from whether employees have access to AI toward what the organisation does around the tool. Two companies can deploy similar systems and create very different capability outcomes. One can use AI to remove administrative friction while deliberately protecting coaching, judgment and live practice. Another can convert every saved minute into additional throughput and assume skills will take care of themselves.

A BMC Psychology study published the same day provides an important counterweight to a simple deskilling story. Using a three-wave, time-lagged survey of 380 full-time knowledge workers, the researchers found that generative AI use was positively associated with proactive behaviour. Greater flow and lower perceived workplace objectification partly mediated that relationship. The design does not establish that AI caused the behavioural change, but it shows why the right question is not whether AI use is good or bad for people in the abstract.

AI can support initiative and make some work feel better while still changing the repetitions through which other skills are maintained.

That is the capability-design problem. Organisations need to decide what they are happy to let the tool do, what they still expect a person to do well, and where the practice for that human work will come from. For negotiation specifically, spoken AI practice matters because preparation and live execution are different tasks.

Designing against deskilling is possible without banning AI

A preregistered experiment published on 17 September 2026 adds a more constructive piece of evidence to this debate. Sebastian Maier, Kai Schwabe, Manuel Schneider and Stefan Feuerriegel tested two interventions with 704 participants using an LLM assistant while practising fraction arithmetic, followed by an unaided test.

One intervention simply rewarded participants for relying less on the model. The other gave metacognitive feedback that made the consequences of offloading explicit. That feedback reduced full-answer offloading and improved subsequent unaided test performance. The reward intervention did not show the same effect.

The task was narrow, the study was short and it did not measure workplace negotiation or long-term skill retention. It would be a mistake to treat it as proof that a particular learning design will preserve complex professional skills.

What it does show is more useful: access to AI and independent human performance do not have to be treated as opposites. The design of the interaction matters.

That distinction is highly relevant to capability building. If a learning system immediately supplies the answer every time the learner hesitates, it may produce a polished result while removing the retrieval, judgment and self-correction that the learner actually needs to practise. A system can instead expose the learner to support while still making it visible when the model did the thinking.

For spoken negotiation practice, the equivalent principle is straightforward. AI can help prepare the case, generate a counterparty and analyse the conversation afterwards. But during the live exchange, the learner still has to decide what to say, retrieve the argument, handle resistance and live with the consequence of that choice. The point is not to make AI less useful. It is to design the learning environment so that usefulness does not replace the behaviour being trained.

Companies may need to separate productivity systems from capability systems

Most organisations are currently asking how to deploy AI to increase output.

That is the correct question for many workflows.

But learning and capability functions need a second question: which parts of the work do we want AI to remove, and which parts do people still need to perform themselves because performing them is how the skill is maintained?

The answer will differ by role.

A lawyer may not need to draft every standard clause manually to remain a strong lawyer. A buyer does not need to create every supplier summary from scratch to remain a strong negotiator. A manager does not need to format every status update personally to remain a strong leader.

But the buyer still needs to handle resistance. The manager still needs to deliver difficult feedback. The leader still needs to make a case when the room disagrees.

Those are not inefficiencies to automate away if the organisation still expects humans to own the outcome.

They are capabilities to preserve.

More automation can still mean more meetings

A particularly revealing counterpoint comes from Anthropic's 16 October 2026 study of its own workforce. The researchers combined employee and hiring records with AI usage, output metrics, calendar patterns, internal communications and 17 interviews. They found that the median employee spent 20 percent more time in meetings than a year earlier. Even among technical individual contributors, meeting time rose 16 percent. The researchers explicitly caution that company growth and a changing mix of roles explain part of the overall increase. The study does not establish that AI caused the additional meetings or improved anyone's conversation skills.

The more interesting finding is what employees describe doing between those meetings. As agents take over implementation work, people increasingly supervise outputs, resolve ambiguities and make decisions that span teams. Human attention becomes a bottleneck, while autonomy and judgment feature more prominently in hiring assessments. At this unusually AI-intensive company, automation has not eliminated the need to coordinate with other people. It has changed the work around that coordination.

This does not contradict the earlier Microsoft evidence that productivity-oriented digital activity grew faster than communication-oriented activity among heavy AI users. The two studies observe different organisations, populations and measures. One tracks relative changes in categories of digital activity; the other describes calendar time and work organisation inside a rapidly growing AI company. Neither measures whether employees have become better at handling disagreement.

For procurement, the implication is concrete. Automating a quotation comparison may save hours, but a supplier's disputed assumptions, a stakeholder's objections or an escalation over delivery still need a decision. Those interactions may be more consequential precisely because routine preparation has become easier. More meeting time is not a capability metric, however. Teams should examine whether buyers can challenge unsupported claims, explain trade-offs, listen under pressure and reach decisions they can defend, then give them opportunities to practise those behaviours aloud.

When asking AI replaces asking a colleague

Workplace knowledge rarely sits neatly in one document. Teams also rely on a shared understanding of who knows what, who has dealt with a problem before and who can be trusted when the answer is uncertain. That network is built in small interactions which may seem inefficient until they disappear.

In an October 15, 2026 Harvard Business Review article, researchers from BetterUp Labs and Stanford describe survey evidence from roughly 1,000 knowledge workers. Some respondents reported preferring to consult AI rather than colleagues, while others said teammates approached them less frequently. Colleague-directed questions were associated with stronger reported coordination and trust. The findings are observational and self-reported; they do not establish that choosing AI caused relationships to weaken. The research also involves a commercial coaching provider, a relevant consideration when interpreting its conclusions.

For a buyer, asking a colleague is often doing two jobs at once. The immediate question might concern a supplier's quality history or why a technical specification cannot move. But asking also reveals who holds that knowledge, builds a channel for future escalation and gives the colleague a chance to disagree with the assumptions behind the question. An AI-generated answer can be accurate and useful while providing none of those relational benefits.

The distinction is not between using AI and collaborating. Procurement teams can use AI to prepare better questions, discover gaps in their brief and reduce administrative work. The risk arises when a plausible answer becomes the reason not to speak to the person whose judgment or cooperation will ultimately be needed. A useful development measure is not how often employees interact, but whether they can still find expertise, invite challenge and resolve a disagreement when it matters.

Practice may become more important as AI gets better

This creates a counterintuitive conclusion.

The more AI can do before and after a conversation, the more valuable deliberate conversational practice may become.

Not because people should reject AI, but because the remaining human work becomes more concentrated in moments where live judgment matters.

A good practice environment can create repetitions that the working day no longer supplies reliably. A buyer can rehearse a price increase discussion without risking the supplier relationship. A manager can practise an escalation conversation before the real employee is sitting across the table. A team can repeat the same difficult moment several times and receive feedback on what actually changed.

This is the logic behind Voice2Evolve. AI does not conduct the real negotiation for the buyer. It provides conversational sparring for procurement negotiation training, with an AI sparring partner and analysis of the conversation that actually happened.

That distinction matters more in an AI-heavy workplace, not less.

The Microsoft study is early evidence about a shift in work patterns, not proof of a decline in human communication skill. But it points toward a question organisations should start answering before the evidence becomes uncomfortable.

If AI increasingly handles the documents, summaries, drafts and routine coordination around work, where will people build the capabilities needed for the conversations that cannot be delegated?

Productivity can be automated surprisingly quickly.

Conversation skill still has to be practised.

Sources

• Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (16 October 2026). How employment is changing at Anthropic. Anthropic Institute.

• Maier, S., Schwabe, K., Schneider, M., & Feuerriegel, S. (17 September 2026). Designing Against Deskilling: Metacognitive Feedback Reduces Cognitive Offloading to LLM Assistants. arXiv preprint.

• London Business School & Sifted (21 September 2026). Solving Europe's AI adoption puzzle.

• Xu, S., Mei, Y., Yao, K., et al. (21 September 2026). More flow and less objectification: Dual influence mechanisms of generative AI use on employees' proactive behaviors. BMC Psychology.

• Gartner (9 September 2026). Prevent AI-Related Skill Loss to Maintain Procurement Cost Goals.

• Culture Amp (17 September 2026). AI at Work Benchmark 2026: 85% Are Encouraged to Use AI. 42% Don't Know Why.

• American Management Association (2026). The New Management Reality: AI, Engagement, and the Invisible Work of Managers.

• Talogy (15 September 2026). Starting a career in the age of AI: The human skills that matter.

• Yu, Y., Chen, Y., Hu, R., Suri, S., & Counts, S. (2026). Adoption of Generative AI in the Workplace: Increasing and Shifting the Balance of Productivity and Communication Activity. Microsoft Research / arXiv.