Digital, Virtual and AI-Enabled Negotiation

Introduction

International negotiation has moved from the conference room to a digitally distributed environment. Negotiators now work through videoconferencing platforms, collaborative documents, instant messaging, electronic signatures, digital data rooms, automated translation, predictive analytics and generative artificial intelligence (AI). These technologies reduce travel, accelerate information exchange and make cross-border participation possible. They also alter how trust is formed, how information is interpreted, how concessions are communicated and how power is distributed.

Digital negotiation is not simply traditional negotiation conducted through a screen. The medium changes the negotiation process itself. A video call may reduce travel and increase scheduling flexibility, yet it can also suppress peripheral cues, create unequal participation and magnify technical interruptions. An asynchronous exchange can provide time for reflection and translation, but it may encourage strategic delay and make relational repair more difficult. AI can identify patterns and draft proposals, but it may reproduce historical bias, expose confidential information or generate confident but inaccurate recommendations.

The central managerial task is therefore not to adopt technology uncritically. It is to design a negotiation system in which digital tools improve coordination and analysis while human negotiators retain responsibility for judgment, relationships, legality and ethical accountability.

Learning Objectives

By the end of this chapter, students should be able to:

1. Distinguish digital, virtual, hybrid and AI-enabled negotiation.
2. Select communication channels according to task complexity, relationship needs and cultural context.
3. Design effective virtual and hybrid negotiation processes.
4. Evaluate algorithmic decision support, analytics, generative AI and automated translation.
5. Identify privacy, cybersecurity, bias, transparency and accountability risks.
6. Apply practical protocols for managing the loss of contextual cues.
7. Critically assess a digitally enabled international negotiation case.

1. The Digital Negotiation Environment

1.1 Core terminology

Digital negotiation uses electronic systems to exchange information, communicate, analyse alternatives or record agreements. Virtual negotiation occurs when participants are geographically separated and interact through digital channels. Hybrid negotiation combines co-located and remote participants. AI-enabled negotiation uses machine-learning or generative systems to support tasks such as preparation, forecasting, translation, drafting or option generation.

These categories overlap. A negotiation may be virtual but not AI-enabled, or hybrid and supported by analytics. The distinction matters because each arrangement creates different risks and requires different controls.

1.2 Channel richness and task fit

A useful principle is channel-task fit. Rich channels communicate multiple cues, permit rapid feedback and support personalisation. Face-to-face interaction is generally rich; video is less rich because it narrows the field of perception; audio, text and structured forms progressively reduce available cues. However, richness is not automatically superior. A written channel may be better for a precise price schedule, a sensitive concession or a record requiring later verification.

Channel selection should consider:

• Ambiguity: How many reasonable interpretations could the message have?
• Interdependence: How much must parties coordinate their actions?
• Relational intensity: Is trust, apology or identity recognition central?
• Cultural and linguistic distance: Will indirectness, silence or hierarchy be easily misunderstood?
• Confidentiality: What information must be protected?
• Time-zone distribution: Can all parties participate fairly?
• Documentation requirements: Is an auditable record needed?

High-ambiguity and high-relational tasks normally require richer interaction, while routine data confirmation can be handled asynchronously. Effective negotiators use a portfolio of channels rather than treating one platform as universally appropriate.

1.3 Reduced cues and social presence

Digital interaction may remove or distort eye contact, posture, pauses, side conversations, room layout and the informal moments in which trust develops. Camera framing can make a participant appear disengaged when the person is reading a document. A delay can be interpreted as hesitation, disagreement or disrespect. A muted microphone can conceal an attempted interruption. Automated captions may omit culturally significant expressions.

The solution is not to guess the meaning of missing cues. It is to make important cues explicit. Negotiators should state the purpose of silence, summarise interpretations, invite dissent, check understanding and distinguish technical problems from substantive disagreement.

2. Designing a Virtual or Hybrid Negotiation

2.1 Preparation as process architecture

In a physical meeting, the room and schedule often provide a natural structure. In virtual negotiation, structure must be deliberately designed. Preparation should cover five layers:

1. Strategic layer: objectives, reservation point, aspiration level, BATNA and possible integrative trades.
2. Relational layer: introductions, trust-building, cultural protocols and rules for disagreement.
3. Technical layer: platform, connectivity, interpretation, document access and authentication.
4. Procedural layer: agenda, speaking order, caucus arrangements, breaks and decision authority.
5. Governance layer: data classification, recording consent, retention, cybersecurity and AI-use disclosure.

The reservation point is the least favourable agreement a party will accept, while the BATNA is the best alternative available if no agreement is reached. These concepts remain essential online, but digital environments can make alternatives easier to investigate and therefore change bargaining power. For example, online procurement platforms may increase a buyer’s knowledge of substitute suppliers, while platform dependence may make a small seller more vulnerable.

2.2 Pre-negotiation protocol

A written protocol should specify:

• names, roles and authority of participants;
• time-zone conversions and maximum session length;
• whether recording, transcription or automated notes are permitted;
• approved platforms and identity-verification procedures;
• document versions and access rights;
• rules for private caucuses and side communications;
• procedures for technical interruption;
• language, interpretation and translation arrangements;
• how tentative statements differ from binding commitments;
• the method for confirming the final agreement.

The protocol should be accepted by all parties before substantive bargaining. This prevents procedural questions from becoming improvised sources of conflict.

2.3 Managing time zones and participation

Time-zone scheduling is an equity issue, not merely an administrative issue. Repeatedly assigning inconvenient hours to one region can communicate low status and reduce cognitive performance. A rotating schedule, shorter sessions and asynchronous preparation distribute the burden more fairly. Breaks should be planned rather than treated as evidence of weak commitment.

In hybrid meetings, remote participants should have equal access to documents, chat, voting, private consultation and the speaking queue. A single room microphone is rarely sufficient. Each participant should be visible and audible, and the chair should actively monitor remote contributions.

2.4 Opening and closing rituals

The opening should establish both relationship and process. Participants can briefly introduce their roles, acknowledge cultural or national contexts, confirm pronunciation of names and review the agenda. A short informal exchange may be valuable, but it should not impose one culture’s preferred style of rapport on everyone.

At the end of each session, the chair should summarise: issues agreed, issues unresolved, assumptions, responsible persons, deadlines and the next decision point. Each party should confirm whether the summary reflects its understanding. The final contract must be validated through the legally appropriate procedure; a chat message or AI-generated summary is not automatically a binding agreement.

3. Negotiation Analytics and Algorithmic Decision Support

3.1 What analytics can do

Analytics may support preparation by organising historical prices, identifying patterns in delivery performance, estimating demand, comparing alternatives and testing scenarios. A multi-issue negotiation can be represented as a set of options with different values to each party. This supports trade-offs rather than single-issue positional bargaining.

A simple utility model is:

Here, U(x) is the negotiator’s overall utility for package x, xᵢ is the value of issue i, vᵢ(xᵢ) is the normalised value of that issue, and wᵢ is its importance weight. The weights are commonly non-negative and sum to 1. This model assumes that the issues can be represented on meaningful scales and that an additive approximation is acceptable. It is a decision aid, not an objective measure of human value.

For example, suppose a buyer evaluates price, delivery reliability and technical support with weights 0.50, 0.30 and 0.20. A package has normalised scores of 0.80, 0.60 and 0.90. Its utility is:

The result, 0.76, has no physical unit. It is meaningful only relative to other packages evaluated using the same definitions and weights. If the issues interact strongly—for example, faster delivery sharply increases cost—the additive model may misrepresent the decision. Interaction terms or a different model may then be required.

Analytics can also estimate expected value under uncertainty. If outcome j has probability pⱼ and utility Uⱼ, expected utility is:

The probabilities must represent the decision maker’s justified beliefs, and the outcomes should be mutually exclusive and collectively exhaustive for the scenario being evaluated. Expected utility does not predict what will happen; it compares uncertain options under stated assumptions.

3.2 Limits of algorithmic recommendations

An algorithm may optimise the wrong objective, rely on incomplete data or treat historical behaviour as a norm. Historical purchasing data may reflect discriminatory supplier selection, unequal bargaining power or temporary market conditions. A model trained on such data can reproduce these patterns while appearing neutral.

Important evaluation questions include:

• What decision is the system intended to support?
• Which data were used, and who is missing from them?
• Are variables proxies for nationality, ethnicity, gender, location or socioeconomic status?
• What error rates occur across relevant groups?
• Can a human challenge the recommendation?
• Is the model stable when market conditions change?
• Can the organisation explain the recommendation to an affected party?

Human oversight must be substantive. A manager who merely approves an algorithmic output without understanding its basis provides ceremonial rather than meaningful oversight.

4. Generative AI in Negotiation

4.1 Appropriate uses

Generative AI can help a negotiator:

• create a preparation checklist;
• summarise authorised documents;
• generate alternative wording for a proposal;
• identify unanswered questions;
• simulate possible objections;
• compare contract clauses;
• translate a draft for human review;
• classify issues by urgency or complexity.

These uses are safest when the input is controlled, the output is advisory and a qualified person verifies the result. AI is particularly useful for breadth: it can suggest options that a busy team might overlook. It is weaker at understanding tacit relationships, political symbolism, organisational history and the difference between a literal statement and a culturally appropriate implication.

4.2 Hallucination, overconfidence and confidentiality

Generative systems may produce fluent but false claims, invented sources, incorrect legal interpretations or fabricated concessions. Fluency is not evidence of accuracy. Every factual, numerical, legal and contractual assertion requires verification against an authoritative source.

Confidential negotiation data should not be entered into a public or unapproved system. Sensitive information includes reservation points, personal data, trade secrets, litigation strategy, identity documents, unpublished prices and information received under a confidentiality agreement. Data minimisation means providing only what is necessary for the task, preferably with names and identifying details removed. Organisations should define retention, access, deletion and vendor-use rules before deployment.

4.3 Human accountability

The negotiator remains responsible for decisions and communications even when AI drafted them. A useful control is a two-person review for high-impact outputs, such as a settlement proposal, employment-related term, regulatory representation or cross-border data clause. The reviewer should check factual accuracy, tone, cultural appropriateness, confidentiality and authority to commit.

AI use may also require disclosure. The appropriate degree depends on the contract, law, organisational policy and expectations of the parties. Concealing automated translation or analysis can damage trust if discovered, especially when the system affects substantive terms.

5. Automated Translation and Intercultural Meaning

Automated translation improves access but does not eliminate linguistic risk. Words carry pragmatic meaning: politeness, status, reluctance, urgency and commitment may be expressed indirectly. A translation can be grammatically correct while changing the force of a statement. Terms such as “perhaps,” “we will consider,” “difficult” or “we expect” may signal different degrees of commitment in different cultures and contexts.

Best practice includes:

1. Agree on the working languages and authoritative version.
2. Use trained interpreters for high-stakes or relationship-sensitive exchanges.
3. Provide glossaries for technical, legal and culturally specific terms.
4. Avoid idioms, sarcasm and unnecessarily complex sentences.
5. Pause after important statements and invite restatement.
6. Record unresolved linguistic ambiguities rather than silently selecting one meaning.
7. Have bilingual legal or technical reviewers validate the final text.

Back-translation—translating a statement into the target language and then independently translating it back—can reveal major discrepancies, but it is not a guarantee of equivalence. The ultimate test is whether the parties share the same operational understanding.

6. Privacy, Cybersecurity and Regulatory Governance

Digital negotiation creates a valuable concentration of information. A platform account, meeting link or shared folder may expose commercial strategy, personal data or legally privileged communications. Threats include credential theft, meeting intrusion, malware, unauthorised recording, insider access, deepfake impersonation and compromised vendors.

A proportionate security framework should include:

• multi-factor authentication;
• unique accounts and least-privilege access;
• encryption in transit and at rest where appropriate;
• waiting rooms and authenticated invitations;
• controlled screen sharing and file permissions;
• secure, tested backups;
• logging of access and material changes;
• patching and endpoint protection;
• a breach-response and notification plan;
• retention and deletion schedules.

Security should be risk-based. A public exploratory meeting and a negotiation involving trade secrets do not require identical controls, but both require informed decisions about risk.

Privacy obligations vary by jurisdiction and sector. Cross-border processing may involve data-protection, employment, financial-services, export-control or records-management rules. Organisations should identify the legal basis for processing, limit collection, respect data-subject rights where applicable and determine where data are stored and transferred. Legal counsel should assess the relevant jurisdictions; technology alone cannot resolve conflicting legal obligations.

7. Bias, Fairness and Transparency

Bias can enter through data, problem definition, labels, model design, deployment or human interpretation. In international negotiation, language, accent, nationality, location and communication style may become unfair proxies for competence or trustworthiness. A sentiment-analysis tool may interpret indirect disagreement as positivity, or classify a direct style as hostility. Such outputs can influence who receives concessions or decision authority.

Fairness is not always achieved by ignoring sensitive attributes. If an organisation does not measure outcomes by relevant groups, it may fail to detect disparate effects. However, collection and analysis of such attributes must itself respect privacy and legal requirements. The correct approach is governed measurement, contextual interpretation and remediation—not automatic reliance on a single fairness score.

Transparency should be calibrated to audience and risk. Parties should know when an automated system materially influences evaluation, translation or recommendations. They should be told the system’s purpose, major limitations and route for human review. Trade-secret protection may limit disclosure of source code, but it does not justify concealing consequential use or refusing all explanation.

8. A Practical Framework: Prepare, Connect, Interpret, Decide, Record

A useful operating framework is the PCIDR cycle:

• Prepare: define objectives, BATNA, authority, data permissions, channel and cultural protocol.
• Connect: authenticate participants, establish rapport, test technology and confirm process rules.
• Interpret: distinguish words from inferred meaning; check translation, silence, emotion and technical artefacts.
• Decide: compare options, challenge analytics, obtain human approval and identify unresolved assumptions.
• Record: document agreements, owners, deadlines, version history, disclosures and follow-up mechanisms.

The cycle is iterative. A cybersecurity incident may require returning from Decide to Connect. A translation ambiguity may require returning from Interpret to Prepare. The framework prevents technological efficiency from displacing relational and governance work.

9. Illustrative Case: The Meridian Components Negotiation

Background

Meridian Components, a German industrial manufacturer, was negotiating a three-year supply agreement with Koyo Precision, a Japanese supplier, and a Mexican assembly subsidiary. The parties used a hybrid format: Meridian’s procurement team met in Frankfurt, Koyo joined from Tokyo, and the subsidiary joined from Monterrey. English was the working language, although several participants were more comfortable in German, Japanese or Spanish.

Meridian used an analytics dashboard that ranked supplier packages on price, delivery reliability and defect rates. A generative-AI tool summarised meetings and proposed contract language. Automated captions were enabled by default.

Problem

The dashboard ranked Koyo’s package below a lower-priced competitor. Meridian’s procurement director therefore pressed for a rapid price concession. During a video meeting, Koyo’s representative said that the proposed delivery schedule was “difficult.” The automated summary recorded this as “supplier accepts schedule subject to minor adjustment.” Meridian treated the statement as limited resistance.

At the same time, the Monterrey team experienced audio delay and could not hear a private discussion among the Frankfurt participants. A confidential spreadsheet containing Meridian’s reservation point was accidentally uploaded to a shared folder accessible to all participants. The AI meeting tool retained transcripts under a vendor policy that had not been reviewed by legal counsel.

Analysis

Several failures occurred. First, the channel was too thin for a high-ambiguity, relationship-sensitive discussion. Second, the parties had not agreed on the meaning of tentative language or the authoritative record. Third, the analytics model weighted price heavily and did not adequately represent the cost of delivery disruption to the assembly subsidiary. Fourth, the hybrid format created an information hierarchy favouring the physically co-located group. Finally, confidentiality and retention controls were absent.

The problem was not that the technology was inherently defective. The organisation had failed to design governance around it. The dashboard’s score was treated as a decision rather than as a conditional model. The AI summary was treated as a transcript of meaning rather than a draft requiring bilingual and contextual review.

Approach and solution

The parties paused substantive bargaining. They revoked access to the confidential file, documented the incident and obtained legal advice on notification and retention. The next session began with a revised protocol:

• all participants used individual authenticated connections;
• the chair repeated every material statement in neutral language;
• interpreters and bilingual reviewers checked key terms;
• no AI system processed confidential material without written approval;
• the dashboard displayed separate scores for each issue rather than one composite ranking;
• the Monterrey team received equal access to the private caucus process;
• any tentative agreement required written confirmation within 24 hours.

Meridian recalculated the package using revised weights and scenario analysis. It recognised that a slightly higher unit price could be offset by lower disruption risk and improved technical support. Koyo clarified that “difficult” meant the schedule was unacceptable without a phased implementation, not that only a minor adjustment was required. The parties developed a phased delivery plan, a quality-improvement programme and a price formula linked to specified cost indices.

Outcome

The final agreement cost more per unit than Meridian’s initial target but produced a more robust total-value package. The parties avoided an unrealistic launch schedule, improved defect reporting and established a joint digital data room with access controls. Meridian retained analytics for scenario comparison but prohibited automated systems from making unreviewed recommendations on reservation points or contractual commitments.

Learning points

1. Composite scores can conceal strategic trade-offs.
2. Translation and summarisation require contextual validation.
3. Hybrid meetings can create unequal participation even when everyone has a login.
4. Confidentiality must be designed before tools are activated.
5. Trust depends on explainable process, not merely efficient software.
6. A pause after a technology failure can protect value better than hurried continuation.

10. Contemporary Developments and Strategic Implications

Emerging systems increasingly combine generative AI, speech recognition, sentiment analysis, digital twins and automated workflow tools. Negotiators may receive real-time prompts about unanswered issues, possible package trades or changes in market conditions. Synthetic voice and video also create impersonation risks, making identity verification more important.

Organisations should avoid an arms race in which each party deploys increasingly opaque systems. If one side uses an AI agent to generate thousands of offers while the other negotiates manually, speed may become coercive rather than productive. Parties may need rules limiting automated messaging, requiring disclosure of bot participation or preserving a human contact for material decisions.

The strategic advantage of AI is likely to lie less in replacing negotiators than in improving preparation, memory and option generation. Competitive advantage will depend on data quality, cultural intelligence, governance maturity and the ability to combine machine analysis with human judgment. Firms should build AI literacy among negotiators, legal teams, information-security staff and senior decision makers rather than delegating responsibility to an isolated technology function.

Best-Practice Checklist

Before a digital negotiation, confirm:

• objectives, BATNA, authority and decision criteria;
• participant identity, roles and time-zone fairness;
• platform security, recording consent and data retention;
• language, interpretation and authoritative text;
• hybrid-equity arrangements;
• contingency plans for outages;
• approved and prohibited AI uses;
• human review of analytical and generative outputs;
• procedures for ambiguity, caucuses and confirmation;
• post-meeting records, access controls and follow-up.

Chapter Summary

Digital, virtual and AI-enabled negotiation changes both the opportunities and the risks of international bargaining. Virtual channels improve access, speed and documentation, but they reduce contextual cues and can produce unequal participation. Effective negotiators therefore match channel richness to task ambiguity, relationship intensity, cultural distance and confidentiality requirements.

A well-designed virtual process requires strategic, relational, technical, procedural and governance preparation. Hybrid meetings need special controls to prevent remote participants from becoming peripheral. Analytics and utility models can clarify trade-offs, but their outputs depend on assumptions, data quality and weights. Generative AI can support preparation, drafting, translation and option generation, but fluent output may be inaccurate, biased or confidentially unsafe.

Privacy, cybersecurity, transparency, fairness and human accountability are central managerial responsibilities. Automated translation and summarisation must be checked for pragmatic and cultural meaning. The PCIDR cycle—Prepare, Connect, Interpret, Decide and Record—provides a practical framework for integrating technology with judgment. The goal is not maximum automation; it is better-informed, more inclusive and more trustworthy negotiation.

Key Terms

Digital negotiation: Negotiation conducted with electronic communication or information systems.

Virtual negotiation: Negotiation among geographically separated participants using digital channels.

Hybrid negotiation: Negotiation combining co-located and remote participants.

AI-enabled negotiation: Negotiation supported by artificial-intelligence systems.

Channel richness: The capacity of a communication medium to convey cues, feedback and personalisation.

BATNA: Best alternative to a negotiated agreement.

Reservation point: The least favourable outcome a party will accept.

Algorithmic decision support: A system that analyses data and recommends, but does not necessarily determine, a decision.

Generative AI: AI that produces text, images, code or other content from prompts or data.

Hallucination: A plausible-looking but false or unsupported AI output.

Automated translation: Machine conversion of content from one language to another.

Back-translation: Translation into a target language followed by independent translation back into the original language.

Data minimisation: Collecting and processing only the data necessary for a defined purpose.

Least privilege: Giving users only the access required for their authorised tasks.

Hybrid inequity: Unequal access or influence between co-located and remote participants.

Algorithmic bias: Systematic distortion or unequal treatment arising from data, design or use of an algorithm.

Explainability: The ability to communicate how or why a system produced an output.

Human-in-the-loop: A design in which a person reviews, approves or can override an automated output.

Caucus: A private discussion conducted separately from the joint negotiation.

PCIDR cycle: Prepare, Connect, Interpret, Decide and Record framework for digitally enabled negotiation.

Review Questions

Short Answer Questions

1. What distinguishes virtual negotiation from hybrid negotiation?
2. Why can a written channel sometimes be preferable to video?
3. Define BATNA and reservation point.
4. Identify two risks of generative AI in negotiation.
5. What is meant by data minimisation?

Long Answer Questions

1. Explain how loss of contextual cues affects trust and interpretation in virtual international negotiation.
2. Discuss the advantages and limitations of analytics and utility models in multi-issue negotiation.
3. Develop a governance protocol for using generative AI to prepare a confidential cross-border negotiation.
4. Analyse why hybrid negotiations may produce unequal participation and recommend controls.
5. Explain how privacy, cybersecurity, transparency and human accountability should be integrated into digital negotiation design.

Analytical / Application Questions

1. A supplier says a proposed delivery date is “challenging” during an interpreted video meeting. Design a process for determining whether this is a polite refusal, a request for adjustment or a minor concern.
2. An algorithm ranks a supplier lowest because of historical delivery data. What diagnostic questions should managers ask before relying on the ranking?
3. Design a PCIDR-based plan for a three-country negotiation involving confidential pricing data, two working languages and both in-person and remote participants.

Multiple Choice Questions with Answers

Question 1

Which negotiation mode combines co-located and remote participants?

A. Asynchronous negotiation

B. Hybrid negotiation

C. Algorithmic negotiation

D. Bilateral negotiation

Correct Answer: B. Hybrid negotiation

Question 2

Which factor most directly indicates that a richer communication channel is needed?

A. The message contains a routine invoice number

B. The parties must clarify ambiguous, relationship-sensitive issues

C. The parties need a permanent written record only

D. The participants have identical technical expertise

Correct Answer: B. The parties must clarify ambiguous, relationship-sensitive issues

Question 3

In the utility example, what is the utility of scores 0.80, 0.60 and 0.90 with weights 0.50, 0.30 and 0.20?

A. 0. sixty

B. 0. seventy

C. 0. seventy-six

D. 0. ninety

Correct Answer: C. 0. seventy-six

Question 4

What is a major limitation of treating an algorithmic recommendation as objective?

A. Algorithms always require face-to-face meetings

B. Historical data may contain bias or reflect inappropriate objectives

C. Algorithms cannot process numerical data

D. Recommendations cannot be documented

Correct Answer: B. Historical data may contain bias or reflect inappropriate objectives

Question 5

Which practice best reduces the risk of generative-AI hallucination in a contract draft?

A. Accepting fluent language as evidence of accuracy

B. Removing all human review

C. Verifying factual and legal claims against authoritative sources

D. Entering more confidential data into a public tool

Correct Answer: C. Verifying factual and legal claims against authoritative sources

Question 6

What is the primary purpose of back-translation?

A. To eliminate the need for interpreters in every situation

B. To identify possible discrepancies between language versions

C. To make a negotiation legally binding automatically

D. To conceal the use of translation technology

Correct Answer: B. To identify possible discrepancies between language versions

Question 7

Which control follows the principle of least privilege?

A. Giving every participant administrator access

B. Allowing users only the access needed for authorised tasks

C. Publishing the reservation point to all attendees

D. Disabling authentication for convenience

Correct Answer: B. Allowing users only the access needed for authorised tasks

Question 8

In the PCIDR cycle, which stage focuses on checking meaning, translation and technical artefacts?

A. Prepare

B. Connect

C. Interpret

D. Record

Correct Answer: C. Interpret

Question 9

Why can a composite supplier score be misleading?

A. It prevents any comparison of alternatives

B. It may conceal trade-offs among price, quality and delivery risk

C. It is always expressed in physical units

D. It cannot use historical data

Correct Answer: B. It may conceal trade-offs among price, quality and delivery risk

Question 10

Which statement best describes human accountability in AI-enabled negotiation?

A. The software vendor automatically assumes responsibility for the agreement

B. Human negotiators remain responsible for decisions and communications using AI outputs

C. No disclosure is ever required when AI is used

D. Human review is unnecessary if the output is statistically likely

Correct Answer: B. Human negotiators remain responsible for decisions and communications using AI outputs