Negotiation: when the other side
has done its homework

The other side of your next negotiation may arrive with a surprisingly detailed view of your organisation. Public announcements, vacancies, filings, product pages, tenders, case studies and executive comments can now be gathered and summarised at a cost that makes preparation viable beyond the largest accounts. Artificial intelligence does not need to uncover a secret to change the conversation. It only needs to make a useful question cheaper to prepare.

That creates an obvious temptation: ask a model to explain what the counterparty really wants, then treat its polished profile as hidden truth. This is the wrong use of the capability. Public evidence is partial, motives are not directly observable and generative systems can present false or weak inferences with confidence.

The more defensible advantage is procedural. Use AI to collect attributable facts, expose a small set of provisional hypotheses and turn each one into a question that the conversation can answer. Record what was actually said or offered. Weaken, replace or retain the hypothesis. Then explore packages across price, timing, scope, risk, service and commitment without surrendering your own constraints or authority.

The governing distinction is simple: an account profile tries to describe the other party; a working negotiation model is designed to be corrected by the other party. Its value does not depend on mind-reading. It comes from making preparation explicit enough to test.

The meeting-ready tool in this article is the Preparation-to-Agreement Loop. It keeps evidence, interpretation, conversation, terms and authority separate. The framework is Andrew’s Intel editorial synthesis, not an empirically validated universal process.

The proposition in one sentence

Do not ask whether the system has understood the counterparty. Ask whether it has produced a better, falsifiable question and a safer way to revise the deal after hearing the answer.

This matters because scale changes the economic threshold for homework. A team may already know how to prepare carefully for a strategic account or major procurement. It may not justify the same hours for a routine renewal, a mid-sized supplier discussion or the long tail of commercial conversations. If AI reduces the first-pass cost, useful preparation can reach more of the work. That is a coverage claim, not a claim that the machine has become the organisation’s best negotiator.

A working negotiation model is designed to be corrected by the other party.

The Preparation-to-Agreement Loop

Use the loop before, during and immediately after a bounded commercial conversation. The purpose is not to score the counterparty. It is to keep the team’s reasoning revisable.

On a narrow screen, each stage is shown as a separate card.

A worked negotiation

Consider a hypothetical manufacturer discussing a new production system with a supplier. Public material shows a recently announced facility, vacancies for particular operating skills and a partner case study describing an installation. The supplier’s AI-assisted brief proposes two hypotheses: the manufacturer may value implementation speed, or it may be staging investment because cash flow and internal readiness are constrained.

Neither hypothesis is a fact. A new facility can be delayed, repurposed or already equipped. Recruitment can reflect replacement rather than expansion. A partner’s marketing copy can exaggerate timing. The brief therefore supplies questions rather than conclusions: What has to be operating first? Which dependency makes the timetable credible? Would the organisation value a phased installation if that reduced execution risk? Is cash timing relevant, or is certainty of delivery the stronger concern?

Suppose the conversation reveals that capital is available, but training capacity is the bottleneck. The original cash-flow hypothesis weakens. The parties can now compare a rapid full installation with a phased delivery tied to training milestones. Price still matters, but so do acceptance criteria, service coverage, timing and the risk of unused capacity.

A profile would have rewarded the team for being right about the manufacturer. A working model rewards it for changing its proposal when it was wrong. That is a more useful standard because the commercial objective is not to win a psychological prediction. It is to find an acceptable agreement, or to recognise efficiently that no such agreement is available.

This example is illustrative. It is not evidence of a deployed product or measured outcome. It shows the mechanism that a real trial would need to test.

Preparation is valuable only if the meeting can overturn it

Human negotiation research offers a relevant but bounded clue. Across three studies published in 2008, Adam Galinsky and colleagues reported that perspective taking improved participants’ ability to identify hidden agreements and create and claim value. These were controlled human studies, not evidence that AI-generated profiles improve commercial negotiations. They support the importance of considering another viewpoint; they do not license a machine to declare what that viewpoint is. ²

A peer-reviewed 2026 PNAS study of an international autonomous-negotiation competition reports 182,812 AI-to-AI negotiations. In that setting, positivity, gratitude and question-asking were associated with reaching deals and with objective and subjective value. These are associations, not causal estimates, and the study does not test human commercial negotiations. ³

The practical implication is narrower. A preparation system should be judged by the quality of the questions and updates it enables, not by the confidence of its starting narrative. A hypothesis that cannot be falsified in the conversation is a story, not an instrument.

What AI changes now

The capability signal is the falling cost of extracting structure from public text. A 2026 Federal Reserve Bank of Dallas working paper fine-tuned and assessed a large language model using 30,000 human classifications, then applied it to nearly 600 million job-vacancy postings across five English-speaking countries. The task was narrow: identifying remote-work practices. It was not negotiation research. It nevertheless demonstrates that messy public text can be converted into a large, structured signal at a scale that manual review would struggle to match. ¹

In a commercial preparation workflow, the useful tasks are similarly bounded. The system can gather documents, identify dated facts, compare changes, surface contradictions, propose alternative explanations, draft questions and organise notes against a pre-agreed structure. Each output should retain a link to the source or be labelled as interpretation.

The management signal is not the existence of a better briefing. It is the decision to redesign the work around revision. Leaders choose what evidence is lawful and appropriate to use, which claims require verification, what the team records, which variables it may trade, where approval sits and how a failed hypothesis is treated. If the performance metric counts the number of briefs generated, the system will produce briefs. If it measures useful questions, model updates, proposal quality, exceptions and realised outcomes, it has a chance to improve commercial judgement.

Capability and management must remain separate. A model may be capable of retrieving and summarising a hundred pages while the organisation remains incapable of correcting the summary, protecting sensitive information or changing its proposal. The first fact does not imply the second.

The failure modes are managerial, not merely technical

1. The profile becomes prophecy. A fluent account narrative anchors the team before the meeting. New evidence is interpreted to preserve the original story. Requiring an alternative explanation and an explicit update log reduces, but does not remove, this risk.
2. Public becomes permissible. Information being accessible does not make every inference appropriate. The National Institute of Standards and Technology notes that generative systems may infer sensitive personal information by combining sources, and that inaccurate inferences can still cause harm. Use organisationally relevant evidence, minimise personal profiling and apply the legal, privacy and employment rules that govern the context. ⁵
3. Tone becomes diagnosis. Hesitation, pace, emphasis and word choice are ambiguous. They can justify a clarifying question. They cannot responsibly be turned into a claim that the other person is lying, afraid, desperate or ready to concede.
4. The system optimises the wrong deal. Understanding the other party is not a substitute for knowing your own utility, constraints and alternatives. An attractive concession can still breach margin, service, risk or authority limits. The model may propose packages; the organisation must define the acceptable range.
5. Assistance quietly becomes autonomy. A 2025 workshop paper comparing humans and four large language models in bargaining settings found the models anchored at extremes and showed limited adaptation to leverage and context. The study has its own scenarios and should not be universalised, but it is a useful counterweight to claims that a fluent system is ready to control a live negotiation. This article requires no autonomous commitment. Human-assisted preparation is enough. ⁴

Some settings impose a higher bar

Public procurement is an obvious example. Cabinet Office guidance updated in April 2026 explains that covered procurements under the Procurement Act 2023 must have regard to the importance of value for money, public benefit, information-sharing and integrity, and must treat suppliers the same unless a relevant difference justifies different treatment. AI-assisted preparation does not displace those duties. Information given to one supplier may need to be shared appropriately; inferred differences cannot become a covert reason for unequal treatment. ⁶

Regulated, employment, healthcare, education and high-consequence contexts create other boundaries. Specialist review may be required. The safest design principle is that the sensitivity of the evidence, the consequence of a mistake and the authority attached to the output should determine the strength of verification and oversight.

Run a bounded preparation trial

Choose a class of negotiation that occurs often enough to compare, matters enough to reward better preparation and remains low enough in consequence for a supervised trial. Renewals, standard supplier discussions or mid-sized account reviews may be more revealing than a theatrical test on the organisation’s largest deal.

Before the trial, define the evidence boundary, the variables that may be explored, the terms that require escalation and the record that must be retained. Use the same Preparation-to-Agreement Loop for assisted and current-practice cases where a fair comparison is possible.

Measure preparation time, attributable-source coverage, questions that materially changed the working model, corrections, unsupported claims caught, package options considered, human interventions, escalation, agreement quality against pre-defined objectives and cases in which no deal was correctly preferred. Do not treat a higher agreement rate as automatically better. A system that helps a team reject an unsuitable transaction can be valuable.

Review failed inferences, not only successful ones. The important learning record is not ‘the AI predicted urgency’. It is ‘we inferred urgency from these facts, asked this question, observed this answer, changed this part of the proposal and stayed within these limits’. That sequence can be audited and improved.

What the evidence does not establish

No reviewed source in this package shows that AI-assisted public-data research improves the net outcome of ordinary commercial negotiations. The Dallas Fed paper demonstrates large-scale classification in a different domain. The Galinsky studies concern human perspective taking in controlled negotiations. The autonomous-negotiation competition is a peer-reviewed AI-to-AI study; its reported language–outcome relationships are associative and do not transfer automatically to human business settings. The bargaining study tests selected models and scenarios, not every present or future system.

The National Institute of Standards and Technology material is risk guidance, not a negotiation trial. The Cabinet Office guidance describes legal objectives for covered public procurement, not a general commercial playbook. The defensible conclusion is therefore about plausibility and design: the cost of structured preparation can fall; negotiation research makes perspective taking and information exchange relevant; current models remain fallible; and organisations can test a human-owned update loop without granting the system authority to commit.

The sentence for the next commercial review

Show me which parts of this preparation are facts, which are hypotheses, what question could prove each hypothesis wrong and who still owns the decision after the answer changes the deal.

If the system cannot answer those questions, it has produced a more polished profile. If it can, it may have produced a more useful negotiation process.

A better brief is not the objective. A safer, revisable route to an acceptable agreement is.

Evidence note

The central framework, worked example, management prescriptions and trial design are Andrew’s Intel editorial synthesis. No commercial vendor source is used. Publication and review status are labelled explicitly; laboratory, task and regulatory boundaries remain visible. Capability evidence is not treated as evidence of management readiness or realised commercial performance.