That Is Not What We Agreed

Sometimes the difficult parts of a negotiation happen between the meetings.

Issues are discussed and resolved. The lawyer opposite me sends through a document recording the outcome. I read it, and I discover that the position recorded is substantially different from the one that we discussed. Sometimes the lawyer has reconsidered the outcome since the meeting; sometimes we simply left with different understandings of it. But there is always a risk that a correction offered in good faith is treated like an attempt to reopen a settled issue.

I encounter this often enough to write about. Everyone agrees that an item was closed, but the disagreement is over precisely what was agreed. The document has already supplied an answer.

I am defending the agreement, but on the page I am the person asking to change it.

The person who holds the pen in a negotiation has an often underappreciated amount of power. It provides the opportunity to establish the initial position against which everything subsequent is assessed. Once a position is documented and circulated, it becomes the thing that everyone identifies as the agreed state of the deal.

We can agree that issues were resolved. That does not mean that we must necessarily agree with the characterisation of that resolution by our opposing counsel. But the common understanding that the issues were resolved at all gives authority to the written account, whether or not it is deserved. Correcting the record is, almost always, a disturbance to the sense of closure that everyone would prefer to preserve.

My first assumption is rarely that the other party intended to record matters incorrectly. It is reasonably understandable for a lawyer to reassess an outcome and reach another conclusion, or to faithfully record their own understanding, which may be disputed. The advantage of writing it down functions in the absence of deliberate exploitation.

And written records are important, because durably recording agreed terms with the ability to share them is necessary in most negotiated transactions. Someone can review a document if they weren’t in the meeting. They can’t recall the discussion.

The work that goes into producing an account of what was said can now be delegated mostly to AI. Transcripts or summaries arrive in my inbox after most of the important meetings that I attend these days. Critically, though, a summary is understood to be someone’s interpretation of the discussion. A transcript is treated as the discussion itself.

I have seen transcripts omit critical details that materially changed how an exchange should be understood. Consider the following hypothetical situation:

In a meeting about a problematic report, one participant says to its author, “I think you put those figures into the report knowing they were wrong.”

The author says, “Yes.”

The discussion moves on. An AI transcript captures the first person’s statement but omits the author’s response.

The transcript still includes the allegation that the figures were included knowingly, but the acknowledgement by the person who included them was missed. A person reading the transcript will see what was said about the author’s conduct, but they will not see the affirmation by the author.

A statement about what someone knew can be disputed as merely an accusation. The acknowledgement, however short it may be, provides additional context that is critical to the interpretation of the event. Removing the response changes the record as it relates to the participation of each party. This example is hypothetical, but my concern comes from a real transcript in which a consequential acknowledgement was omitted. The omission materially changed what the record showed about an exchange where that acknowledgement mattered.

One word can have that effect.

The word could be totally surrounded by hundreds or thousands of others that were correctly transcribed. Ironically, the correctness of all of those other words lends legitimacy to the completeness of the record. But the importance of that one word shares no relationship with the amount of space it consumes inside the transcript. An acknowledgement, denial, or qualification can be the most critical item inside a record, while accounting for practically none of its length.

The transcript in such a situation will often appear correct. It’s coherent: one person makes a statement and the discussion moves on. Unanswered statements appear all the time in negotiations and other meetings. Nothing in the transcripts I receive flags uncertainty, an inaudible moment, or how confident the system was in what it recorded.

Audiovisual recordings are helpful, when they are available. They enable a party that suspects a transcript is incorrect to challenge it, but the transcript itself still serves as the initial documented account. Compounding the problem is the fact that often the only record available to the participants is the transcript, due to corporate policies or personal aversions to more invasive recording.

People can still challenge transcripts. A response can exist in memory. But in such a circumstance, recollection ends up in competition with a document created for the purpose of serving as an unbiased record of what happened. A memory cannot be replayed, and there is nothing to point at if someone asks for evidence that the transcript missed something.

The absence of something from a transcript does not establish that it did not occur. Yet that absence can be used to question the account of the person who remembers it. The omission becomes a reason to doubt the person who noticed it.

Whether the record is a summary that selects and organises what occurred, or a transcript that attempts to fully reproduce everything that was said, the resulting document can only provide one account. That account favours whichever party it happens to agree with, and burdens the party who has to challenge it.

The AI tool does not need to have an interest in the outcome of a negotiation or meeting for it to produce this problem. Omissions, summary errors, or incorrect transcriptions can offer a benefit to one party over another without deliberate intervention. Removing a human author with an interest does not remove the consequences attached to building the record.

The problems in these documents travel, too. In my hypothetical scenario above, a summary may go further and suggest that the author was given no opportunity to respond to the accusation. Now, not only has the affirmation been excluded from the record, but the author has been positioned as if they were denied something that should have been given. A future reader of that summary has their perspective shaped by a series of automated steps that distorted what occurred.

By that time, the actual disagreement concerns a summary written by someone who wasn’t in the initial conversation and is relying on a document whose accuracy they may have no reason to suspect. Meanwhile, the person who heard the acknowledgement is stuck trying to re-establish what happened in the meeting.

Human memory has never been a reliable solution to this problem. My own experience with summaries produced by lawyers can easily make that case, and I accept that my recollection may be wrong, too. AI could produce better records than humans taking notes. People can also challenge accurate records because they have misremembered what occurred. Complaints should be examined without being granted the full weight of acceptance.

But a history of general reliability cannot settle whether a particular response was given. Even an otherwise excellent transcript may miss a critical detail. Acknowledging that systems make errors is of little use if the output is treated as the conclusive record when errors are challenged.

I’m deeply concerned by the level of authority we give to these documents. Automated transcription seems like a small administrative task, but when we rely on those transcripts to determine what occurred, the power of holding the pen falls out of human hands. We are giving AI a more consequential role than many of us realise.

Written accounts should be something we can rely on. That depends on being able to challenge them. A transcript may be more accurate overall than a person’s recollection, but where they disagree about a consequential detail and no independent record is available, we cannot settle that disagreement simply by referring back to the transcript. Its general accuracy does not establish its accuracy at the point in dispute. In that situation, it should face the same scepticism as the recollection it contradicts.

When I correct a counterparty’s summary, I am asking for the document to reflect the agreement. I need and expect that request to be taken for what it is, and not treated as a request for a different agreed position.

When correcting the record is treated as changing the agreement, the record has become the agreement.