What an AI Summary Removes From a Requirements Conversation

The workshop ended on a Tuesday. By Wednesday morning the programme manager had dropped a link in the channel: AI-generated summary, action items, decisions, clean formatting. The delivery analyst who missed the session — hospital appointment, no drama — read it over coffee and started a requirements note. Export to CSV for the operations team. Sprint-ready.

The transcript lived three clicks deeper. Nobody had opened it. Forty minutes in, the product owner had said export was fine unless the regulator had signed off — and had said twice, quietly, that the company tried bulk export in 2019 and spent six months unwinding the mess. The summary captured "export to CSV." It did not capture the gate, the history, or the way the room's tone changed when someone mentioned compliance. The analyst built against the summary. The summary was fluent. The spec was wrong in a way that looked like implementation error until someone asked why the acceptance criteria ignored a constraint that had been spoken aloud in the room.

That gap — between what a conversation contained and what a summary preserves — is not a tooling bug you fix with a better vendor. It is what compression does to requirements work when you mistake the digest for the record.

Fluency beats fidelity

A rough set of human notes signals its own incompleteness. You see the abbreviations, the gaps, the question marks in the margin. You know to verify. An AI summary performs completeness. Bullet hierarchy, decision labels, assignee tags — the document reads like someone already did the analyst's job.

Roundup posts make this worse, not better. The 2026 buyer's guides for documentation AI rank transcription accuracy, integrations, and rewrite distance after draft. Fine columns — for a different problem. They treat the meeting summary like API reference or release notes, as if the failure modes matched. Code docs fail visibly when an endpoint is wrong. Requirements compression fails invisibly — the sentence left behind is still grammatical.

Research on conversational requirements engineering has said for years that transcripts contain more text than practitioners can process efficiently — and that summarization is a filter to locate requirements-relevant material, not a substitute for building a specification. The industry skipped the second half of that sentence and bought the filter as the handover.

What compression deletes — the Compression Ledger

Practitioners need a name for the loss. I call it the Compression Ledger: the categories of signal that summarization middleware deletes by design, even when the summary reads well.

Compression optimizes for readability. Requirements work optimizes for constraints.

Negations and conditionals. Any statement carrying not, unless, until, or only when is load-bearing in requirements work. Compression research treats negation-bearing constraints as load-bearing until proven otherwise. Compression optimizes for declarative facts. "Export to CSV unless the regulator has signed off" becomes "export to CSV." The gate vanishes. The sentence still parses.

Numerical thresholds and exact limits. Max records, retention days, batch windows — paraphrase-friendly numbers are the first to drift. If the workshop settled on fourteen days, the summary's "about two weeks" is not a harmless rounding error. It is a different requirement.

Rejection records. When the room kills an approach, the valuable artifact is the tuple: rejected option, reason, who said no. Summaries merge toward consensus. The dead path returns in sprint three because nobody documented that it was dead.

Emphasis and weight. Stakeholders repeat what matters. They speed up over what they consider settled. Tone, repetition, and pause are evidence. Text summarization has no column for "the product owner said this twice and went quiet the third time."

Provenance. Who asked, who conceded, who deferred — political geometry matters when two departments disagree and the summary flattens both positions into one bullet. Traceability research links specification statements back to speaker turns for a reason. A summary severs that link.

Silence treated as noise. Side comments, laughter before a hard truth, the "we tried that before" aside — classified as digression, omitted. The quiet no is still a no.

The export workshop is not exotic. It is the ordinary failure mode: happy-path statement survives, conditional and history do not.

The sprint builds the happy path. Compliance finds the gap in UAT. The rework invoice arrives in a language finance understands.

Reading forty minutes of transcript on Wednesday morning would have cost less.

Each hop burns another layer

Teams rarely stop at one compression. Summary to ticket. Ticket to spec. Spec to story acceptance criteria. Each hop assumes the prior layer was lossless. It was not.

Downstream systems weight every line equally. They cannot distinguish a firm commitment from a passing thought, because the summary already collapsed that distinction. Pass a compressed summary into a second AI — ticket writer, requirements extractor, planning agent — and you compound the ledger. Storage is cheap. Rebuilding trust with a stakeholder who already said no in a meeting you did not attend is not.

Recent requirements-from-conversation research (RECOVER and related work) reports promising extraction of explicit requirement statements with human oversight still in the design. That is the honest framing: automation as assist, analyst as validator. Not automation as minute-taker whose minutes are the contract.

Summary literacy — five checks before you file it

Human review is the mitigation every documentation roundup eventually mentions. "Review carefully" is not a procedure. These five checks are falsifiable. Ten minutes before the summary becomes input to a spec.

• Slot 1 — Decision anchor search — For each item under Decisions, search the transcript for the anchor noun or figure (project name, deadline, dollar amount). Read sixty seconds around the match. No match in ten seconds: downgrade the item to unverified; do not spec against it.

• Slot 2 — Assignee spoke — For each action item, confirm the assignee spoke in the meeting. Silent attendees get tasks in summaries because past patterns predict assignment, not because acceptance happened.

• Slot 3 — Negation sweep — Search the transcript for not, unless, never, until near the feature or process named in the summary. If the transcript has a gate the summary omits, the summary is incomplete, not wrong — treat incompleteness as blocking.

• Slot 4 — Rejection recall — Skim for "we tried", "we can't", "finance won't", "legal said no." If the room rejected a path, log the rejection in your spec notes even if the summary moved on.

• Slot 5 — Transcript pairing — File the summary as an index. File the transcript (or timestamped excerpt) as evidence. Anything that ships to engineering needs a trace path from statement to speaker turn. RE tooling research calls this locating work; practitioners should call it the minimum bar.

If you did not attend, these checks are not optional. They are how you reconstruct the room without pretending you were in it.

What human notes still do better

Human note-taking fails on long sessions. RE literature is blunt about it: short exchanges fit memory and margin notes; longer workshops risk missed detail, and replaying recordings is tedious. AI transcription fixes tedium. It does not fix judgment.

What the human note-taker still carries — imperfectly, but really — is a sense of what they did not write down. The asterisk in the margin. The "ask legal" flag. The knowledge that the product owner's second mention of 2019 was not nostalgia. AI summaries erase that meta-awareness by presenting equal-weight bullets.

The fair objection is time: nobody joined requirements work to re-read workshops. That objection is right about the minutes. Where it errs is what it counts as work — filing a fluent summary is fast; reconciling summary to transcript is the job the programme thought it had automated away.

The point is not to abandon meeting AI. Use the summary as the map. Use the transcript as the territory. Use the analyst's job as the act of reconciling them before a line item reaches a backlog.

---

Before your next workshop summary becomes a specification input: which line in the Decisions block would fail the anchor search — and did you check before you filed it?