The Case Against AI Summaries for Deep Work

The Compression Problem

Every AI summary is lossy compression.

60 minutes of conversation → 500 words.

Information is removed. By definition.

The question isn't whether things are lost—it's whether what's lost matters.

What Summaries Lose

Exact Wording Summary: "Client expressed concern about timeline." Actual: "If we miss the March deadline, this whole deal is dead."

Different implications. Different urgency.

Voice and Tone Summary: "Client was frustrated with current solution." Actual: "I want to throw my laptop out the window every time I use this."

Same information, completely different emotional weight.

Tangents Summaries keep "main points." Tangents get cut.

• •What people actually care about
• •Information they think is secondary but isn't
• •Context that explains the main points

Hesitation and Uncertainty "I think" vs "definitely" "We're planning to" vs "we're committed to"

Summaries often flatten these distinctions.

When Summaries Are Fine

• •"What action items came out of this meeting?"
• •"Who owns what?"
• •"When's the next call?"

Summaries work great here. You need facts, not nuance.

When Summaries Fail

• •Creating content in someone's voice
• •Understanding someone's actual position
• •Building long-term context
• •Providing AI with complete information

These need the raw material, not the processed version.

The Summary-of-Summary Problem

Here's something subtle.

When AI answers questions about past meetings, it often reads summaries—not transcripts.

You ask: "What did Client X say about competitors?" AI reads: Its own previous summary mentioning competitors. You get: A summary of a summary.

Each layer of compression loses fidelity.

Full transcripts → single search → original words. Much more reliable.

The Storage Argument

"Full transcripts are huge and hard to manage."

Actually... no.

A 60-minute transcript is maybe 10,000 words. About 60KB. A year of weekly client calls: ~3MB.

Storage is not the constraint.

A Hybrid Approach

You don't have to choose.

Save full transcripts (for depth when needed). Generate summaries (for quick reference). Search transcripts directly (for accuracy).

The summary is an index. The transcript is the source.

Use the right one for the task.

My Rule

For anything that will be read once and forgotten: summary is fine.

For anything I might reference again: full transcript.

For anything where precision matters: always check the source.

The cost of keeping transcripts is near zero. The cost of not having them when you need them can be significant.

Default to keeping everything. Let search handle retrieval.