Michael Daugherty on Quill and the AI Chief of Staff

Michael Daugherty on Building Quill Into an AI Chief of Staff

Michael Daugherty is building Quill around a deceptively difficult question: what should happen after a useful conversation?

A transcript preserves words. A summary compresses them. Neither one necessarily tells a founder what to decide, an engineer what to fix, a seller what to enter in the CRM, or a manager what to revisit before the next one-on-one.

In Venture Step episode 108, Daugherty explains how that gap led Quill from private meeting notes toward a broader AI chief-of-staff product. The vision is compelling because meetings contain operating context that usually disappears. It is demanding because context becomes risk as soon as a system can remember, infer, and act.

The same meeting does not create the same work

Quill began as a meeting note taker, but Daugherty did not like the default idea behind most recap products. A single bot joins the call and gives everyone the same result.

Consider a product sales call. The buyer needs an internal case: what the product does, how it fits, what evidence remains missing, and who must approve it. The seller needs a different working view: decision makers, objections, promised follow-up, timing, and the next CRM update.

Both views must agree about what happened. They do not need to select the same details.

Daugherty says Quill uses meeting type, the user's role, and saved formats to recommend a document. In his recruiting example, the product can apply a recurring evaluation rubric and assemble evidence from the conversation. An internal meeting can produce a different output without forcing the user through a separate copy, prompt, and paste routine.

That distinction has research support. A Microsoft study of an LLM-powered meeting recap system found that fixed-length summaries do not meet every recap need. Its seven-user evaluation is too small to prove a universal design, but it supports the idea that highlights and structured minutes serve different contexts.

The stronger model is one shared decision record with several role-specific next-context views.

From note taking to context

The next step is not a longer summary. It is memory across conversations.

Daugherty describes Quill's home screen using upcoming meetings and prior interactions to help a user prepare. If the last conversation ended with a promise, the system can surface it before the next call. Cross-meeting search can consolidate what different people have said about the same project, customer, or decision.

Quill's current documentation describes meeting recording, local transcription, speaker recognition, recommended documents, cross-meeting intelligence, model choice, and opt-in sharing. The current homepage positions the product as an AI chief of staff.

The episode extends beyond those current descriptions. Daugherty imagines a system that learns which suggestions a user accepts, understands how connected tools are organized, creates its own recurring extensions, and eventually coordinates specialized agents. Those statements describe a direction. They should not be mistaken for a current feature guarantee.

flowchart LR
    A["Conversation"] --> B["Shared meeting record"]
    B --> C["Role-specific next context"]
    C --> D["Cross-meeting memory"]
    D --> E["Recommended action"]
    E --> F["Human approval"]
    F --> G["Verified tool write"]
    G --> H["Future preparation"]

The interesting product is the whole loop. The dangerous product skips the approval and verification states.

Why Daugherty calls it a chief of staff

Daugherty makes a word-order distinction in the interview. He says an AI chief of staff sounds like a system that manages people. He prefers "chief of AI staff" for a system that helps one person coordinate a growing collection of AI agents and projects.

Quill's current public site now uses "AI chief of staff." The labels have converged in marketing, but the distinction remains useful.

A software chief of staff should manage context and coordination for a person. It can prepare a meeting, surface a commitment, recommend a next step, draft an artifact, and invoke an approved tool. It does not inherit the person's managerial authority or professional accountability.

Daugherty is explicit about the responsibility gap. People cannot blame a failed deliverable or destructive action on an assistant that apologizes. Judgment still belongs to the human and organization that chose the system, gave it access, and accepted the result.

Read [[What Is an AI Chief of Staff]] for the category and authority model.

The AngelList operating context

Before Quill, Daugherty spent roughly a decade at AngelList. Quill's About page says he helped build Syndicates and worked as AngelList Venture grew beyond $1 billion in assets supported.

AngelList's current company history records the 2013 launch of Syndicates and the later billion-dollar platform milestone. A current syndicate page identifies Daugherty as a lead alongside Judy Yang.

The connection matters because venture infrastructure depends on coordination across founders, fund managers, investors, legal entities, records, deadlines, and approvals. Meetings can contain high-value context, but the canonical work still lives in systems with defined owners.

The product lesson is not that every conversation should become the source of truth. It is that a conversation can become a better input to the systems that already own the record.

MCP makes the integration legible, not automatically safe

Daugherty describes Quill connecting to Linear, Airtable, Obsidian, Notion, and other tools. He also describes a local MCP server that lets compatible applications search Quill meeting data.

Model Context Protocol gives an AI application a standard way to discover resources, prompts, and schema-defined tools. The current MCP documentation describes tools as typed operations a model can request.

In Daugherty's example, an agent might notice that a user discussed discarding an old laptop, inspect an Airtable base used for electronics, and recommend deleting the corresponding row. The user sees the recommendation and approves it.

That last step carries the authority.

The protocol can describe the delete operation. It does not know whether the meeting statement was final, whether the row represents an asset record that must be retained, whether the user owns the table, or whether deletion is reversible. Official MCP security guidance treats consent, scopes, redirect validation, and local server execution as security boundaries.

The practical rule is simple: capability describes what can be done; policy and approval decide what may be done now.

Local-first changes the trust relationship

Quill records and transcribes on the user's computer by default. Its documentation says users can choose local models, personal API credentials, private endpoints, or Quill cloud processing.

That architecture can reduce vendor dependence and preserve offline access. It also lets users choose different model routes for different kinds of work.

Daugherty compares the trust effect with Obsidian. People are more willing to invest in a system when they know they can reach and move their data. The foundational local-first software essay makes the same broader case for offline work, longevity, collaboration, privacy, and user control.

Local does not mean invulnerable. A stolen endpoint, lost disk, forgotten key, untested backup, external model request, or accidental share can still expose or destroy information.

Quill's privacy policy says encrypted copies may be stored when sync or sharing is enabled. In the episode, Daugherty describes users discarding an old computer and then discovering that local ownership also meant the vendor could not recover the only copy. The company had built encrypted sync and was rebuilding its recovery and multi-device experience at the time of the conversation.

Control includes the power to keep data and the responsibility to recover it.

Recording is only the first consent decision

A meeting assistant can operate without placing a bot in the participant list. That improves the experience. It can also make recording less visible.

Federal and state rules are not identical. Federal law includes a one-party-consent provision in 18 U.S.C. 2511, while California Penal Code section 632 generally requires every party's consent for a confidential communication.

Location, subject matter, workplace policy, privilege, and contract can change the analysis. Recording consent also does not automatically authorize indefinite retention, cross-meeting profiling, model processing, sharing, or writes into another system.

Local processing does not erase participant rights. It changes who holds the data.

Customer development becomes part of the architecture

Daugherty describes advanced Quill users connecting the product to their own workflows and requesting new MCP functions. The company can learn from those power users, make a capability available inside the product, and reduce the technical work required for the next group.

That is a useful startup advantage. A focused team can treat the meeting workflow as its core data model rather than a feature that ships during one quarterly cycle.

It also creates a discipline problem. The loudest customer request is not automatically the product strategy. A new connector or extension can expand the permission surface faster than it expands durable value.

The right loop combines observed use, repeated need, safe defaults, clear ownership, and evidence that the workflow improved.

The episode's durable idea

Quill's larger bet is not that meetings need prettier notes. It is that conversation can become structured context for the next decision.

That requires more than an accurate transcript. The system must know which record is shared, which view is personal, which memory is stale, which tool is canonical, which action is merely suggested, who can approve it, and how to correct the result.

Daugherty's most useful word is control. A good AI chief of staff should give a person more control over preparation and follow-through without quietly taking control over people, records, or consequences.

Read [[Why Generic Meeting Summaries Fail]], [[What Local-First Means for an AI Meeting Assistant]], and [[How to Turn Meetings Into Actions Without Losing Control]] next.

AI assisted with research organization and drafting. Dalton Anderson remains responsible for the analysis and publication decision.