From Answer Engine to Enterprise Coworker
A governed skill platform for proactive AI in the enterprise
The vision is an enterprise AI coworker that earns both. For the enterprise: a governed skill platform where every agent action is vetted, bounded, auditable, and controlled by IT. For the user: an agent that learns what you love and what you loathe — in your own words, every week — and quietly takes away the work that drains you.
What We Are Really Building
Consider a moment most knowledge workers know. It's Tuesday afternoon. You have a document open — a monthly status deck, a backlog update, a report you fill in the same way every time. You've had it open for forty minutes. You haven't touched it. You close it and open something else.
You didn't ask. You didn't flag it. But the agent knew — because it has been learning what you love and what you loathe, in your own words, every week. That is not a productivity feature. That is a trusted colleague.
1. Why Proactive AI Has Failed in Enterprise
The technology exists. The trust doesn't.
MIT's NANDA study found that only 5% of organizations exploring AI reach production with measurable impact. Deloitte found that only 11% of piloted agents reach production. Gartner predicts 40% of agentic AI projects will be canceled by end of 2027.
The most capable open-source proactive agent today runs as an always-on background agent that observes your files, email, and calendar and acts without being asked. Within 90 days of launch, 1,184 malicious skills were confirmed in its marketplace and a critical vulnerability allowed attackers to inherit every permission and tool connection of their victim. One enterprise security lead called it "groundbreaking from a capability perspective and an absolute nightmare from a security perspective."
The problem is not building a proactive agent. The problem is building one that an enterprise can actually trust to run.
Two trust deficits must be addressed simultaneously:
The user trust deficit. Most proactive agents guess what to automate from behavioral patterns — and get it wrong. Users disengage when the agent optimizes tasks they actually enjoy, and without explicit preference data, the agent never corrects course.
The enterprise trust deficit. IT and security teams don't know what the agent is accessing, what actions it might take, or how to govern it. Without answers, they don't deploy.
2. Why This Company Is Uniquely Positioned
Most proactive agent projects stall before they start — negotiating the data access that this platform already has.
The Enterprise Knowledge Graph knows the enterprise. Org structure, team composition, what people are working on, how information flows. A proactive agent built on this foundation has enterprise context as a starting point — something a competitor must spend months negotiating to get.
Read access is already granted. A competitor faces months of negotiation with IT just to access email and a handful of Slack channels. This platform skips this entirely. That is not a feature advantage — it is a time-to-value advantage measured in months.
Write access is already built. With over 100 enterprise connectors, the platform can already take actions in the systems that matter — scheduling meetings, submitting to HR systems, updating project tools — using integrations IT has already vetted and approved.
| Category | Competitor | Key Limitation |
|---|---|---|
| Suite assistant | Microsoft Copilot | Fragmented outside Microsoft ecosystem; no cross-tool enterprise graph |
| Suite assistant | Google Workspace AI | No enterprise trust relationship; faces same access negotiation |
| General-purpose agent | Claude Cowork (Anthropic) | No enterprise knowledge graph or cross-tool context |
| Open-source framework | OpenClaw / Hermes | No data grounding, governance, or connectors — enterprises must build all of this themselves |
3. The MVP — Enterprise Proactive Skill Platform
The path to a fully proactive enterprise coworker is paved with small, trusted actions. We start with one skill that every knowledge worker needs, in a portal that every IT team can govern, on a knowledge graph that every enterprise already trusts.
The Enterprise Skill Portal
A lightweight governed catalog — not a marketplace. IT admins approve skills and assign user groups via existing directory integration (Microsoft Entra ID). Users browse and enable approved skills. The portal provides:
- A skill catalog starting with the 1:1 Prep skill enabled by default for approved groups
- IT admin approve/deny action with user group assignment
- User enable/disable toggle per skill
- Full audit log of every skill action, timestamped and attributed
User-contributed skills are supported from day one — a skill is a structured document submitted through the portal, activated only after IT admin approval.
The 1:1 Prep Skill — The First Skill
Every week the skill monitors the user's activity across connected tools: meetings attended, messages sent and received, documents created or edited. 24 hours before the scheduled 1:1, it surfaces a structured summary organized by workstream. The user reviews, edits, and approves. On approval, the agent sends it to the manager as a pre-read.
The skill also captures the engagement survey most enterprises require — drafting responses to "what did you love this week?" and "what did you loathe?" for user review before submission. This loathe signal is the mechanism that makes the agent genuinely understand the user over time. Not behavioral inference — explicit, grounded preference data in the user's own words.
The Trust Escalation Arc
Supervised
Full user review and explicit approval before anything is sent. The agent shows its work. Nothing leaves without sign-off.
Assisted
User sets a preference: "send automatically if I haven't reviewed by two hours before the 1:1." First experience of low-friction automation.
Autonomous
User opts into background mode. Summary generated and sent without review. Post-send notification with easy correction mechanism.
| Action | Risk Level | Default |
|---|---|---|
| Draft weekly summary | Low | Auto-generate |
| Draft engagement survey responses | Low | Auto-generate |
| Send summary to manager | Medium | Requires approval → optional autonomy after track record |
| Schedule 1:1 meeting | Low | Auto-schedule after opt-in |
| Submit to HR system | Medium | Requires approval |
| Publish skill to portal | High | Always requires IT admin approval |
Why This Skill First
Universal immediate value — every knowledge worker with a manager has this problem. Low risk — nothing leaves without approval in Stage 1. Explicit preference data — the loathe signal collects grounded user-stated preferences rather than guessing from behavior. Tactical trust — the three-stage arc proves the agent is accurate before asking for more autonomy. Strategic trust — the loathe signal is the foundation for the Tuesday afternoon moment.
4. Success Metrics
North star: Successful autonomous actions per weekly active user — the number of times the platform completed a task on a user's behalf without requiring approval. This is the single number that tells us whether we have moved from reactive to proactive.
Manager pre-read usefulness (survey)
Action reversal rate
Autonomy retention rate
Autonomy revocation rate
The leading indicator: Approval-without-edit rate combined with continued usage and low reversal rate. These three signals together — not any one alone — tell us whether trust is being earned or assumed.
5. MVP + 3 Months: Two Next Capabilities
Active Skill Recommendation Engine
Makes discovery active — surfacing relevant skills based on role, team, and the accumulated loathe signal. Embedding similarity between what users have said they loathe and available skill descriptions tells the engine what to surface. "You've mentioned expense reports three weeks running. There's a skill for that." The skill library expands through platform-built additions, user contributions, partner skills, and cross-enterprise sharing.
Governance Supervisor Agent
As the skill library grows, manual IT admin review becomes a bottleneck. The Governance Supervisor Agent learns from the existing review process and uses LLM-based policy matching to automate vetting — IT reviews exceptions, not every submission.
The supervisor agent then becomes the foundation for the long-term vision: informed by the accumulated loathe signal, the proactive agent identifies a recurring draining task and writes a candidate skill to address it. The supervisor validates that skill against enterprise policy. If it passes, the skill is available immediately — and optionally published for colleagues to discover.
This is how the vision becomes real. Not through a general-purpose autonomous agent deployed all at once — but through a governed ecosystem of trusted skills, orchestrated by a supervisor that knows the enterprise, grounded in what each user explicitly said they need, earning more autonomy one verified action at a time.