Opening Haven
Agent Operations Intelligence

Agent operations, run as one system.

Your people and your AI handle the same customers on different stacks, scored by different definitions of good. Haven reads both against one standard, reasons through a model of how operations actually behave, and drafts the fix. Nothing ships without your call.

Get your operation's first read

5 minutes. Free. The read in your inbox today, your place on the launch list with it.

Below A sample operation, replayed. The desk is yours. Everything you see is what an operator sees on a normal Tuesday morning.
Operator home Tue · 09:14
Replay

Morning. Here's what Haven surfaced while you were away.

Since you left overnight → now
8 signals overnight 1 worth your time now 12 drafts pending your call volume last 24h · humans 38% · AI 62%
Worth your time Surfaced 27 min ago

Refunds are up 31% since Friday's dispatch delay.

Roughly €12k in exposure if the trend holds another five days. The APAC shift sees it first. No SLA breach yet; there is headroom this week.

source · Measure, refund-rate anomaly traced · 4 functions drafts ready · 2
How the signal moved One root cause · four functions touched · overnight
  1. F4 · Measure Refund-rate anomaly on the dispatch-delay cohort Source
  2. F3 · Perform Tone drift on refund denials, both populations read Read
  3. F1 · Enable Refund macro v2 drafted, empathetic opener Drafted
  4. F5 · Improve SLA review on the billing tier Suggested

This is the layer. A QA tool would have scored the tone. An analytics dashboard would have charted the refunds. Neither connects them, and neither drafts the fix.

What you're looking at Drafts are Haven's unit of work: not an alert, a written intervention with the evidence attached. Each one carries the same anatomy: the score against your standard, the baseline it broke from, the projected impact, and the window in which Haven verifies the fix held.

Pending your call.

12 active · 2 from this signal · 1 verified this morning
Coaching note · for a billing-shift agent (human) 22m ago Pending

Led with policy on ticket #58291, a disputed refund. Third instance in 7 days, outside this agent's own 30-day baseline. Suggested note leads with acknowledgment before the policy line.

Metric 04 · Empathy in opener · threshold 0.65
Today 0.42
Own 30d 0.71
Threshold 0.65
Peers 0.78
Projected, if applied CSAT 4.2 → 4.5 escalations −2 / wk repeat contact −4 pts
Verifies on next 5 escalations · 7-day window
Macro update v2 · for the AI cluster on refund denials 20m ago Pending

Refund-denial replies trending terse since macro v1. Drift is localized to this sub-intent; every other intent is on baseline. v2 adds the same empathetic opener the human standard already carries.

Metric 04 · Empathy in opener · threshold 0.65
Last 200 0.51
Cluster 30d 0.74
Threshold 0.65
Human peers 0.79
Projected, if applied CSAT 4.1 → 4.4 repeat contact −7 pts escalations ~−8 / wk
Verifies on next 200 closes · 48-hour window
Macro update v1 · for the AI cluster on first-touch greetings Applied yesterday · 08:21 Applied · verified

Greeting template was missing on first-touch tickets. You applied v1 yesterday morning. The cluster has since closed 200+ replies on it.

Metric 01 · Greeting present · threshold 0.85
Pre-apply 0.76
Post-apply 0.92
Threshold 0.85
Human peers 0.94
Verified · lift held over 200 closes · 48h window
Why this isn't a black box Every read traces back through an explicit model. The section below is the actual reasoning behind the refund signal, edge by edge.
The operating graph · how Haven got there

Four edges, walked in order. Nothing invented.

  1. Dispatch delay, Friday. An operational event enters the graph as a known stressor.

    edge · ops-event → refund-intent · prior 0.82 · lag 2–4 days · effect +0.6σ

  2. Refund intent rises on schedule. The anomaly lands inside the predicted lag window, which is why it surfaced as signal and not noise.

    edge · refund-intent → tone-pressure · prior 0.64 · lag 0–2 days

  3. Tone degrades where volume concentrates. The AI cluster takes the refund-denial spike; its empathy score breaks baseline first. The humans show the same drift, smaller.

    edge · tone-pressure → standard-breach · observed in both populations

  4. The drafts target the cause, not the symptom. Macro v2 for the cluster, a coaching note for the shift, an SLA review suggested downstream. One root cause, three interventions, each with a verification window.

    drafted · 2 pending your call · 1 suggested

The graph holds roughly fifty relationships like these across the seven functions, each with a prior, a lag, and an effect size. The LLM does the reading and the writing; the graph decides what's plausible; your operation's data sharpens both the longer Haven runs. Inspectable end to end.

Where this operation sits, today.

7 functions · mapped Reactive → Emerging → Defined → Optimised
What this is not

Three categories already sit in your stack. None of them is a layer.

Haven reads, reasons, and drafts across all seven functions, humans and AI both. You decide what ships.

Get your operation's first read.

Five minutes of questions. A read of your operation against the seven functions, like the one above, in your inbox today. Free. No call, no follow-up.

Start the first read
The platform is next · one message when it opens
Enter a valid email to continue.
Invalid access code.
You're on the list.
One message when the platform opens. No drip.
Privacy Policy

Who we are

Haven is a product of The How Company B.V., registered in Amsterdam, the Netherlands. We are the data controller for any personal data collected through this website. For the full privacy policy, see haven-does.com/privacy.

What we collect on this page

If you join the waitlist, we collect your email address. If you enter an access code, we set a single signed cookie (haven_access, 30-day lifetime) so you don't have to enter it again. The cookie is strictly necessary for the access feature and contains no tracking data.

Why we collect it

Your email is used to send a launch notification when Haven opens and occasional product news. The legal basis is your consent (GDPR Article 6(1)(a)), recorded with the exact wording of the checkbox you agreed to. The access cookie operates on legitimate interest under Art. 6(1)(f) and falls within the strictly-necessary exemption of the ePrivacy Directive.

Sub-processors

Supabase (hosted on AWS, EU West) stores the waitlist record and consent log. Netlify hosts the site and runs the form-handling functions. Resend sends the confirmation email. Netlify and Resend are US-based companies. We rely on the EU-US Data Privacy Framework and Standard Contractual Clauses for these transfers.

How long we keep it

We retain your email until you unsubscribe, you ask us to delete it, or 12 months after collection if Haven hasn't yet opened, whichever comes first.

Your rights

Under the GDPR, you have the right to access, correct, or delete your personal data, withdraw consent, request portability, and lodge a complaint with the Dutch Data Protection Authority or your local supervisory authority. To exercise any of these, email privacy@haven-does.com.

Data security

All data is transmitted over HTTPS and encrypted at rest. Access to production data is limited to a single administrator with multi-factor authentication. We log function activity and review access patterns regularly.

Last updated · 28 April 2026 · v1.1