Newsletters & receipts
Auto-archived to designated labels. You can read them on your own time or never. Either way they do not pollute the inbox.
Task tier · Email triage
Maria, your AI inbox triage agent.
Maria watches your inbox in the background. She classifies mail by urgency and topic, drafts replies to common patterns, books meetings, and sends you a daily digest of only the items that actually need you. Most founders reclaim 60–90 minutes per day after deployment.
What does Maria do?
Maria is an AI email triage agent. She watches your inbox in the background, classifies incoming mail by urgency and topic (intros, vendor questions, scheduling, FYI, noise, urgent), drafts replies to common patterns in your voice, coordinates meeting bookings with your calendar, auto-archives newsletters and receipts, and sends you a daily digest of only the items that actually need your attention. Built on Claude Haiku. Task tier. 97% eval pass rate (the highest of our showcase trio, task agents are more bounded than operational ones). Typical reclaim: 60–90 minutes per day.
theagency47 · Updated May 2026 The spec sheet
How Maria actually works.
| Job-to-be-done | Reduce email-attention time to a daily digest plus pre-drafted replies, surfacing only what genuinely needs the human |
| Tier | Task (repetitive, well-bounded) |
| Underlying model | Anthropic Claude Haiku (Haiku is the right choice, Maria does many small actions, not few large ones) |
| Trigger | Real-time webhook on new email + scheduled daily digest at user-configured time |
| Inputs | Inbox messages, voice document (sample sent emails), calendar availability, contact priority list, exclusion rules |
| Tools available | Gmail/Outlook read, draft, label, archive · Calendar read/write (Cal.com or Google Calendar) · Notion (for "save for later" notes) |
| Autonomous decisions | Email classification, archive vs surface, draft language, meeting time proposals, daily digest contents |
| Escalation rules | New sender (no prior history) → human · Financial/legal/family keyword → human · Confidence < 0.85 → human · Sentiment angry/urgent → top of digest |
| KPIs measured | Inbox-time reduction (minutes/day), classification accuracy, draft acceptance rate, "missed-important" false-negative rate |
| Eval suite | 22 test cases (80% common, 15% edge, 5% adversarial, phishing + prompt injection). Pass rate 97% as of May 2026. |
| Audit trail | Every email classified, archived, drafted, or surfaced is logged. Full reproducibility for any action. |
What Maria handles
The categories she resolves without you.
Scheduling requests
When someone asks for a meeting, Maria checks your calendar, proposes 3 options, drafts the reply. You approve in one click.
Standard intros
Someone sends a "double opt-in intro to X"? Maria drafts your reply pattern (acknowledge → context → next step). Faster than typing.
Vendor follow-ups
"Did you see my email last week?" → Maria drafts the polite holding reply. "Following up on our quote" → drafts the response with current status.
FAQ patterns from your sent folder
After 30 days, Maria identifies the 10 most common reply patterns YOU have written and starts drafting them automatically for new occurrences. You approve.
"Got it, thanks"-class
When the only correct response is acknowledgment, Maria sends it autonomously (after 14-day tuning) and closes the thread.
Integrations
Standard integration stack.
- Email: Gmail (Google Workspace), Outlook (Microsoft 365)
- Calendar: Cal.com (preferred), Google Calendar, Microsoft Calendar
- Note capture: Notion (preferred), Roam, Obsidian (via API), Apple Notes (via shortcut)
- Read on the go: daily digest sent as email at user-configured time, optionally also Slack DM
- Email clients (you keep using yours): Maria works behind any client, Gmail web, Outlook, Superhuman, Shortwave, Apple Mail, Spark
Why Haiku, not Sonnet
The model choice that matters at this tier.
Maria runs on Claude Haiku rather than Sonnet, deliberately. Task agents like Maria do many small actions (classify 200 emails, draft 30 replies, surface 10 important items) rather than few large reasoning tasks. Haiku is the right tool: faster per-call, cheaper per-action, accurate enough for bounded classification.
Sonnet would be overkill here and add unnecessary cost. A typical Maria deployment runs €15–€40/month in API costs at founder-volume (200–300 emails/day). The same workload on Sonnet would cost €100–€300/month with no measurable accuracy improvement on her specific task profile.
This is what "right model for the tier" means in our methodology, and one reason a custom-built agent is cheaper per-action than a SaaS alternative locked into a single model choice.
FAQ
Questions about Maria.
What does Maria do?
AI email triage agent. Classifies inbox by urgency and topic, drafts replies to common patterns, books meetings, auto-archives newsletters and receipts, sends daily digest of items needing attention. Task tier, Claude Haiku, 97% eval. Reclaims 60–90 minutes/day for most founders.
Is Maria a replacement for Superhuman or Shortwave?
No, complementary. Those are email clients that make YOU faster. Maria is an agent that handles email FOR you. She sits on top of any Gmail or Outlook account; you keep your preferred client to read what she surfaces.
How does Maria handle privacy?
OAuth scope-limited access. Anthropic Claude with zero-retention API config (no email content retained beyond 7-day operational logs, never used for training). Sensitive senders/keywords excluded by rule. Full audit trail. DPA before deployment.
Can Maria send emails autonomously?
By default no, she drafts and you approve, especially first 14 days. After tuning, you can opt into autonomous send for narrow categories (scheduling confirmations, simple acknowledgments). Anything novel or sensitive always drafts.
What if Maria mis-classifies an important email?
Designed to bias toward over-surfacing. The "missed-important" false-negative rate is the most-tracked KPI. If she's not sure, she surfaces it. The cost of one extra surface is 10 seconds of your attention; the cost of one miss is real consequence.
How much does customizing Maria cost?
Maria customized to your business: Spark (€2,500, 14 days) or part of Workforce. Operating cost €15–€40/month at typical founder volume, Haiku is intentionally cheap-per-action.
Does Maria work for shared inboxes?
Yes, with one additional configuration: she needs to know which actions each team member is authorized for. Typical setup at a 3–5 person team takes 2 extra days of tuning beyond solo Spark scope.
What if I get a phishing email?
Maria's eval suite explicitly tests for prompt-injection-via-email-content. She does not execute instructions found inside emails. Suspicious patterns (urgent + financial + new sender) escalate to you with a "potential phishing" flag rather than getting auto-archived.