Workflows first, AI only with a reason
Every feature has one job: keep you in control and keep costs down. ThinkShip combines a deterministic workflow layer (cron jobs, triggers, data matching) with a purpose-built escalation layer for topic maturation, self-organization and passive operation — AI only switches in once the workflow layer alone isn't enough.
Workflows save tokens
Recurring checks run as a plain cron job — 0 tokens as long as there's nothing to do. AI only switches on once something real has actually been found.
Plain language, not prompt engineering
People describe what they need the way they would to a colleague — no tool training, no prompt handbook.
Only as much AI as needed
The system decides per task whether a simple workflow is enough or real AI judgment from one or more agents is genuinely needed.
An organization that grows with you
A coordinator agent leads team leads, team leads lead specialists — a real reporting line, not a flat tool list.
Three-tier approval
Every action can be approved once, always, or never — at the level of topic, role, or single action.
Budgets & token caps at every level
Maximum values per run, agent, topic and organization, visible live on the taskboard.
Data-residency routing
Tasks involving sensitive data can be strictly routed to local models by policy — cloud models stay out of reach.
True multi-tenancy
Each tenant stores its own encrypted provider keys. No shared credentials, no data mixing between tenants.
Self-hosted, your infrastructure
An installer sets up the entire stack on your own hardware or in your own cloud environment.
Passive mode
Runs as a deterministic workflow (cron/trigger) in the background — costs nothing as long as nothing happens. Only a real match switches AI on.
Access reviews for roles & permissions
Elevated access rights that haven't been reconfirmed in over 90 days land automatically on a review list — confirm or revoke in one click.
End-to-end encryption
Sensitive content stays encrypted throughout, with a key that belongs to the tenant alone — not even the operator can read it.
More than a chat interface
These capabilities are already part of the concept and current development, not just an idea for later.
Multiple workflow entry points
Cron job, webhook, a file or feed change, an external event, or a written request — a workflow can be triggered in different ways, not just through chat.
Agents are evaluated on a regular basis
Every role has its own eval set of golden tasks. New agents run on probation with close review; anyone who doesn't hold the success rate gets retired instead of quietly continuing.
Simplification on your terms
The system spots recurring patterns and suggests what could be automated — what actually gets simplified is always your call, never a silent autonomy escalation.
Variable & object picker in the workflow editor
Pick fields straight from a connector's response or an earlier step instead of guessing or retyping field names.
Action buttons follow real permissions
If someone lacks the permission for an action, the button is disabled — with a visible reason, not just grayed out for show.
Audit log with transactional integrity
Every audit entry is written in the same database transaction as the action itself — no silent gaps in the trail.
Data erasure on request
Knowledge entries can be permanently removed from the knowledge base and vector index — supporting the GDPR right to erasure.
SIEM/log aggregation via OpenTelemetry
The same audit signal, structured for export to an external log or SIEM system (e.g. Splunk, Elastic).
Encrypted internal communication (mTLS)
Traffic between internal services (NATS, model broker) runs over mTLS with an internally operated certificate authority.
Prompt pseudonymization
Pre-configured terms (e.g. customer or company names) are swapped for random placeholders before they reach an AI model — the response is automatically restored afterwards. Placeholders are fresh per run and never stored.
What happens when an agent fails too often?
Never let it quietly continue — automatic action instead. See “Agents are evaluated on a regular basis” above.
And what if a role keeps getting it wrong? Every task runs through a golden-task eval. If the success rate drops below the configured threshold, the role is automatically retired — a new instance takes over, and the event is logged to the audit trail at the same time, instead of quietly continuing.
Planned next
Clearly marked apart: these points aren't available yet — they're the next stages of development.
Low-code workflow editor
Assemble and adjust recurring workflows visually instead of only describing every step in text.
Distributed, grouped schedulers
Load balanced across multiple scheduler nodes for high throughput of recurring jobs, grouped by tenant or priority.
Regular workflow health checks
Automated health checks for configured workflows, so a silent misconfiguration surfaces before it becomes a problem.
Telephony integration
A phone call can be routed directly to an AI agent while structured notes are created in parallel.
A growing integration network
Connections to common systems (CRM, ERP, ticketing, calendars and more) are added step by step rather than promised in one big bang.
Automatic pseudonymization
Complements the manual term list with automatic detection – validated patterns (IBAN, email, phone number) plus a locally-run NER model for names and places in free text, before prompts (including knowledge-base context) reach an AI model.
What this looks like day to day
See how a request travels from the first sentence to a finished result.