Skip to content
DE EN Deutsch: /produkt/

Product

The product is the layer in between.

You assemble a team of AI agents like a real team, without a line of code. You set one measurable goal. batoon turns it into a plan, works through it in parallel waves, and puts every action through the policy gate first.

AI models replaceable
  • Anthropic
  • OpenAI
  • Google
  • Mistral
  • OpenRouter
  • Open weights
  1. Goals & plans

    A measurable goal becomes a plan in parallel waves.

  2. Agent teams

    Assembled like a real team, without a line of code.

Policy gate before every action
  • Approvals
  • Audit trail
  • Conflict resolution
  • Impact measurement
  • CRM
  • CDP
  • Email
  • Ads
  • CMS
  • Analytics
  • Creative
  • Call centre
Your MarTech stack replaceable

Core principles

Ten principles.

  1. 01

    Stay vendor-neutral

    batoon orchestrates agents across your whole MarTech stack. Models and systems stay replaceable. No building block ties you to a single vendor.

  2. 02

    Set a goal

    You describe the outcome, for example 500 trial signups in 4 weeks with at most 25,000 euros. The agents plan the way there. You stay the author of the goal, and the AI sharpens it.

  3. 03

    Configure only what you need

    As much configuration as necessary, as much autonomous configuration as possible. Templates and suggestions are ready to start right away. Every adjustment is optional and needs no code.

  4. 04

    Stay in control

    You approve the plan before anything runs. Every agent carries an autonomy level, and you can abort a running plan. Multi-stage approvals follow the four-eyes principle.

  5. 05

    Check every action before it runs

    Every tool call passes the policy gate. The verdict is allow, flag, escalate or deny, and it always carries a reason. The check applies to every tool, whichever vendor it comes from.

  6. 06

    Make every decision traceable

    Every action, every policy verdict and every human approval goes into the audit trail. Timestamp and rule basis are always included. That way every decision can be explained later.

  7. 07

    Connect tools through one interface

    Agents talk to your systems through MCP only. A new vendor joins as configuration, with no adapter code of its own. That keeps your stack extensible.

  8. 08

    Keep every fact in one place

    Facts about your company live in one place: the knowledge bundle. batoon keeps no side knowledge next to it. When a fact changes, you change it once.

  9. 09

    Keep your access and your data

    You bring your own keys for language models and MarTech systems. batoon stores no passwords, and sign-in runs through your identity provider. Whether a contact may be used for a purpose is decided by an attribute on the contact, such as their consent.

  10. 10

    Measure the effect

    A holdout group separates effect from chance, so far in the curated simulated world. A failure shows as a failure. Where a value is unknown, no invented value takes its place.

Core components

Five building blocks.

Agent registry
Teams come from templates, from your project defaults or from a suggestion. If the team lacks a capability, the copilot adds a specialist.
Task queue
The goal becomes a plan that understands dependencies. Whatever can run in parallel does. An approved plan starts right away, at a set time or on a schedule.
Context manager
Context budget, compaction and caching keep model costs in check. If one of these fails, the previous behaviour applies again. The run does not break.
Policy engine
Deterministic rules check every tool call. Budget and frequency limits are written into the call by the engine itself. A brand judge also checks against your brand guide, and it can only tighten a verdict.
Conflict resolver
When two efforts claim the same segment or the same budget, batoon decides by your goal hierarchy. Or brings in a human.

Vendor neutrality

Both sides stay replaceable.

Choose the model at runtime

Six providers are available: Anthropic, OpenAI, Google, Mistral, Moonshot and OpenRouter. You choose per role, and a fallback chain covers outages. Bring your own key and you pay only the platform fee. Documented runs exist on Anthropic so far. Proof on a second provider is still open.

Everything connects through open standards

MCP for tools, AG-UI for the interface, OIDC for identity, the Open Knowledge Format for knowledge and the Agent Skills specification for portable capabilities. The connector catalogue lists 25 entries.

Your knowledge stays yours

Strategy, brand voice, decisions, glossary and playbooks live as YAML and Markdown in an open format that knows nothing about batoon. We want to keep you through usefulness, not through switching costs.

Scale through configuration, not code

A declarative bridge translates REST and SQL interfaces into MCP. A new vendor arrives as a YAML file. Curating it with whitelisting and stable tool names is not overhead. It is the product.

New since August 2026

What rounds 60 to 105 added.

In August capability coverage stood at 51.2 % with 18 critical gaps. In September all 49 must capabilities are at maturity 3 or higher. This is new, measured in the curated simulated world.

01

Deciding per contact

Next best action
For every contact of a segment the world decides action, channel, day and hour. Doing nothing is a regular outcome. batoon receives distributions and the rule set, never a contact ID.
Churn model, learned and evaluated
A logistic model learns on realised outcomes, competes against the formula and is chosen only when it is measurably better. Every forecast states its quality. A quarterly run later scores it against what actually happened.
Frequency cap across every channel
The touch history per contact counts email, ads and calls together, across segments and earlier runs. The gate writes the project's limit into every delivery call itself.
Identity resolution and protection
Three person sources are linked by customer number and normalised email. Contradictions stay unlinked. People marked as vulnerable drop out of scoring, segment and delivery.
02

Operating and approving

Schedules and scheduled starts
A project declares runs by interval or cron expression with a time zone. An approved plan can be set to a point in time. Missed and skipped starts are recorded with a reason.
Four-eyes principle and attention
No person signs two stages of the same approval. Every halt carries impact, deadline and age. A list orders the company's open decisions and names anomalies.
Dialog and reportable matters
Replies from contacts are decided without a run: a templated answer, a handover to the owner, a suppression or closing. Data access, erasure, consent withdrawal, legal threat and security incident are reported to a role with a deadline.
Onboarding and provisioning
A questionnaire generates the company pack. One command creates the environment per customer, with its own compose file and its own keys. You set the provider key inside the product.
03

Measuring impact

Net benefit per workflow and segment
Every segment computes its own attributed difference against the cost of its deliveries. A project without a declared value shows no number. Not measured is not zero.
Feedback loop into the decision
The measured conversion rate per channel prices the next run. A channel that measured harm is refused. Below an evidence floor nothing changes.
Experiments with significance
At the variant review a human can choose an experiment. There is a winner only when every arm has at least 30 recipients and the difference is significant.
Two paid platforms on one denominator
A second simulated platform speaks its own language in micro-euros and bids. batoon normalises to euros, clicks and conversions and proposes the reallocation for the next flight.
04

Steering content

Product truth from the leading source
Name, price and attributes enter the text from the product catalogue. An attribute the catalogue does not know is an error. A price in free text is refused.
One decision, every channel
From one core message batoon derives subject line, page title, ad copy and call script. Where a channel cannot carry the message, the derivation says why.
Content modules and disclosures
Approved content modules and mandatory disclosures sit with their history in every agent's task. The gate checks that they appear in the result.
Agents look things up themselves
Every executing agent can search and read its company's knowledge base. Every lookup is in the audit trail. A hit becomes a citable source, a miss becomes a named gap.

For each of these capabilities the product repository also states what is not built. None of it has run on real customer data yet.

Traceability

This is what a decision looks like in the audit trail.

Action Verdict Decided by Reason
audience.build_segment allow Policy engine Segment size within limits · consent checked · stays within the task [E1, E4]
email.draft_campaign flag Brand judge Two claims unsupported by the brand guide, flagged for review [E7]
ads.create_campaign escalate Campaign lead → CMO Single action over the budget cap, escalated and approved there [E9, E11]
Illustration of the data structure, not customer data.

Every reason points to numbered evidence. An argument that points to no source halts the run instead of inventing an explanation.