Pick your role. Read what concerns you.

Mid-size and large companies whose marketing runs across several systems: CRM, email, ads, CMS, analytics, often a call centre too. As soon as several systems and several people are involved in one campaign, the coordination problem batoon solves appears.

We are entering through regulated industries: insurance, banking, pharma and energy. There, governance decides whether a campaign may start at all. The capability catalogue batoon measures itself against consists of 13 scenarios from these industries, from an insurer to a hospital group.

C-level Marketing IT & architecture Compliance & privacy Agency

batoon models a marketing organisation with several roles. The marketing lead sets goals and approves plans. The CMO decides on budget overruns and high risk. Compliance and data protection define rules, consent requirements and reporting duties. IT runs the stack and connects systems.

Rights come in three levels: read, approve, configure. They follow from the role in the company and the role in the project together. Whoever signs an approval is read by the server from the session.

C-level Marketing IT & architecture Compliance & privacy Agency

Most likely through an agency. An agency with many mid-market clients runs each client as its own company in batoon and brings the marketing expertise smaller teams often lack.

Whether smaller companies would pay themselves is an open question. We test it in the validation phase with conversations in three segments.

C-level Marketing Agency

Yes. Every company is its own tenant in batoon, with its own knowledge base, roles, connectors and environment. A person can be a member of several companies and picks the active one after login.

A company grants guest access only to itself. An agency cannot give itself access to a client the client has not entered itself.

Agency

Not for the daily work. You write a goal in your own words, review the plan and approve. An onboarding questionnaire generates the company pack.

Rules, knowledge and configuration live as YAML and Markdown. Whoever adjusts them in detail works with text files in a git repository. That is work for IT or an agency, and it needs no programmer.

C-level Marketing

Every system in your stack is getting its own AI agents. Each one works inside its own silo. The agent in your email tool knows neither your ads budget nor your call centre. Rules that span channels cannot be guaranteed by any silo agent: budget, contact frequency, consent, control groups.

batoon is the layer in between. You set a goal. A team of agents works across the whole stack, and every action passes a policy gate first. According to Deloitte only 21 % of companies have a mature governance model, while 74 % want to deploy agents by 2027.

C-level Marketing IT & architecture Compliance & privacy Agency

It only sees its own channel. A holdout in the email platform excludes those held back from the email. The call centre rings them anyway. A frequency cap in the email tool counts no ad and no call.

Then there is the position. The vendor enforcing the rules also sells the channel being decided on. batoon sells no channel and no model.

C-level Marketing IT & architecture

A chat harness models one user. Whoever acts approves. There is no instance where somebody else has to agree, no threshold by amount or risk, and no control group across channels. Its spend limit caps what the model costs, and not what it spends.

batoon models the organisation. The plan goes to the campaign lead, the budget overrun to the CMO, the EU segment to legal. The differentiation upwards and downwards has its own page.

C-level Marketing IT & architecture

With limits in code, not in the prompt. The policy gate checks every action against the cap per single action and against the project's total budget, reserves amounts and counts spend. A single action over the cap escalates to the responsible role, usually the CMO.

Since September the gate also checks the expected net benefit of every delivery. It uses the channel's measured conversion rate where one exists. A channel that measured harm is refused.

C-level Marketing

Through a touch history per contact that counts email, ads and calls together, across segments and earlier runs. The gate puts the project's weekly limit into every delivery call itself. Whoever has reached the limit is skipped, and the run reports it.

On top come quiet hours per channel, consent per channel, suppression lists and a protection mark for vulnerable people. Whoever is marked drops out of scoring, segment and delivery.

Marketing Compliance & privacy

The conflict resolver detects the conflict before anything runs. It decides by the goal hierarchy you have defined. Where the hierarchy gives no answer, the run halts and brings in a human.

Marketing

Through four levers: coordination time, stack utilisation, approvability and budget discipline. For each, the ROI page states the pain with its source, the mechanism in the product and the metric a pilot measures it by.

The page also carries a worked example you fill with your own numbers: campaigns, people, hours, hourly rate, media budget. The defaults are assumptions. There are no measured values from production use yet; the first pilot delivers them.

C-level Marketing Agency

With a control group that holds in every channel. batoon keeps the holdout clean in email, ads and the call centre at the same time. The result is the attributed difference per segment against the cost of the deliveries, as net benefit per workflow.

At the variant review you can choose an experiment. There is a winner only when every arm has at least 30 recipients and the difference is significant. All of this has so far been measured in the curated simulated world, not on real revenue figures.

C-level Marketing

A deterministic check before every tool call. It checks budget caps, frequency caps, forbidden claims, consent, suppression lists, risk classes, data purposes and whether the agent stays within its own task. A tool with no connected connector behind it is refused before any other rule.

Every verdict comes with a reason: allow, flag, escalate or deny. A brand judge also checks text against your brand guide and may only tighten a verdict. If the judge fails, the deterministic rules still apply.

Marketing IT & architecture Compliance & privacy

Five situations halt a run. Plan approval, an escalation, a conflict, variant sign-off and a scheduled start. Which role may decide is set in an approval matrix per project, by risk and budget size.

Multi-stage approvals follow the four-eyes principle. No person signs two stages of the same approval. Every halt carries impact, deadline and age, and a list orders the company's open decisions. Who decided is read by the server from the session, never from the browser.

C-level Marketing Compliance & privacy

Everything a run did, appended only and never overwritten. Every action, every gate verdict, every human decision with person and role, every model call with its cost, every knowledge lookup by an agent.

Anything that changes something carries a checked rationale with numbered evidence. An argument without evidence halts the run. A tool result that was truncated for the model is marked as truncated in the trail, so nobody assumes the model saw more.

C-level IT & architecture Compliance & privacy

Contact data stays in the source systems. Scoring, segmentation, per-contact decisions and delivery run there and answer with numbers. Neither the orchestrator nor the language model nor the audit trail ever sees a contact ID.

The same boundary is the scale boundary. In the simulated world 1,000,001 contacts pass through the whole churn-rescue chain in 180 seconds, and every tool result stays under 700 bytes.

IT & architecture Compliance & privacy

Yes, in the simulated world. For every contact of a segment, action, channel, weekday and hour are decided. Whoever has already converted gets nothing. Whoever reacted comes earlier. Doing nothing is a regular outcome.

The churn forecast behind it learns on realised outcomes and competes against a formula. It is chosen only when it is measurably better, and every forecast states its quality. Not built: an offer catalogue and per-contact decisions in ads.

Marketing

Yes. A project declares schedules by interval or as a cron expression with a time zone. The orchestrator creates every due run exactly once, even with two processes running. Missed, skipped and not-started due times are recorded with a reason.

An approved plan can be set to a point in time, up to 90 days ahead. Until then the run waits at a second halt that anyone with approval rights can bring forward or cancel at any time.

Marketing IT & architecture

Replies to campaigns are decided without a run. A rulebook reads the intent and picks one of four reactions: an answer from a template with product truth, a handover to the responsible owner as a call task, a suppression for further mailings, or closing. Every decision carries its latency against the project's time budget.

Reportable matters run separately. Data access, erasure, consent withdrawal, legal threat and security incident go with a deadline to a role you define. Where the duty requires it, the contact leaves marketing at once. Limits: a keyword rulebook rather than a language model, email only, no live provider.

Marketing Compliance & privacy

In four ways. The brand judge checks every text against your brand guide. Approved content modules and mandatory disclosures sit with their history in every agent's task, and the gate checks that they appear in the result. Name, price and attributes enter the text from the product catalogue, and a price in free text is refused.

From one core message batoon derives the text for each channel. Where a channel cannot carry the message, the derivation says why. A text that drops a claim or the call to action escalates to a human.

Marketing Compliance & privacy

That is a runtime setting. Six providers are in the catalogue: Anthropic, OpenAI, Google, Mistral, Moonshot and OpenRouter. You choose per role, separately for planning, chat and execution, and enter your own key in the product. A fallback chain covers outages.

Every task of a plan gets a processing depth by money and impact. Light tasks run on the smaller model, deep ones on the planning model. Documented runs exist on Anthropic only so far.

C-level IT & architecture

Yes, per workflow and channel. The holdout measurement of every completed run feeds a learning ledger. The measured conversion rate prices the next run in the gate. Below an evidence floor of 30 treated and 10 held-back contacts nothing changes.

Limits: only channels with a return channel learn, calls keep their declared rate. Costs are never learned. Nothing learns across workflows. The knowledge base stays read-only for agents.

C-level Marketing

The workspace is a chat. You write a goal, say 500 trial signups in four weeks with a €25,000 cap. The copilot sharpens the goal, proposes a team, adds missing specialists and starts the run. Every step appears as a card in the chat.

Approvals appear as cards with reason, rule and required role. You decide in the chat or in the side panel. The pages Runs, Knowledge, Guardrails, Skills, Schedules, Content modules, Dialog, Paid and Settings add overview and configuration to the chat.

Marketing

The goal becomes a plan with tasks, dependencies and processing depth. You approve the plan, right away or for a point in time. The agents work in parallel waves. Every tool call passes the gate first. Halts go to the responsible role.

At the end stands the result with KPIs, holdout comparison, net benefit, model cost per role and the complete audit trail. Mission Control shows the run live, the outcome page the result, the audit page every decision.

C-level Marketing

Every halt carries impact, deadline and age. A list orders all open halts of the company: overdue first, then by impact, then by waiting time. Added to that are anomalies nobody can decide any more but somebody must see: failed runs, missed due times, overdue reports.

The Runs entry in the navigation carries the number of open decisions. Not built: delivery by mail or Slack.

C-level Marketing

Yes. A deterministic mode demonstrates the same three use cases without an API key, with the same cards, halts and audit trails. It is the default after the first start of the container stack and the basis of every demonstration.

Switching to a real model is an explicit setting. A stored key alone switches nothing.

Marketing IT & architecture Agency

Yes, a REST API with an OpenAPI schema and a reference page inside the product. Every route carries a description, the right it needs, the possible errors and a typed response model. A test fails as soon as schema and application drift apart.

The chat runs over the open AG-UI protocol. You report external events through a dedicated door that can trigger a run. Schedules can be created and removed through the API.

IT & architecture Agency

Your knowledge lives as Markdown in the Open Knowledge Format: strategy, brand voice, decisions, glossary, playbooks, team. You maintain it locally or in a git repository that batoon connects per company. The copilot and every agent in a run can search and read it.

Rules sit in the project's guardrails, in content modules with approval status and in campaign briefs. Recurring procedures are stored as skills following the open Agent Skills specification. None of it knows anything about batoon, and all of it can be taken along.

Marketing IT & architecture Compliance & privacy

The connector catalogue lists 25 entries: ten simulated systems for demo and test, thirteen OAuth vendors from HubSpot, Salesforce, Mailchimp, Meta Ads, Google Ads, GA4, Segment and Braze to Canva, Figma, Adobe Express, Dropbox and WordPress, plus two BigQuery connections.

Five are verified against the real API: Canva, Mailchimp, Meta Ads, HubSpot and Google BigQuery, each read-and-draft only. Nothing is sent. The configuration for WordPress.com is written but not yet verified against the real API. The remaining OAuth entries are catalogue rows with no tools behind them yet.

Marketing IT & architecture

In three stages, by preference. A vendor with its own MCP server is mounted directly. A vendor with only a REST API gets a YAML file for the declarative bridge, with no new code. Either way a third step curates: a whitelist of tools and stable names for policy and audit.

OAuth secrets live in a self-hosted vault. The access token never leaves the bridge process. It reaches neither the model nor a tool's arguments nor the audit trail.

IT & architecture

Today as a container stack per company, via Docker Compose. A provisioning command creates the environment with its own keys and its own compose file. The stack consists of orchestrator, web interface, simulated world and model proxy, optionally an identity broker, vault and bridge.

Where the stack runs for a pilot is something we settle with the design partner. A shared hosted service does not exist yet.

C-level IT & architecture Compliance & privacy

batoon is an OIDC relying party. It stores no passwords and runs no user directory. An identity provider is authoritative only for the companies that name it themselves. Nobody enters a foreign company through a foreign provider.

There are three rights: read, approve, configure. They follow from company and project role. The browser round trip with PKCE runs against a Keycloak broker with a federated customer directory. It has never run against the provider of a real corporation.

IT & architecture Compliance & privacy

Customer data stays in your source systems. Per company, batoon holds the configuration, the knowledge base, audit trails, chat histories and operational data such as the learning ledger, schedules and reports. Contact IDs never reach batoon.

The language model sees the goal, the plan, knowledge excerpts, briefs and truncated tool results. Which provider that is, you decide with your own key. The data protection assessment for that provider remains your task.

IT & architecture Compliance & privacy

The path is defined: questionnaire, company pack, repository as a mandatory step, connectors, a first run in deterministic mode, then with a model. A reliable duration exists only after the first pilot. So far no company has walked the path outside our own tests.

C-level IT & architecture Agency

A licence per company, tiered by usage: projects and teams, connected systems, role coverage and compliance features. No per-seat pricing. A usage component with an included allowance comes on top. Model costs pass through with your own key.

We validate price anchors with the first design partners. There are no public prices yet for that reason.

C-level Agency

With traceability and enforcement. Every decision sits in the audit trail with reason and evidence. Consent applies per channel, data purposes are declared and checked at the gate, and a new purpose is a halt with high impact. Reportable matters go with a deadline to a role.

batoon is not legal advice and carries no certification. What it delivers is the record an audit needs.

C-level Compliance & privacy

An agent reaches only tools with a connected connector behind them, and only objects within its own task. Forbidden claims, missing consent and suppression lists lead to denial. High risk and budget overruns lead to a halt.

We name two limits. The brand judge is advisory and can, at worst, let a deliberately crafted text pass, in which case the deterministic rules still apply. And whoever changes guardrails does so today with one signature, without the four-eyes principle.

C-level IT & architecture Compliance & privacy

A fallback chain switches to the next model, and finally to the same model over a second billing route. Context budget, compaction and caching fall back to the previous behaviour on any error. A run does not break because of it.

The chain is tested. Its third stage depends on a valid key, which was missing at last check. It does not bridge an outage of the whole provider.

IT & architecture

batoon measures itself against its own catalogue of 96 enterprise capabilities, eight layers and 13 industry scenarios. As of 5 September 2026: 66.8 % capability coverage, all 49 must capabilities at maturity 3 or higher, 13 of 13 scenarios feasible by that catalogue. On 25 August it was 51.2 % and 2 of 13.

Built are three use cases end to end in a curated simulated world, five real systems read-and-draft only, and multi-tenancy with login. The full status is on the status page.

C-level Marketing IT & architecture Compliance & privacy Agency

No. There are no paying customers and no results from production use. We are looking for two to three design partners for paid pilots with a say in the product, ideally from regulated industries.

C-level Marketing IT & architecture Compliance & privacy Agency

No live execution. Nothing is sent, no campaign launches. No real customer data. No enterprise IdP in use. Documented runs on one model provider only. We set the assessment ourselves.

24 of the 96 capabilities sit below maturity 3, all in the should and could classes. For every built capability the product repository states what is not built.

C-level Marketing IT & architecture Compliance & privacy Agency

Write to florian@batoon.ai. Name your industry, the systems in your stack and when you want to introduce agentic marketing. A pilot runs on your real stack, with acceptance against the same catalogue IDs batoon measures itself against.

C-level Marketing Agency