Skip to content

Security & governance pack · for IT, data protection & legal

What IT wants to know before agents get write access.

Agents that write into CRM, ads and website belong in every proper security review. This pack answers the questions upfront: permissions, data flows, GDPR, prompt injection and the EU AI Act, plus a pre-filled questionnaire. No form required, and it's made to be forwarded.

  • No form
  • PDF · DE & EN
  • As of October 2026
Read the video as text

The video has no speech, only music. Scenes and on-screen text:

  1. 0:00Logo: security & governance · RevPrime®.
  2. 0:04“Agents in your accounts?” Marketing wants agents (CRM, ads, website), agents are meant to write (in live systems), the IT review blocks everything (questions without answers). Before an agent may write, IT wants to know four things: What may it do? Where is the data? Who approves? How do I stop it?
  3. 0:128 agents, 10+ tools with access, 1 security review: “Without rules, IT rightly says no.”
  4. 0:18“No agent changes anything without a go. Every level documented, revocable at any time.” The areas with their highest level: CRM L3, LinkedIn & Google Ads L2, GA4 read only (L0), website & CMS L2, Slack & inboxes L2, knowledge base L2, enrichment L2, the model runs via your account. Everything runs in your accounts, under your contracts.
  5. 0:26“One change, one path”, shown on a proposal by the paid media agent (demo data): read (L0), analyze (creative B3: cost per lead +38% in 7 days), propose (L1: shift budget from B3 to B1), effect & risk (about −18% cost per lead, low risk, rollback possible, checkpoint after 7 days), your go, write (L2) and change log with before/after. This agent's permissions: read, propose, write after go for budgets and bids.
  6. 0:42“One change log, complete.” Entries from 06:00 to 07:10: pipeline data read (L0), proposal with effect and risk (L1), CRM note under an approved rule (L3), a finding waits for a go, an email saying “ignore all rules” is treated as data and not executed. The proposal waits for your go and gets approved.
  7. 0:50“Permissions, step by step.” L0 read, L1 propose, L2 write after go, L3 approved rule. Outreach, publishing and budgets always need a single go.
  8. 0:58“How the security review runs.” Before the start the security pack goes to IT; in setup permissions and narrowly scoped access per agent and tool; in operation stoppable any time; at project end a documented handover, access removed, tokens rotated, no copies at igromo.
  9. 1:06“Reviewable for IT and data protection.” L0–L3 per agent and tool, a change log for every change, no shadow copies, a pre-filled security questionnaire as a PDF. Applies to every agent: external effect only after a go, change log, data isn't instructions, stoppable any time.
  10. 1:16“Forward the security pack to your IT.” No form, PDF in German and English.

At a glance

Six answers for the busy reviewer.

  • Runs in your accounts

    Tools, knowledge base, hub and agent setup run in your environment. Models, enrichment and tools run on your own contracts.

  • Reading isn't writing

    Agents read for the analysis. Write access is granted individually per tool and agent.

  • No go, no external effect

    Outreach, publishing and budgets need a single go. Irreversible actions always do.

  • Complete change log

    Every change with before/after, reason, approval and timestamp. Rollback where technically possible.

  • Data isn't instructions

    Agents treat content from emails, chats and web pages as information. Rules only come from the knowledge base.

  • Stoppable at any time

    Revoke an approved rule, pause an agent, remove access: the agent stops immediately.

You're in marketing and want to introduce agents?

Forward this page to IT and data protection before the topic hits the review. It usually saves the first round of questions.

Forward by email

01 · Permission ladder

Permissions are granted step by step.

Every agent starts at the bottom and climbs only as far as its job requires. The level is set per tool, documented and can be revoked at any time.

  1. L0

    Read

    Agents read CRM, ads, web and docs for the analysis.

    Reading pipeline data for the weekly report

  2. L1

    Propose

    Numbered proposals with effect, risk and checkpoint. Nothing changes in the tool.

    “Pause creative B3, shift budget to B1”

  3. L2

    Write after go

    After a single go the agent writes the change straight into the tool and logs it.

    Publish a landing page, start a sequence

  4. L3

    Approved rule

    Narrowly scoped routines run on a rule approved once, which can be revoked at any time.

    Write CRM notes, maintain audiences

Always a single go, even with an approved rule

  • Messages to external people (outreach, sequences, emails)
  • Publishing (landing pages, posts, ads going live)
  • Budget and bid changes
  • Anything that can't be rolled back

Typical starting configuration of the core agents

AgentReadsWritesMax.
Signal AgentCRM, web, intentScores and notes in the CRM (rule), alert to salesL3
Angle & Content AgentKnowledge base, existing assetsDrafts; publishing only after goL2
Paid Media AgentLinkedIn Ads, Google Ads, GA4Optimizations, budgets, bids (per go)L2
Pipeline Reporting AgentCRM, ads, analyticsReports and dashboards in the hubL1
Outreach Routing AgentCRM, outreach toolsRouting, contact hygiene (rule); sequences after goL3
Inbox & Ops AgentApproved inboxes and channelsTasks with an owner after goL2
Product Truth AgentProduct docs, release notes, websiteFlags on contradictionsL1
QA & Brand Guard AgentLive pages, ads, emailsFindings; fixes after goL2

Guide values. Which level an agent actually gets per tool is set and documented in the project. Many teams deliberately start with L0 and L1 only.

02 · Data flows

Where data comes from and where it goes.

An agent run always follows the same path. The concrete data-flow document with your tools, fields and access is created in the project and handed over.

  1. 1

    Your tools

    Read through dedicated, narrowly scoped access to CRM, ads, web and inbox

  2. 2

    Agent run

    Model via your account, with the context this run needs

  3. 3

    Knowledge base

    Rules, angles and decisions as files in your repo or workspace

  4. 4

    Proposal in the hub

    What, why, effect, risk (waiting for a go)

  5. 5

    Write & log

    Change in the tool, before/after in the change log

AreaExamplesAgents readAgents writePersonal data
CRMHubSpot, SalesforceCompanies, contacts, deals, stages, activitiesNotes, scores, properties, workflows, deal stagesYes (business contact data)
AdsLinkedIn Ads, Google AdsCampaigns, cost, performanceAudiences, creatives, bids, budgets (budget always with a go)Low (audience lists)
Web & analyticsGA4, Search Console, FibblerPages, traffic, search queries, company-level visits– (read only)Low (aggregated)
Enrichment & outreachClay, Sales Navigator, HeyReach, InstantlyCompany and role data, sequence statusEnrich lists; sequences only after goYes (business contact data)
Website & CMSWebflow, Vercel, Lovable, HubSpot CMSPages and contentLanding pages and content, only after goNo
Communication & workSlack, Teams, Outlook, Gmail, Notion, ClickUpApproved channels, inboxes, tasksTasks, summaries, draftsYes (message content)
ModelClaude (default), others after testingThe context of the given run–Depends on the run (via your account)
Knowledge base & hubYour repo or workspaceRules, angles, decisions, metricsUpdates after the owner's goMinimal (aggregated)

A selection of typical tools. Which tools, fields and inboxes an agent actually sees is set by you in the project. New tools are only added when the revenue logic calls for it.

03 · GDPR & data processing

Data stays where it is. Under your contracts.

Because RevPrime® runs in your accounts, your data map changes less than you'd think: the data lives in the tools you've already reviewed.

  • Controller

    You

    You decide on the purposes and means of processing and on every approval in the system.

  • Processor

    igromo

    Where igromo has access to personal data while building and running the system, a data processing agreement under Art. 28 GDPR governs it.

  • Your processors

    Model and tool providers

    Anthropic, HubSpot, LinkedIn & co. run on your accounts and your contracts, just as they do today.

Data minimization
Agents only get access to the fields, channels and inboxes their job requires. The hub stores aggregated metrics and status instead of personal data.
No special categories
RevPrime® processes business contact data (name, role, company, business email, activities). Special categories under Art. 9 GDPR are not part of the setup.
No automated individual decisions
Scores and routing prepare decisions; the system makes no decisions with legal or similarly significant effect within the meaning of Art. 22 GDPR.
Legal bases
Account research and B2B outreach often rely on legitimate interest. The balancing test is yours. Your rules (e.g. on email marketing) live in the brand, claims & legal module of the knowledge base, and every agent follows them.
Third countries
Many tools in the stack are US providers. Transfers run on your contracts with those providers, e.g. based on the EU-US Data Privacy Framework or standard contractual clauses. Where providers offer EU data regions, they can be used.
Documentation for DPIA and records
Whether a data protection impact assessment is needed is your decision. I provide the system description, data flows and measures as input for the DPIA and your records of processing.
Data subject rights
Because the data lives in your systems, your existing processes for access, rectification and erasure apply. Agents don't create shadow copies.
End of project
I hand the system over with documentation. My access is removed and tokens are rotated. I keep no copies of your customer data.

04 · Prompt injection & agent security

Security rests on the architecture, not on the model.

No model today is fully immune to prompt injection. RevPrime® therefore limits what a deceived agent can do in the first place and doesn't rely on the model spotting a manipulation.

Hidden instructions in emails, web pages or forms

Data isn't instructions: external content is treated as information and never becomes a rule unchecked. Rules only come from the knowledge base, with the owner's go.

Data exfiltration through outbound actions

No external effect without a single go. Recipient and content are part of the proposal. Write access only to the tools the agent needs.

Overly broad permissions

Least privilege per tool and agent along the permission ladder. Start read-only, more rights only once the agent has proven itself.

Wrong or outdated claims go live

Approved claims in the brand, claims & legal module. The QA & brand guard agent checks everything before and after go-live, the product truth agent matches against product docs.

Error cascades and runaway cost

Budget and bid changes always need a go. Routines run with thresholds from the approved rule and can be paused at any time.

Compromised credentials

Separate, narrowly scoped access per tool. No secrets in the knowledge base or prompts. Revoking or rotating access stops the agent immediately.

That said, an agent can make mistakes even within approved routines, such as a wrong CRM note. That's why these routines are narrowly scoped, every change is in the change log, and anything with external effect stays with the single go.

05 · EU AI Act

Where agentic marketing sits in the AI Act.

An assessment from an architecture point of view, not legal advice. It helps your legal team get to the right questions faster.

  1. Aug 2024

    Entered into force

  2. Feb 2025

    Prohibited practices and the AI literacy duty (Art. 4) apply

  3. Aug 2025

    Obligations for providers of general-purpose AI models apply

  4. from Aug 2026

    Further obligations, incl. transparency (Art. 50), apply step by step

For some obligations the EU has initiated adjustments and deadline shifts through the digital omnibus package. Your legal counsel checks the current status.

  • Risk class

    Typical RevPrime® use cases (content, reporting, campaign monitoring, account research and company-level scoring) do not, in my assessment, fall under the high-risk areas of Annex III, such as employment, creditworthiness of natural persons or critical infrastructure. Use cases in those areas are not part of RevPrime®.

  • Roles

    You use the system in your company and are therefore usually the deployer. The model provider (Anthropic by default) is the provider of the general-purpose model. The allocation of roles is documented in the project.

  • Transparency

    Chatbots that talk directly to people must disclose that they are AI; synthetic images, audio and video (deepfakes in particular) must be labeled. With RevPrime®, messages to external people only go out after a single go. Your labeling rules live in the brand, claims & legal module and apply to every agent.

  • AI literacy

    Art. 4 requires people working with AI systems to be sufficiently trained. Onboarding, playbooks and the documented handover are part of every project and contribute to this.

  • Oversight and traceability

    Human approval, a change log with before/after and clear owners are formally required mainly for high-risk systems. RevPrime® has them built in anyway, because otherwise leadership won't trust the system.

This pack describes the architecture and way of working of RevPrime®. It is not legal advice and does not replace a review by your data protection and legal counsel. The proposal and the data processing agreement are binding.

06 · Security questionnaire

The usual questions. Already answered.

These questions come up in almost every vendor and security review. The answers hold for every RevPrime® project; anything that depends on the specific setup is marked as such.

You have your own questionnaire?

Send it over. I'll fill it in, usually before the first meeting with IT or data protection.

Send questionnaire

01 · Company & accountability

Who is igromo?
igromo e.U., based in Heiligeneich, Lower Austria (EU), is the founder-led studio of Mario Maier: management consultancy and IT services from one source. Contact: hello@igromo.com.
Who has access to our systems?
Me, Mario Maier, personally. Every engagement runs founder-led and senior-only. Other people only get access with your prior consent and under the data processing agreement.
Does igromo hold certifications like ISO 27001 or SOC 2?
No. I run no platform of my own that stores your customer data. The data stays in your tools, which are covered by their own certifications and your contracts. My measures are described in this pack, and the data processing agreement sets out the details.

02 · Access & identities

How do the agents access our tools?
Through the tools' official interfaces (APIs, connectors, MCP) with dedicated, narrowly scoped access per tool. Shared personal logins are not used. The rights an agent gets follow the permission ladder and are documented.
Who grants and revokes permissions?
You do. Access is created in your tools, and you can revoke it at any time. Without access an agent can neither read nor write.
Where are credentials and tokens stored?
In the secret management of the respective runtime, not in the knowledge base, in prompts or in the repo. Exactly where is set and documented per project.
How is igromo's own access secured?
I use personal accounts in your tools with multi-factor authentication wherever the tool supports it. I adopt your requirements, e.g. SSO or access only through specific accounts.

03 · Data & storage

Where is our data stored?
In your tools and accounts. The knowledge base lives as versioned files (Markdown, YAML) in your repo or workspace; the hub stores aggregated metrics and status. I don't copy customer data into igromo's systems.
Which personal data do the agents process?
Business contact data from CRM and outreach tools (name, role, company, business email, activities) and content from inboxes and channels you approve. No special categories under Art. 9 GDPR.
Is our data used to train models?
The agents run on your own model account under your contract terms. Under commercial API and business contracts of the major providers, inputs are not used for training by default. Check this in your contract with the provider.
How long is data retained?
The data stays in your systems and follows your deletion policies. How long the change log is kept is set by you in the project.

04 · AI & agents

Which model do the agents run on?
Claude by Anthropic by default. The system is designed to be model-agnostic; other models and agent harnesses are possible. Every combination is tested and signed off for the specific deployment, so switching isn't a single click.
Can agents act without approval?
Only within routines you have explicitly approved once, such as audience upkeep or CRM notes. Anything with external effect (outreach, publishing, budgets) and anything irreversible needs a single go.
What happens when an agent is wrong?
You reject the proposal, and the rejection with its reason goes into the decision memory. Executed changes are in the change log with before/after and can be rolled back where technically possible.
Can we stop an agent immediately?
Yes. You can revoke any approved rule, pause any agent and remove any access. After that the agent no longer acts.

05 · Operations & incidents

Is every action traceable?
Yes. The change log records every change with agent, before/after, reason, approval and timestamp. Agent status, pending approvals and anomalies are visible in the hub.
What happens in a security incident?
I inform you without undue delay, lock the affected access and analyze the incident together with you. Deadlines and contact channels are set out in the data processing agreement.

06 · Contract & exit

Is there a data processing agreement?
Yes, as soon as I have access to personal data. You can use your own template.
Who owns the system?
You do. Knowledge base, prompts, playbooks, hub and integrations run in your environment and are handed over with documentation. Usage rights are set out in the proposal.
How big is the lock-in risk?
Low. The knowledge base consists of open files, the system is designed to be model-agnostic and runs in your accounts. Your team can keep running it after the handover.

To pass on

All of it on a few pages, as a PDF.

The same content as a print-friendly PDF that you can attach to the review request, file in vendor management or print for the meeting.

This pack describes the architecture and way of working of RevPrime®. It is not legal advice and does not replace a review by your data protection and legal counsel. The proposal and the data processing agreement are binding.

Next step

Ready for marketing that runs itself?

A 30-minute diagnosis call: I look at your setup and show you where pipeline leaks and which agents would have the biggest leverage first. You get that even if RevPrime® isn't a fit.

Free · directly with Mario · reply within 24h

Rather send a quick message? Message me on WhatsApp