AI Agents
Use AI for your business.
I build AI agents for marketing, sales and internal teams as a temporary team member: in your systems, aligned with IT and the AI team, approved by your people.
Felix Zeeb, Zurich. Since 2019 I have been building marketing and sales systems from SMEs to large groups, today mostly AI agents. In my own companies I sell myself, for example at NativeAI, with a B2B CRM I built. About me
Companies I have had the privilege to work for since 2019: Swisscom · NZZ · Allianz · Suva · PwC · CH Media · Groupe Mutuel · Mobility and many more. Reference calls on request.
Anatomy of an agent service
The same architecture behind the sales agent, the multichannel agent and the Business Context Layer.
Plan your project and experience how I work.
What can go wrong and what prevents it
✓ Guardrails in code and tests before every release catch errors, and a person approves before anything goes out.
✓ Models in Switzerland or the EU without training on your data, self-hosted for sensitive data. Personal data is masked.
✓ Code, runbook, tests and infrastructure stay with you. Your team is prepared from day one should you want to end the collaboration, and should I suddenly drop out, a partner I have worked with for years steps in.
✓ Sign-off by IT. Architecture and security agreed up front. Service account with minimal rights instead of full access.
✓ Swiss data protection act, FINMA, professional and insurance secrecy built in from the start; approvals in several stages.
✓ You use the model of your choice and can switch it at any time; no lock-in, code and infrastructure stay with you.
✓ Usage and approval rate measured from day one; the team shapes it in the review rounds.
Who is responsible for what?
During the build I work in your repository, tracing and eval stack, and where they do not exist yet, I set them up for you. As long as I run operations, I deliver releases with a report. After that:
| Task | Business team | MarTech or AI team | IT | Me, until handover |
|---|---|---|---|---|
| Daily operations: review and approve drafts | ● | |||
| Flag errors, maintain knowledge and examples | ● | |||
| Test cases and evaluation | ● | ● | ||
| Releases after a passed regression test | ● | |||
| Code review of every pull request | ● | ● | ||
| Deployment, keys and access, monitoring | ● |
responsible on request
Repository, runbook and infrastructure as code are with you from day one.
Example: review of a planned agentic AI implementation with the AI team of a telecoms group.
The learning cycle
Generate, evaluate, optimise, also after go-live: live data shows cases nobody thought of. On the right, the report before every release.
- ConversationThe agent works in daily operations.
- TracingEvery case is logged.
- Automatic evaluationA checker model rates every case.
- Review by the business teamThe business team reviews without writing prompts.
- Test caseNotable cases become test cases.
- OptimizationExamples and rules are improved.
- Regression testWith synthetic users against all cases.
- ReleaseAnd round again.
Drafts sent without changes. The share of drafts the team approves without correction. When it rises, the team saves time and the agent gets content and tone right. The other values must not drop meanwhile.
One system, three layers
Data, CRM and AI belong together and depend on each other: one stack, not a menu of offers.
The context layer as the foundation
Everything that defines your company, in one place: data warehouse and events, ERP, website, shop and app, brand guidelines, template and content libraries, product knowledge and the definitions of your audiences. Without this context, agents guess.
CRM as the engine
Salesforce Marketing Cloud, HubSpot, MoEngage, Dynamics. Lifecycle, journeys, triggers.
Agentic AI as the accelerator
Request agents, campaign agents, assessments. In your channels, with a human gate.
Building blocks for agentic systems, open source
I developed these Claude skills myself and make them available to everyone who works with agentic systems. Anyone who wants to develop them further with me is warmly invited.
- google-ads-audit ↗Audits a Google Ads account like a PPC pro: findings, action plan, experiment backlog. Method from real audits.
- ga4-growth-strategy ↗Cleans GA4 data, reconciles it with real orders and derives a prioritised marketing and sales strategy.
Try it yourself
The AI Project Planner is an agentic flow itself: one agent asks, a second writes the plan, a third checks it. You get a plan including an effort estimate for internal use and see along the way how I build such systems.
CONTACT
Let's talk about your project.
Tell me about your idea: on a video call or over lunch, in Zurich and across German-speaking Switzerland. I'll tell you straight away how I would build it.
I reply within one working day with a proposed time.
Prefer email? felix.zeeb@geistzeit.ch