How AfM's Marketing System Evolved from 2023 to 2026

Why tell this story

Articles about AI marketing systems tend to describe an end state: agents that run everything, humans who only approve. Ours did not arrive there, and we do not claim it has. This is how AI for Marketing's system actually developed, drawn from our published case records and the documents we run the business by. Where a stage is not documented, we leave it out.

One fact frames the whole story. AI for Marketing Ltd was incorporated in August 2025. The client work from 2023 to mid 2025 described below (ToastPal, Mumshandmade and the global technology platform) is work Jakub Cambor delivered as a freelance consultant before founding AfM. It is where the method came from, and it is labelled that way.

The timeline

Stages in AfM's system, from published case records and internal documents

Period	What happened	What it taught us
October 2023 to October 2024	Twelve months of AI-assisted paid search for ToastPal across the US, UK, Australia and Canada (freelance, before AfM)	Structured testing matters more than clever tools; the case record lists the absence of a testing framework as a starting problem
September 2024 to January 2025	Meta and Google advertising built from zero for Mumshandmade, a premium knitwear brand (freelance, before AfM)	Tracking, audiences and creative have to be built before optimisation means anything
October 2024 to March 2025	A content experiment with no human review step on Harmonance, the founder's own project	Unreviewed AI content can earn search traffic, and it is a risk we accept on our own brand, not a client's
November 2024 to September 2025	About 11 months of paid media for an anonymised global technology platform, plus inbound lead automation (freelance, almost entirely before AfM)	Quarterly strategy reviews matter, and so does the handoff after the click: more leads created a processing bottleneck
2026	One-off no-code agent builds: Allegiance Industries (running daily) and HEAVENSAKE (delivered in full, not deployed by the client)	Simple inputs, explicit exception paths, outputs designed as drafts for the client to approve; delivered is not the same as deployed
2026	Delivery moved onto AfM's own internal operating system, run as one managed service	The system holds context and prepares work; a person decides, checks and approves

Stage one: service work with AI assistance

The earliest documented work is paid media Jakub ran as a freelance consultant. The ToastPal record runs from October 2023 to October 2024, and the Mumshandmade record from September 2024 to January 2025. The ToastPal record describes AI assistance for ad copy volume and data analysis, but the work was client service in the ordinary sense: a person running campaigns, testing and reporting.

The lasting lesson was about method rather than technology. The ToastPal record names "no systematic testing framework" as one of the problems at the start. Structured testing, not the choice of AI tool, is what made improvement repeatable. We have written that up separately as scientific AI marketing testing.

Stage two: automation where the work was repetitive

During the paid media engagement for an anonymised global technology platform (November 2024 to September 2025, about 11 months), the case record reports cost per lead falling from $270+ in the engagement's own first quarter, on LinkedIn only, to $8.86 blended across Meta and LinkedIn in the final quarter. Most of that fall came from moving the main budget to Meta: LinkedIn on its own was still $201 a lead at the end. Success created a new bottleneck: leads arrived faster than the team could research and file them. The answer was an inbound processing workflow that enriches each lead and files it in the CRM, with low-confidence records flagged for a person.

The 2026 builds for Allegiance Industries and HEAVENSAKE, AfM's first as a company, followed the same shape: a simple input, an AI agent inside a no-code orchestration layer, and an explicit route for the cases the agent could not handle. The HEAVENSAKE agent was designed so that every output is a draft for the brand team to approve. It was delivered in full, but HEAVENSAKE decided not to deploy it into their live content workflow, so it has no traffic results. The lesson is plain: a build the client does not adopt produces nothing, and a service in which AfM runs the work does not depend on the client adopting software.

Stage three: testing autonomy on our own brand

Harmonance is Jakub's own project, so it is not a client result. Between October 2024 and March 2025 he ran a content system on it with no human review step at all, deliberately, to see what unreviewed AI content could do. The case record reports 10,976 organic clicks and 770,576 impressions over 152 days.

The experiment showed that the approach can earn search visibility. It also clarified a line we still hold: running without review is a risk an owner can choose for their own brand. It is not a risk we take with a client's reputation.

Stage four: from tools to an operating system

The founder's documented path runs from no-code tools such as Make.com and Relevance AI to repository-backed systems with a database, server-hosted workflows and an operator channel. The move was driven by a practical problem: context scattered across tabs, chats and tools cannot be searched, checked or reused.

The system today holds client context, records meetings, prepares drafts, analysis and reports, and tracks what was approved. The documents that govern it are explicit about limits:

• • every client-facing artefact goes to Jakub first, and nothing is sent externally without an exact approval;
• • every AI-generated reply to a prospect or client stays an unsent draft until he edits and sends it himself;
• • outbound sending is held behind a daily owner approval that the sending scripts check before they run.

Honesty about progress became a rule too. An internal review in July 2026 found that progress tracked on paper had run ahead of what was working in practice. The response was to judge the system by what it demonstrably does, not by what its checklists say.

Stage five: one managed service

The offer went through several names before it settled. What it is now is simple: AfM runs a client's marketing, using its internal AI operating system and the founder's expertise. The system retains context and supports production; a person supplies judgement, prioritisation and review. Clients buy managed work and measurable improvement. They are not buying code, and they do not have to operate software.

What is deliberately not automated

• • Decisions about what to prioritise for a client.
• • Approval of anything that leaves the business: emails, ads, published pages, reports.
• • Replies to people.
• • Changes to spend.

Some of this may change as specific workflows earn trust. Some of it should not change at all.

Lessons you can use

1. • Start with service and method; automate the repetitive steps once they are understood.
2. • Design the exception path before the happy path.
3. • Test autonomy where you own the risk.
4. • Judge the system by its outputs, not its checklists.
5. • Keep approval wherever money, reputation or a relationship is at stake.

For how the service runs day to day, see how it works; the full case records are on the case studies page.

FAQ

Is AfM's system autonomous?

No. It prepares work, holds context and keeps records, but a person decides priorities and approves everything that goes to a client or the outside world. The one unreviewed experiment described here ran on a founder-owned brand.

Can clients buy the system itself?

The service described here is managed: clients buy the work and the improvement, not the software. Any other arrangement would be a separate conversation about scope.

Why use no-code tools for some client builds?

For narrow workflows such as prospect research, no-code orchestration with an AI agent is quick to build and easy for the client's team to understand. The approach is chosen per workflow, not as a rule.

Why mention an internal review that found problems?

Because a system that only reports its successes cannot be trusted to report a client's results either. Stating where progress was overstated is part of the standard we apply to our own work.