Our AI Manifesto
IBM famously said that as a computer can never be held accountable, it can never make a management decision. This is at the core of the Type A ethos when building an agentic digital marketing agency.
3 Power Laws
Traditionally, agencies have always used technology to automate things. From pulling report data into dashboards, to setting up crawl alerts to building advanced Zapier automation workflows, agencies have always automated scheduled, repeatable and deterministic work.
With the advent of AI we are now able to supercharge these automations so we can provide better service to our clients and drive results much faster.
When understanding AI usage at Type A do 3 things:
- split tasks into HIL and AFK
- break deliverables down to their smallest primitives
- view all work as a database that can have several outputs
HIL vs AFK
HIL stands for human in the loop. AFK mean away from keyboard. Using these two concepts you are able to start building self learning agentic systems that have human creativity and judgement at every step.
Deliverables as primitives
A deliverable like ‘keyword research’ contains smaller tasks like ‘build a keyword universe’ and even smaller actions like ‘map keywords to urls’. The smallest output of our deliverables are called primitives. Some primitives are reused across deliverables. For example the primitive of ‘chunking URLs according to intent’ is used across tech audits, onpage audits, reporting, keyword research, etc.
Views of a database
If we consider all digital marketing work, it’s really just lots of databases that are being viewed with different filters to create different outputs. A site crawl is a database, GA and GSC data are databases, content on the clients site is a database, competitor data is a database. Deliverables are just views of these databases combined.
These databases, filtered and combined to create a different output.
Frameworks bring it together
Type A have frameworks for completing our work. It allows us to plan more effectively, bug check quickly and deliver consistent high quality, predictable output for our clients. The delivery is bespoke but the methodology is standardised. Think of it the same way you build a house - the order of foundations, insulation, electric and plumbing doesn’t change from house to house but the house size, layout and decoration are totally bespoke.
Guardrails
HIL is built into every primitive. This means every AI artefact that is created has felt a human hand. As the work is chained together the output from a HIL task A is used in the AFK task B and then HIL task B is used in AFK task C. This recursive loop means that the AI learns and gets better and is constantly fed better and better data to provide better and better results.
Our frameworks have guardrails by design. Firstly, as the primitives are so small, the blast radius for probabilistic LLM issues like hallucinations are inconsequential and close to zero. Secondly, as each step is pre-mapped and built on human input the AI never makes a decision, it only builds in a tiny sandbox, supervised by a Type A specialist.
Primitive Checks
Our HIL primitive checks have the following stopping points, once the HIL is complete the following things fire:
AI diff check
To confirm human changes were actually made to the primitive.
Spec check
To confirm the changes makes sense against the primitives schema and specifications.
Weekly why?
A slackbot the collates all changes made and asks the human why they were made so it can be fed back into the client profile.md which is the master rule book for producing client work.
Client Verify
The AI calls the client on the phone each month to confirm our changes are correct and aligned with their business.
Persistent Memory and Client Control
AI has access to basic databases like search console and GA which are great at saying “what happened” but they struggle to say ‘why’ something happened and they are completely unable to give any context about the clients business.
To solve this we have 2 self learning files that auto-update with the freshest information about the clients business:
The clients business
Context on the clients overall business, their goals, tone of voice, brand and everything that makes them, them.
The granular detail
Novel information about more granular topics (the story of the clients business, products and services).
The client as the human in the loop
To keep this information fresh, up to date and factually correct we have human maintenance loops that route questions and answers to the people in the clients business that are best placed to answer them.
AI calls specific client contacts if there has been any profile changes to get confirmation that they are correct.
AI calls all client contacts gather feedback and log as issues on github.
When the team do not have enough context to complete a task or need clarification on clients in-house style, product or service they can ask a slackbot to ping them client for the information.
Across All Surfaces
The magic to making everything work in harmony is having data collection available across all surfaces, diff checked, synthesised and compared to our knowledge base on a regular basis. Every entry point for feedback is collected to feed our self learning system from:
With this level of data ingestion, we are always up to date with the clients industry and a step ahead of the competition.
Databases
The databases powering all of this work are the most important parts of the chain. If the data is wrong then it can have a serious knock on effect.
Our approach to have a Directed A-cyclical Graph or DAG across all of our data. This ends up in a medalian bronze, silver, gold database structure that looks something like:
Why this approach wins
Most agencies will build individual tools to do individual tasks; a content briefing tool for this sort of content, a meta data re-writer for that process. Which has a few major issues:
they don’t share context
they don’t self learn
the client is never in the delivery loop
Type A’s approach to building a fully agentic agency utilises a monorepo that has full, deep context of everything in the clients world with AI that works to strict guardrails, non-intrusive ways for the AI to talk directly to the client all underpinned by smart frameworks with a human hand build in by default.
Experience our agentic capabilities today.
Get in touch for a free AI visibility audit or request a proposal.
Request a proposal