The Context Imperative

AI has a credibility problem. For all the widespread usage, few people trust AI generated content to be faithful to their original query. Even fewer consider it to be relevant to business needs. As a result, extensive QA processes have been devised to hide the origin of AI content. Yet, review processes are merely band aid solutions. Highly tuned prompts help mitigate any inaccuracies within AI output, but they can’t provide the human experience needed to show insight and prove value. For AI to be truly useful, the challenges facing AI output need to be fixed at source.

Enter Context

In the lab, AI models are trained on the entire breadth of human knowledge. That is incredibly useful for research and discovery, but becomes a problem when you need a specialised agent to guide decision making. Many agentic AI projects fail because the underlying AI algorithm lacks the detailed experience and institutional know-how needed to support the process being automated. Human knowledge workers need extensive training before they can reliably manage a complex business process. Agents are no different.

Over the last twelve months, context has become a leading tech industry buzzword. Model makers promote it as the answer to the much publicised accuracy and adoption challenges confronting them. Power users see it as the solution to slop. Context provides the detailed institutional background needed for AI to work effectively. It adds the personalised intelligence and industry insight otherwise lacking in poor quality AI generated content. Without context, AI doesn’t know what your business does or who you’re talking to. It can’t align output to your business strategy or to your specific strengths and weaknesses. That results in generic content to match the generic web information used to train the underlying large language model (LLM). 

Retrieval Augmented Generation

Turning AI into a useful decision-making engine requires feeding it with all the business-specific intelligence not shared publicly. Much of that information will already be published internally though. It will be collected and organised for human consumption in SaaS platforms such as Salesforce or SharePoint. It just needs to be made accessible to AI. This should be seen as a content and data engineering challenge rather than an integration problem. The underlying integration work will already have been carried out by the platform developers, but most AI models also require content owners to update the content itself. As such, data scientists have developed a set of best practice approaches for sharing in-house datasets and private information with LLMs, which are collectively known as Retrieval Augmented Generation (RAG).

At its most basic, RAG is about organising internal business documents so AI models can use them to support business queries. A business analyst might rewrite internal strategy documents to follow a detailed FAQ format, so that they are easily searchable by AI. A revenue operations team might open up the corporate CRM system for AI indexing using an MCP integration. A data engineer might add contextual information to corporate dashboards so that AI can interpret and monitor KPIs. RAG is ultimately about tagging content so that it is easier for AI to categorise, as well as about adding missing business context to reduce hallucinations. Often this requires breaking out long documents into an FAQ format; a strategy that content teams are also using to improve website rankings within ChatGPT. 

Small Language Models

RAG makes AI models easier to customise. For all the success of Claude and ChatGPT, many IT teams want to replace them with internal models. That's often driven by security requirements, but internal models are also more responsive to business needs. App developers have very little control over the content returned by a ChatGPT API call. A bespoke small language model can be tweaked for the required use case simply by updating the relevant training material. Increasingly, automated feedback loops form part of the training process. Allowing models to learn from user feedback and successful agent executions with minimal developer intervention.

AI projects often fail despite initial good results. Getting AI to automate a task once is relatively easy. Getting AI to successfully automate the same task every time is hard. All too often, an agent can complete 90% of a task 90% of the time. Teams struggle to close that final 10% because the AI hasn't been trained in sufficient depth on the relevant process. A custom model can help close that gap because it removes irrelevant knowledge causing hallucinations, and replaces it with the detailed business context required to complete the task.

Much of the time, you don't even need a bespoke model. The leading public models work as well, if provided with all the relevant background information in the system prompt. That does require careful prompt tweaking, as well as adding links to plenty of secondary documentation. Training a model is a long and complex task, regardless of model size. Collating the information needed to automate a process with AI takes far longer than actually developing the relevant automation workflow. However, it is essential if AI is to become the foundation of enterprise automation.

Marketo Adds Picklist Management

Marketo has always had a strange relationship with picklists. The system will give you a list of options to choose from when configuring a smart list. However, admins have no way of specifying the options displayed in the smart list editor. They're either synced from Salesforce or collated based on existing values. In the latest release, Marketo admins will have more control over the picklist values displayed to users. Adobe are adding a picklist editor to the field management screen in admin. About time.

Operationalising ABM in the Real World: Your Step-by-Step Guide

Finding it difficult to navigate the complexities of ABM. A new eGuide from CRMT Digital cuts through ABM hype, explaining how to implement and run an ABM program in practical terms. Boost ROI with practical approaches to account-based targeting, personalisation and engagement.

Marketing Operations Roadmap Matrix

A Marketing Operations Roadmap is an essential strategic tool that aligns marketing priorities with your overall business goals. Any roadmap begins with understanding where your marketing organisation is right now, and then defining where you want to get to. This free Marketing Operations Roadmap Matrix from CRMT Digital allows you to do exactly that.