How AI is Transforming Crisis Communications: SFT and RAG Explained
The purpose of this article is to use crisis communication examples to point out what the concepts Supervised Fine-Tuning (SFT) and Retrieval-Augmented Generation (RAG), which are closely related to artificial intelligence, mean.
When a crisis hits, every word matters. Whether it’s a product recall, a data breach, or a reputational threat, communications professionals need to respond fast, carefully, and consistently. And often under extreme pressure.
Today, artificial intelligence (AI) is becoming a powerful partner in this work. Two technologies in particular are changing the way crisis communications, marketing, and PR agencies operate:
- Supervised Fine-Tuning (SFT)
- Retrieval-Augmented Generation (RAG)
If those terms sound technical, don’t worry. This compact article breaks them down in a way that anyone in the field can understand. You’ll also see how your agency or department could use them to gain a real competitive advantage.
Supervised fine-tuning: Teaching AI to speak your language
So, what is SFT, aka supervised fine-tunig? Think of SFT like giving a smart assistant a crash course in your agency’s voice, style, and best practices. Start with a pre-trained AI model, like ChatGPT or other language models. Then, you train it further on your own examples, things like past press releasesm successful social media posts, client-specific messaging, and award-winning campaigns for instance.
The goal would be to customise the model so it sounds like your agency every time it writes. It doesn’t “learn new facts,” but it learns how you like to say things.
One good example could crisis communications, and how to deal with it.
Imagine your agency specialises in high-stakes PR: product recalls, executive scandals, cyber breaches, etc.
You could fine-tune a model on rather straight forward outcomes like crisis press releases, apology letters, internal memos, and media Q&A templates.
Now, when a real crisis happens, you can generate ready-to-use, on-brand drafts within seconds, even in the middle of the night, all reflecting your tone, structure, and sensitivity.
The results speak for themselves. A custom AI helps you respond faster, more consistently, and with less risk of tone-deaf or sloppy messaging.
Retrieval-augmented generation: Letting AI pull the right information
While fine-tuning teaches the AI how to speak, Retrieval-Augmented Generation (RAG) helps the AI find the right things to say. How? Based on real, up-to-date information from your very own files.
Here’s how it works:
You upload important documents into a searchable system, e.g. brand guidelines, legal templates, client playbooks, regulatory documents.
When you ask the AI a question or prompt it to write something, it searches your stored content first. Then it generates a new answer or document based on what it found.
In other words, RAG connects your internal knowledge with the AI’s writing power making the outputs far more accurate and grounded.
Let’s take crisis communications again as an example. A contamination issue hits a major food brand you work with. Instead of guessing, the AI immediately pulls past similar crisis statements, industry-specific legal wording, and uses the client’s tone-of-voice guide
It then generates an internal staff memo or media statement that’s perfectly aligned and backed by real examples. The results should be encouraging. AI responses are faster, smarter, and safer, because they’re based on your agency’s or firms real expertise.
Why this matters for PR and Marketing agencies and/or departments?
| YOUR NEED | HOW AI HELPS |
| Respond faster under pressure | AI drafts ready in seconds |
| Maintain the right tone | AI trained on your style (SFT) |
| Use the correct facts and phrasing | AI pulls from your documentation (RAG) |
| Help new staff | AI amplifies your best practices |
| Scale your team’s output | AI lets one strategist do the work of several |
Whether you’re building proactive messaging or reacting to fast-moving crises, these tools allow you to move faster without sacrificing quality or control.
But, AI still doesn’t replace the human touch. Just like AI as a whole, supervised fine-tuning and retrieval-augmented generation aren’t about replacing communication professionals, but about giving you a powerful new tool or an extra brain that thinks like your agency/department and remembers your best work, ready to act instantly when it matters most.
If you’ve ever wished you could clone your best writer, strategist, or crisis manager… this might be the next best thing!
Postscript: a slightly more technical approach
Supervised Fine-Tuning (SFT)
SFT is a process where you take a pre-trained language model (like GPT) and further train it on a specific dataset that relates to your task or domain. This helps the model behave more predictably and appropriately for your use case.
Important note: You’re not adding new knowledge to the model per se, but you’re guiding its behavior like teaching it how to respond more like you want it to.
Example: Say you’re a law firm. You might fine-tune a model on legal Q&A pairs, so it responds like a legal assistant instead of a general-purpose chatbot.
Tech side: Requires some Python, but it’s mostly about configuring settings, choosing the right dataset, and running the process using available tools.
Why this matters?
- If you already work in tech or data, both of these are quite learnable. You’re likely familiar with things like Python, APIs, or databases, which are the main building blocks.
- These approaches are key to making LLMs useful in real business settings, especially when privacy, accuracy, or domain-specific knowledge matters.
- These are high-demand skills in the job market right now. Companies want people who can make language models work with their own data.
Pretrained Language Model
Domain-specific training data
Fine-tuning Process (Python + config)
Custom Fine-Tuned Model
Ask a question → and get customised response!
If you’re a fashion e-commerce company, you fine-tune the model with your product descriptions and customer support chats. And now it speaks your brand language when answering customer questions.
Visual Workflow: SFT in a Comms/Marketing/PR Agency
Pretrained Language Model
Add your agency's style + tone examples
TRAINING DATA
- Award-winning press releases
- Email campaigns for luxury brands
- Social posts with high engagement
- Media pitches with perfect tone
Fine-Tune with Python + Config Tools
Agency-branded AI assistant: knows your style & voice! Oh, techwise, you’d just need…
- A few hundred (or thousand) high-quality examples (your best content!)
- Open-source tools like Hugging Face + LoRA or PEFT for cost-effective fine-tuning
- Python know-how or a hired ML partner to run the tuning job once: and then you’re good!
About the main picture: it is generated with Midjourney. Alt text: AI-generated image of a woman typing on a computer at a desk. Translucent spheres resembling soap bubbles, filled with data, float in the air.