The Week at Vinton #5

Clients appreciate knowing that meetings, workshops, and status updates will be recorded, transcribed, and summarised using AI, but more importantly, an AI notetaker that keeps that data secure in Salesforce and isn’t used to train a public LLM.

Consultants and clients can focus fully on discussions without furiously taking notes, improving the quality of live interactions. Automated summaries and next steps means the project can keep moving at a pace.

Written records of conversations, decisions, and actions are automatically generated, helping prevent misunderstandings later on.

With everything documented, it’s easier to onboard new team members mid-project or hand off deliverables seamlessly at the end.

Rather than viewing AI notetakers as “optional,” progressive firms are weaving them directly into their client experience. Will using AI notetakers become a standard expectation in every professional services agreement?

I had an interesting conversation with my son and his friend last weekend about how we still, mistakenly, feel that our interactions with ChatGPT or similar tools are helping to train them so that they get better with every conversation we have.

It *feels* like that’s what’s happening, but it’s not.

Large language models (LLMs) aren’t absorbing every conversation they have in real-time and getting smarter by the minute. That would be wildly inefficient and hugely expensive! Although ChatGPT’s terms say they “may use your content to train our models” (and to be clear they don’t use content from their business offerings, like our Enterprise API integration, to train at all), they aren’t doing so for every tiny aspect of a conversation.

What can make AI feel like it’s learning is when you turn on memory – which I’ve done in my personal ChatGPT subscription – but even then, it’s remembering key facts I’ve shared across sessions so it can personalise its responses and maintain continuity over time. It’s not learning, but it is retaining helpful context I’ve agreed to share.

For example, ChatGPT knows I’m a business leader in a SaaS business that’s a Salesforce partner. It knows that our Vinton solution is only suitable for Salesforce users. It remembers this and tailors future responses accordingly, without needing to be reminded every time.

RAG (Retrieval-Augmented Generation) is helping to deliver better answers. Instead of relying solely on what random data the model was trained on months (or years) ago, RAG is about giving the model access to external knowledge sources in real time (company documentation, customer conversations, product manuals) and having it pull in relevant information to answer a question more accurately.

Imagine asking, “What’s our refund policy for enterprise customers?” RAG lets the model search your internal knowledge to include the up-to-date answer. Imagine Agentforce in Salesforce – it’s grounding its answers on something real, current, and specific from your Salesforce data – not just “best guess” text generation.

Combining memory and RAG, you get AI that feels like it’s learning and evolving with you. It’s not being “trained” like we used to think of it, but rather it’s being fed the right data at the right time, and contextualising the answers based on memory and retrieval.

With tools like memory and RAG, we’re not making AI models smarter in the traditional sense. We’re making them more useful. More aligned. More like an actual teammate who remembers your goals, understands your business, and shows up prepared.

I had a genuinely enlightening conversation with a client this week. They’re a charitable organisation that regularly applies for grants.

Up until the end of last year, the unspoken rule of grant applications was simple:

The more you write, the more passionate and committed you must be to your cause, the more likely your application will be considered for the grant.

But then… AI…

Last year a US nonprofit, Candid, that provides data and insights about the social sector, conducted some research on where foundations stand on AI-generated grant proposals. 57% don’t know whether they’ve received applications created with Gen AI, but 23% say they will not accept GenAI content, and 67% are undecided.

Our client mentioned that this year they are seeing many grant applications with a very limiting cap on the total characters of the application text, and that’s because they know AI can help generate compelling submissions in seconds. The long, heartfelt submission has lost its value as a differentiator.

Instead of text volume, they’re asking for authenticity in other ways. Applicants are being encouraged to supplement short written submissions with presentations, videos, and other creative ways to show why their mission matters.

• Less (AI-generated) text.
• More (human-generated) genuine content.