🌟 Vasilij’s Note
The AI landscape took a practical turn this week – major releases are all about getting real work done. OpenAI’s latest model is pushing boundaries in everyday tasks, and chatbots are evolving into full-blown app platforms. I also spotted big players like Salesforce weaving advanced AI into business workflows, making “autonomous agents” less of a buzzword and more of a Tuesday afternoon. Below I dive into what GPT‑5 means for us and highlight new tools (and skills) to keep your team ahead. Let’s jump in!
In Today's Edition:
This Week in Agents | What Changed
ChatGPT becomes an app platform → OpenAI launched ChatGPT Apps with a new Apps SDK, allowing developers to embed interactive mini-apps directly inside ChatGPT. Early partners (Booking.com, Canva, Expedia, etc.) are live, enabling users to browse real estate, design slides and more – all in-chat.
Impact: This shifts ChatGPT from a Q&A bot to an app store of AI services, giving businesses access to 800M weekly users without building a traditional app. openai.com
Salesforce bets big on AI agents → Salesforce expanded its AI suite with Agentforce 360, integrating GPT-5 and Claude models. Employees can now interact with company data through ChatGPT, Slack, or Salesforce apps. A new Agentforce Commerce feature even lets merchants sell via ChatGPT’s Instant Checkout.
Impact: Enterprise-grade AI agents are going mainstream. Teams get out-of-the-box AI copilots across support, sales, and commerce – tightly integrated with CRM data. salesforce.com
Claude gets skills → Anthropic introduced “Agent Skills,” essentially plugin packs (instructions, scripts, data folders) that Claude can load automatically for tasks like Excel manipulation or brand-style enforcement.
Impact: Highly customizable agents. Your Claude agent can become a specialist in your workflows without custom code – improving speed, consistency, and accuracy. claude.ai
We are using Google Workspace at AiGentic Lab. I will be doing a video on Google Agents in the upcoming week or two. If you are interested to see it in action, subsribe here no to miss it: https://www.youtube.com/@vasilijnevlev
Upskilling Spotlight | Free Google Training
Google: Prompting Essentials (Coursera, ~4 hours) — Outcome: Master the art of writing effective prompts for generative AI. This quick course walks through Google’s approach to prompting, covering techniques for common business tasks and a five-step method for crafting effective. By the end, you even build a simple homebrew AI agent and earn a shareable certificate. coursera.org
Vanderbilt Univ: Generative AI Strategic Leader (Coursera, ~40 hours) — Outcome: Gain a leader’s perspective on implementing AI and AI agents in an organisation. This longer specialisation covers how AI systems work, scaling AI in business, and decision-making with AI. It emphasises practical uses of AI agents (the chatbots that act autonomously with custom data) and includes a full module on ChatGPT prompts. Great for managers who need to drive AI adoption strategically. (certificate upon completion) coursera.org
Maker Note | What I built this week
This week, my team and I automated part of our sales process using an AI agent. We set up a custom agent to handle post-demo follow-ups – drafting personalised emails and updating our CRM notes after sales calls. It’s a work in progress, but here are a few takeaways so far:
Data accuracy is everything - We had to ensure the agent pulls fresh CRM data for each prospect; even small data lags led to awkward emails. Lesson: integrate your databases carefully so the AI isn’t working with stale info.
Define the guardrails - We gave the agent a strict template and tone for emails (and approval steps) so it stays on-brand. Early on, it wrote overly lengthy emails – a quick adjustment to its instructions fixed that. Lesson: even autonomous agents need clear guidelines and the occasional human check, especially customer-facing ones.
Focus on one workflow at a time - Rather than automating the whole sales cycle, we started with follow-up emails (a contained task with clear success metrics). This made it easier to measure impact – and indeed, our reps saved about an hour each day.
Next week, we will be expanding the agent to handle meeting scheduling as well.
Operator’s Picks | Tools To Try
OpenAI Connector Registry — Best for giving agents secure access to internal apps (Salesforce, Drive, Slack, etc.) with minimal setup. Built into AgentKit.
Why try it: eliminates heavy IT integration work; enables agents that actually act on your business data. openai.comClaude Agent SDK — For teams building Claude-powered agents with memory, tool permissioning, and multi-step task coordination.
Why try it: Claude’s long context window + reliable reasoning shine in complex workflows. claude.com
Deep Dive | GPT‑5 in Action: Should Your Business Upgrade?
Read full report here: https://openai.com/index/introducing-gpt-5/
TL;DR;
GPT‑5 is here, and it’s a major leap – faster and smarter with expert-level reasoning across coding, writing, math, and even health. It combines what used to be multiple models into one unified system, so it knows when to give you a quick answer versus when to “think”. For mid-size businesses, GPT‑5 can potentially streamline everything from code dev to content creation. But new power comes with new considerations: there’s buzz (and some backlash) about its changing style and the cost/ROI of upgrading. Let’s break down what’s new and how to decide if GPT‑5 deserves a spot in your workflow.
On paper
Big brain boost: GPT‑5 is touted as OpenAI’s smartest, fastest, most useful model. Benchmarks back that up – it outperforms previous models in key areas and scores at state-of-the-art levels on coding, math, and multimodal. In plain English, it solves hard problems more reliably now (fewer errors, more logic) and even handles images or graphs with better accuracy.
Expertise across tasks: This model isn’t just a better chatbot – it’s an all-in-one AI worker. OpenAI specifically improved its performance in three common business use cases: writing, coding, and health. It can draft and edit complex documents with more nuance (even decent poetry, see below), write front-end code or debug large codebases in one go, and provide more context-aware answers on medical or technical queries. It also drastically cuts down “hallucinations” and off-target rambling, meaning you spend less time sifting out wrong answers.
Unified & “agentic”: GPT‑5 introduced a real-time router system. Instead of you picking a turbo vs. heavy model, it automatically routes easy questions to a quick responder and harder tasks to a deeper reasoning mode (and even taps GPT‑5 pro extended reasoning for ultra-complex asks). This not only speeds up responses, it also enables more agent-like behavior – GPT‑5 is better at multi-step workflows and using tools autonomously. In evaluations, it showed significant gains in carrying out instructions and coordinating tools to complete tasks end-to-end. In short, it’s closer to an autonomous assistant that can figure out how to get to an answer (by calling APIs, searching, etc.) rather than just spitting out an answer.
Open options: Interestingly, alongside GPT‑5’s release, OpenAI also open-sourced two smaller models (20B and 120B parameters) under Apache. These aren’t GPT‑5’s equals, but they let businesses fine-tune and run decent models on-premises. It shows OpenAI acknowledging the need for private, cost-effective solutions – and it means you might not have to use GPT‑5 for every task. For instance, you could run an open 20B model internally for simple FAQ bots (no data leaves your servers) and reserve GPT‑5 for the heavy lifts like complex analysis.
In practice
Productivity gains (and adjustments): Early adopters report real boosts in productivity. Everyday tasks – summarising reports, drafting proposals, debugging code – are faster and more “one-shot” with GPT‑5. A rough idea can turn into a polished memo in minutes. That said, users have noticed GPT‑5’s style and tone can differ from its predecessors. In fact, when OpenAI rolled it out to the ~700 million ChatGPT users, some fans of the old GPT-4 (or “4o”) complained the new model’s personality was more straightforward and less quirky. For business use, a more direct tone is fine (even preferred), but if you had carefully crafted prompt workflows for GPT-4, expect to re-tune them for GPT-5’s style and length. It’s a bit like a new employee – more capable, but you’ll brief it differently to get the best results.
Costs & access: The good news – GPT‑5 is available to all ChatGPT users, even free tier (with some limits), so you can try it right now in the ChatGPT interface. Plus subscribers get higher usage quotas, and Pro subscribers can access “GPT‑5 pro” for those really tough jobs. API access is also rolling out, likely at a cost similar to or slightly above GPT-4’s pricing. For businesses, if you’re already on ChatGPT Enterprise, you probably have GPT-5 by default. The key consideration is value for money: GPT‑5 can be overkill (and more expensive) for trivial tasks that older models handle fine. It shines in complex scenarios – e.g. generating a whole marketing campaign package with images, or performing deep analytics on customer feedback. Leverage it where its extra brainpower actually saves you time or money; otherwise, using cheaper models for routine stuff might remain a smart strategy.
Trust and verification: OpenAI has worked on making GPT‑5 more factual and less prone to going off the rails. It even reduces sycophancy, meaning it’s less likely to just agree with false user assumptions. Still, no AI is perfect. In our internal tests, GPT‑5 occasionally confidently gave an incorrect stat or misinterpreted a niche query (though less often than GPT-4 did). The advice here stays the same: use GPT‑5 as a super-smart assistant, but keep a human in the loop for critical outputs. One technique some teams use is the “SIFT” method from media literacy – Stop, Investigate, Find better info, Trace sources – applied to AI output. In practice, that might mean double-checking the key facts it gives, or having it show sources (with tools or by asking it). The bottom line: GPT‑5 will confidently get more right – and also confidently get some wrong. Your workflow should assume both.
Issues / Backlash
“Tool or crutch?” With great power comes great... overreliance. One emerging concern is that junior staff (or any of us, really) might lean too heavily on GPT‑5 and lose skills or judgment. For example, if every line of code or email copy now comes from GPT‑5, are we rubber-stamping things we don’t fully vet? Some managers are already rotating tasks that don’t use GPT to keep people sharp. It’s worth instilling a mindset in your team that GPT‑5 is a starting point or accelerator, not the final arbiter. Make review steps explicit in your processes (especially for legal, financial, or brand-sensitive content).
Model behavior changes: OpenAI’s rapid updates mean the AI’s behavior can tweak over time (sometimes without warning). We saw this with GPT-4 earlier, and GPT‑5 will be no exception as OpenAI fine-tunes it or applies new safety rules. This can be a “backlash” point if an update suddenly makes the AI less verbose or more cautious, and it affects your outputs. Mitigation: maintain a simple prompt library or tests for your key use cases. If something feels off one day, you can quickly A/B test a previous prompt to see if the model changed. OpenAI is getting better at communication here, but it’s ultimately their model – stay adaptable.
Competitive landscape: GPT‑5 is arguably the new gold standard, but rivals are close on its heels. Anthropic’s Claude 4 (and Claude Sonnet 4.5 coding model) claims some superior coding ability, and Google’s Gemini is expected to be a heavyweight. This “backlash” is more of a market one: businesses might feel nervous about lock-in with one vendor’s AI if another leapfrogs in 6 months. We already see strategies like Salesforce offering both OpenAI and Anthropic models. One pragmatic approach is to stay model-agnostic in how you build AI into your workflows. Use modular integrations (or platforms like Dust, featured below) so you can swap out the backend model if needed. Competition will drive better pricing and features – not a bad thing – but avoid over-committing to proprietary formats or prompts you can’t reuse elsewhere.
My take (what to do)
Give GPT‑5 a pilot run: If you haven’t yet, identify one or two processes in your business and experiment with GPT‑5 there. Good targets are things that were a stretch for GPT-3.5/4 – like generating a full draft of an ebook chapter, analysing a complex dataset, or writing functional code modules. Measure any improvement (speed, quality) against your usual approach. This will tell you quickly where the ROI lies. In our case, GPT‑5 cut a proposal writing process from 3 hours to 1 hour, simply because it needed less back-and-forth prompting to get a solid result.
Upskill your team: New capabilities mean a small learning curve. Host a brown-bag session on “GPT-5 tips” for your staff. For example, teach how to invoke its “extended reasoning” mode in ChatGPT (e.g. by saying “let’s think this through step by step” or using the GPT-5 Pro if available) when faced with a hard problem, versus keeping prompts concise for quick answers. Share a cheat-sheet of prompt examples that leverage GPT‑5’s new strengths (like its ability to analyse an image + text together, or generate more creative outputs). Investing an hour now in training could save each employee many hours down the line in more effective AI use.
Mind the costs & set guidelines: As with any powerful tool, set some internal guidelines for GPT‑5 use. For instance, decide when it’s appropriate to use the API for automated tasks – perhaps high-value tasks like drafting client reports, but not for trivial one-liners where cheaper models suffice. Implement basic usage tracking if possible (ChatGPT Enterprise provides some analytics). This helps prevent surprise bills and also surfaces which teams are getting value. Also consider data sensitivity: remind folks what not to input into any third-party AI without clearance. OpenAI has enterprise privacy promises, but a good rule is to avoid pasting unredacted customer data or proprietary code unless you’re using their approved secure channels or have a data processing agreement.
Keep an eye on the horizon: Finally, recognise that GPT‑5 is a milestone, not the endgame. AI capabilities will keep advancing. It’s less about just adopting GPT‑5 and more about adopting an AI-forward strategy. Encourage your team to continuously share new use cases they discover. Maybe one developer finds GPT‑5 + the new ChatGPT Apps can semi-automate market research; another uses it to simulate a customer for training new hires. Make a forum (Slack channel, weekly stand-up segment) for these AI hacks. By creating a culture that iterates on how to best use tools like GPT‑5, your business won’t just have an AI – it’ll develop an AI advantage.
Spotlight Tool | dust.tt
Dust — The “agentOS” platform to build your own AI agents. Dust lets you create custom AI agents in minutes, connected to your internal data and apps.
Purpose: Help teams automate complex workflows with AI workers instead of just chatbots.
Edge: Dust agents can take actions across tools on your behalf – not just draft content. For example, a Dust agent can automatically create a Jira task, update a Salesforce record, or schedule a meeting based on an email. Under the hood, it uses protocols like Anthropic’s MCP for secure data/app access (think of it as a “USB-C for AI” connecting to your databases). It also has built-in permission controls and respects data boundaries (no data retention by the AI model), addressing enterprise security needs.
In short, Dust provides the building blocks of an AI agent workforce: you focus on defining the tasks and rules, and the agents handle the busywork across your cloud software.
Let me know if you would like to see a demo!
Try it: dust.tt
Did you find it useful? Or have questions? Please drop me a note. I respond to all emails. Simply reply to the newsletter or write to [email protected]
