AI Workflow Enhancement

Last week, I described four design patterns for AI agentic workflows that I believe will drive significant progress: Reflection, Tool use, Planning and Multi-agent collaboration. Instead of having an LLM generate its final output directly, an agentic workflow prompts the LLM multiple times, giving it opportunities to build step by step to higher-quality output. Here, I'd like to discuss Reflection. It's relatively quick to implement, and I've seen it lead to surprising performance gains. You may have had the experience of prompting ChatGPT/Claude/Gemini, receiving unsatisfactory output, delivering critical feedback to help the LLM improve its response, and then getting a better response. What if you automate the step of delivering critical feedback, so the model automatically criticizes its own output and improves its response? This is the crux of Reflection. Take the task of asking an LLM to write code. We can prompt it to generate the desired code directly to carry out some task X. Then, we can prompt it to reflect on its own output, perhaps as follows: Here’s code intended for task X: [previously generated code] Check the code carefully for correctness, style, and efficiency, and give constructive criticism for how to improve it. Sometimes this causes the LLM to spot problems and come up with constructive suggestions. Next, we can prompt the LLM with context including (i) the previously generated code and (ii) the constructive feedback, and ask it to use the feedback to rewrite the code. This can lead to a better response. Repeating the criticism/rewrite process might yield further improvements. This self-reflection process allows the LLM to spot gaps and improve its output on a variety of tasks including producing code, writing text, and answering questions. And we can go beyond self-reflection by giving the LLM tools that help evaluate its output; for example, running its code through a few unit tests to check whether it generates correct results on test cases or searching the web to double-check text output. Then it can reflect on any errors it found and come up with ideas for improvement. Further, we can implement Reflection using a multi-agent framework. I've found it convenient to create two agents, one prompted to generate good outputs and the other prompted to give constructive criticism of the first agent's output. The resulting discussion between the two agents leads to improved responses. Reflection is a relatively basic type of agentic workflow, but I've been delighted by how much it improved my applications’ results. If you’re interested in learning more about reflection, I recommend: - Self-Refine: Iterative Refinement with Self-Feedback, by Madaan et al. (2023) - Reflexion: Language Agents with Verbal Reinforcement Learning, by Shinn et al. (2023) - CRITIC: Large Language Models Can Self-Correct with Tool-Interactive Critiquing, by Gou et al. (2024) [Original text: https://lnkd.in/g4bTuWtU ]

We sent 4,495 AI SDR emails in 2 weeks and achieved the #1 response rate on our platform. But here's what nobody tells you about making AI SDRs actually work... The Metrics: ✅ 4,495 personalized messages sent in 14 days ✅ Highest response rate on our entire platform ✅ $700,000 of pipeline opportunities opened ✅ Meetings booked daily (literally got one this morning) ✅ Outperformed all our historical human SDR averages — mostly ✅ Better results than some of our human AEs The Reality Check First We had unfair advantages. SaaStr has been around since 2012, we've sold $100,000,000 in sponsorships, and people know our brand. We targeted our existing database—website visitors, past attendees, lapsed accounts—not cold lists. We spent 2 weeks doing basically nothing else: 90 minutes every morning, 1 hour every evening training our AI, plus real-time responses throughout the day. 👉What Actually Works: 1️⃣ Your AI has to add real value, not just volume There's no way we could send 4,495 good emails ourselves manually in two weeks. The key is each one has to be at the level we would write ourselves. Bad: "Hey [NAME], saw you visited our website" Good: "Congrats on your new VP role at Oracle. Since you attended SaaStr London last year, thought you'd want to know about our 2025 VC track with speakers from a16z and Sequoia..." 2️⃣ Your data is messier than you think We trained our AI on 20+ million words of SaaStr content, but still found: - Opportunities never logged in Salesforce - Missing context from AEs who never used the system - Customer relationships that existed nowhere in our CRM We literally spend time every day finding things that were missing and manually adding them to AI's knowledge base. 3️⃣ Human-in-the-loop isn't optional When prospects respond to your AI, YOU have to respond instantly at the same quality level. We have it hooked up to Slack—our phones go off at all hours because SaaStr is global. The AI creates an expectation of responsiveness. You better match it or they'll know it was "just an AI email." 5️⃣ This is additive, not replacement We still do personal emails, marketing campaigns, and have human SDRs. Results by campaign type: - Website visitors: Hit or miss - Cold outbound: Ranked 4th out of 4 campaigns - Lapsed renewal accounts: Really good results 🏋🏽♀️ The Uncomfortable Truth: It's MORE work, not less. You get 10x better output, but it requires S-tier human orchestration. E.g., we're running 30+ personas across different campaigns. 🔮 Bottom line: AI SDRs work incredibly well, but only with proper training and orchestration. After 60 days of daily improvements, you'll have something you're proud of. But you can't skip the daily 30-45 minute audit process. Full breakdown with all our tools and processes at link in comments.

Agentic AI can get expensive fast. Not because the model is always wrong. Because every agent step can add more context, more tool calls, more reasoning, and more tokens. If you want scalable agents, token efficiency becomes architecture. Here are 6 ways to save tokens in agentic workflows: → 𝗣𝗿𝗼𝗺𝗽𝘁 𝗖𝗮𝗰𝗵𝗶𝗻𝗴 Reuse repeated system prompts, instructions, and context instead of processing the same input again. Best for stable instructions, long policies, and repeated workflows. → 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗖𝗮𝗰𝗵𝗶𝗻𝗴 Reuse previous answers when a new query has the same meaning, even if the wording is different. Best for support, FAQs, internal knowledge, and repeated user requests. → 𝗥𝗼𝘂𝘁𝗶𝗻𝗴 Send each task to the right model, tool, or workflow based on complexity, cost, and accuracy needs. Simple tasks should not use your most expensive model. → 𝗞𝗲𝗲𝗽𝗶𝗻𝗴 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗖𝗹𝗲𝗮𝗻 Remove irrelevant, outdated, duplicated, or noisy information before sending context to the model. Clean context improves both cost and output quality. → 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗖𝗼𝗺𝗽𝗮𝗰𝘁𝗶𝗼𝗻 Compress long conversations or documents into shorter summaries while preserving critical facts. Useful when agents need memory without carrying the entire history. → 𝗟𝗮𝘇𝘆-𝗟𝗼𝗮𝗱𝗶𝗻𝗴 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 Load only the information needed at the moment instead of dumping everything upfront. The lesson is simple: More context is not always better. Better context is better. Token savings come from smarter caching, cleaner context, better routing, and loading only what the agent actually needs. Save this if you are building AI agents, RAG systems, automation workflows, or production GenAI apps.

When working with multiple LLM providers, managing prompts, and handling complex data flows — structure isn't a luxury, it's a necessity. A well-organized architecture enables: → Collaboration between ML engineers and developers → Rapid experimentation with reproducibility → Consistent error handling, rate limiting, and logging → Clear separation of configuration (YAML) and logic (code) 𝗞𝗲𝘆 𝗖𝗼𝗺𝗽𝗼𝗻𝗲𝗻𝘁𝘀 𝗧𝗵𝗮𝘁 𝗗𝗿𝗶𝘃𝗲 𝗦𝘂𝗰𝗰𝗲𝘀𝘀 It’s not just about folder layout — it’s how components interact and scale together: → Centralized configuration using YAML files → A dedicated prompt engineering module with templates and few-shot examples → Properly sandboxed model clients with standardized interfaces → Utilities for caching, observability, and structured logging → Modular handlers for managing API calls and workflows This setup can save teams countless hours in debugging, onboarding, and scaling real-world GenAI systems — whether you're building RAG pipelines, fine-tuning models, or developing agent-based architectures. → What’s your go-to project structure when working with LLMs or Generative AI systems? Let’s share ideas and learn from each other.

🔍 Technical AI Series by Brindha Jeyaraman Part 2: Why GPUs Are Fast (and Why They’re Still Under-Utilised) GPUs dominate AI workloads because they’re designed for massive parallelism. Thousands of lightweight cores execute the same instruction across different data elements exactly what neural networks need for matrix math. And yet, in real-world training pipelines, it’s common to see 40–60% GPU utilisation. Why? Because GPUs don’t operate in isolation. Here are the usual culprits 👇 🔹 CPU-Bound Preprocessing Data loading, tokenisation, and augmentation often run on CPUs. If this stage is slow, the GPU simply waits. 🔹 Inefficient Data Loaders Single-threaded pipelines, poor shuffling strategies, or Python overhead can starve GPUs of data. 🔹 CPU–GPU Synchronisation Overhead Frequent synchronisation points and blocking calls introduce stalls between kernels. 🔹 Memory Access Patterns Non-contiguous tensors and frequent memory allocation/deallocation reduce effective throughput. The result? You pay for expensive accelerators that spend a significant portion of time doing nothing. High-performance AI training requires: 1. Asynchronous data pipelines 2. Careful CPU–GPU coordination 3. Stable tensor shapes 4. Minimised synchronisation points Owning AI performance means owning the entire pipeline, not just the model. In the next post, I’ll explain why attention the core of modern LLMs is fundamentally a memory problem, not just a compute one. #GPUComputing #AIInfrastructure #MLOps #SystemsEngineering

I'm excited to announce the launch of AI Refinery for Sovereign and Agentic AI, a groundbreaking platform that deepens our partnership with NVIDIA. This first-of-its-kind platform champions data sovereignty and operational resilience through physical AI, paving the way for enhanced competitiveness in the journey toward agentic AI.   As I've mentioned before, I firmly believe that AI presents a unique opportunity for Europe to reinvent its economy, drive productivity, resilience, and competitiveness, and support future growth. I'm incredibly proud to see the momentum our clients are gaining, including Public Power Corporation, Roche, Kion Group, Noli, and Nestlé.   Nestlé, for instance, is launching a new AI-powered in-house service that will generate high-quality product content at scale for eCommerce and digital media channels. This initiative exemplifies the transformative potential of AI in driving business efficiency and innovation. The expansion of our AI Refinery platform is particularly significant for European organizations, enabling them to accelerate the deployment of AI agents while addressing their sovereignty concerns. This is especially crucial for the public sector and critical infrastructure industries, such as energy, telecommunications, and defense.   We continue to support our clients in maintaining control over their critical data and leveraging innovative AI solutions through this expanded AI Refinery platform. More details here: https://lnkd.in/dvekqfB6 #Noli #Nestle #PublicPowerCorporation #KionGroup #Roche #AgenticAI #AI #Accenture

Most developers are missing the real value of Claude Code. It is not just an AI that writes code. The actual power comes from how you configure it around your engineering workflow. After exploring it deeply, one thing became very clear to me: prompting is only a small part of the picture. What actually makes Claude Code powerful is understanding when to use what. Here’s the mental model that made it click for me: 1. CLAUDE.md Your project memory. Best for stack, architecture, coding conventions, and commands Claude should always know. 2. Skills For recurring patterns that should automatically activate when relevant. If you repeatedly say “remember to do this,” it probably belongs here. 3. Commands For workflows you intentionally trigger, like code reviews, scaffolding, debugging, or repetitive tasks. 4. Subagents Specialized assistants for focused responsibilities like security reviews, testing, or architecture discussions without polluting the main context. 5. MCP Servers The bridge to external systems. Useful for live docs, GitHub, databases, APIs, Jira, Slack, and other tools Claude normally would not access. 6. Hooks Deterministic automation. Great for formatting, validations, running tests, or preventing risky edits automatically. 7. Plugins A way to package workflows and reuse them across projects or teams. One thing I feel developers may get wrong early is trying to configure everything at once. The smarter approach is to start lean. Begin with CLAUDE.md, then commands, then skills. Add MCPs, hooks, and subagents only when friction starts appearing in your workflow. That’s when Claude stops feeling like just another coding assistant and starts feeling more like an engineering partner. Have you explored Claude Code yet? Curious what has actually improved your workflow.

A nice review article "Transforming Science with Large Language Models: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation" covers the scope of tools and approaches for how AI can support science. Some of areas the paper covers: (link in comments) 🔎 Literature search and summarization. Traditional academic search engines rely on keyword-based retrieval, but AI-powered tools such as Elicit and SciSpace enhance search efficiency with semantic analysis, summarization, and citation graph-based recommendations. These tools help researchers sift through vast scientific literature quickly and extract key insights, reducing the time required to identify relevant studies. 💡 Hypothesis generation and idea formation. AI models are being used to analyze scientific literature, extract key themes, and generate novel research hypotheses. Some approaches integrate structured knowledge graphs to ground hypotheses in existing scientific knowledge, reducing the risk of hallucinations. AI-generated hypotheses are evaluated for novelty, relevance, significance, and verifiability, with mixed results depending on domain expertise. 🧪 Scientific experimentation. AI systems are increasingly used to design experiments, execute simulations, and analyze results. Multi-agent frameworks, tree search algorithms, and iterative refinement methods help automate complex workflows. Some AI tools assist in hyperparameter tuning, experiment planning, and even code execution, accelerating the research process. 📊 Data analysis and hypothesis validation. AI-driven tools process vast datasets, identify patterns, and validate hypotheses across disciplines. Benchmarks like SciMON (NLP), TOMATO-Chem (chemistry), and LLM4BioHypoGen (medicine) provide structured datasets for AI-assisted discovery. However, issues like data biases, incomplete records, and privacy concerns remain key challenges. ✍️ Scientific content generation. LLMs help draft papers, generate abstracts, suggest citations, and create scientific figures. Tools like AutomaTikZ convert equations into LaTeX, while AI writing assistants improve clarity. Despite these benefits, risks of AI-generated misinformation, plagiarism, and loss of human creativity raise ethical concerns. 📝 Peer review process. Automated review tools analyze papers, flag inconsistencies, and verify claims. AI-based meta-review generators assist in assessing manuscript quality, potentially reducing bias and improving efficiency. However, AI struggles with nuanced judgment and may reinforce biases in training data. ⚖️ Ethical concerns. AI-assisted scientific workflows pose risks, such as bias in hypothesis generation, lack of transparency in automated experiments, and potential reinforcement of dominant research paradigms while neglecting novel ideas. There are also concerns about the overreliance on AI for critical scientific tasks, potentially compromising research integrity and human oversight.

I've watched organizations rush to implement AI tools across their revenue functions, often with mixed results. Today, I'm sharing a crucial insight: the companies seeing transformative results are not those with the most advanced tech stacks. Instead, they deploy AI with surgical precision at the intersection of efficiency and trust. In my latest piece, I break down specific AI tools reshaping revenue operations and offer strategic guidance on implementing them without eroding the customer trust that underpins sustainable growth. Key takeaways: 🎯 Conversation Intelligence Platforms (Gong, Chorus): Not just for call analysis, but for scaling successful behaviors while maintaining authentic customer interactions 🎯 Predictive Lead Scoring (MadKudu, 6sense): Allowing targeted deployment of human capital against high-probability opportunities (with critical guardrails) 🎯 Personalization Engines (Mutiny, Optimizely): Creating tailored experiences without increasing operational complexity or crossing the "creepy line" 🎯 Content Generation (Jasper.AI, Copy.ai, Claude.ai): Achieving velocity without sacrificing quality (but still requires human oversight to be more, well, human). 🎯 Customer Journey Orchestration (Drift, a Salesloft company, Qualified): Creating guided buying experiences that feel personalized while operating at scale 🎯 AI Assistants (Grok, ChatGPT): Rapid iteration and testing of multiple approaches before committing resources The most successful revenue organizations aren't those using the most AI but those using AI most strategically. There is a competitive advantage in knowing where NOT to automate - in preserving human connection where it creates differentiating value. What AI tools are you implementing in your revenue operations? And more importantly, how are you measuring their impact beyond efficiency metrics? Read more here: https://lnkd.in/e4Ang6Nj __________ For more on growth and building trust, check out my previous posts. Join me on my journey, and let's build a more trustworthy world together. Christine Alemany #Strategy #Trust #Growth

Most LLM-infused RAG-based vector DB searches could be done just as easily with SQL or SPARQL queries. We must stop overengineering workflows just to get the agentic label. All that accomplishes is making products more complex and expensive. Use the simplest approach to create the outcome, and more use cases will be feasible. Use AI as an augmentation layer when reinventing workflows. Remember, in an agentic workflow, not every step must be handled by an LLM. The power of agents is their ability to use tools like people do. Leverage tools, information architecture, and other models to keep costs down. I had a startup founder ask about building a recommendation system with LLMs yesterday. In ten minutes, I helped them save millions in AI API costs by working through a simple workflow evaluation with them. We discovered that Step 1 was the only place that an LLM could provide value. They wanted to enable users to make their requests with natural language, so the app could extract more information to use for personalization. Intent detection, multistep conversation (asking clarifying questions), and gathering granular information from unstructured data are all great LLM use case categories. What happens next doesn’t have to use expensive models. The value creation came from reinventing the user workflow, not the operational workflow. A menu-based or categorical selection workflow isn’t feasible for this use case due to the complexity of the subject domain. The only way to extract all the information required to serve a high-quality recommendation was conversationally. The rest of the workflow was fairly straightforward. No vector DBs or RAG required. Once all the details were extracted, the rest of the workflow fit into conditional statements and logical steps. That’s the power of AI workflow evaluations. It’s just as valuable to define what AI shouldn’t do to make the unit economics work for more use cases. The upfront work to make the data more structured and define the logic could also benefit from an LLM’s help. That one-time cost can save the business from having to use an LLM repeatedly as part of the product’s workflow. Just because we could use AI doesn’t mean we should. I advise clients to run workflow evaluations and only use AI when it provides more value than an alternative, lower-cost approach.