Business Process Automation

Few Lessons from Deploying and Using LLMs in Production Deploying LLMs can feel like hiring a hyperactive genius intern—they dazzle users while potentially draining your API budget. Here are some insights I’ve gathered: 1. “Cheap” is a Lie You Tell Yourself: Cloud costs per call may seem low, but the overall expense of an LLM-based system can skyrocket. Fixes: - Cache repetitive queries: Users ask the same thing at least 100x/day - Gatekeep: Use cheap classifiers (BERT) to filter “easy” requests. Let LLMs handle only the complex 10% and your current systems handle the remaining 90%. - Quantize your models: Shrink LLMs to run on cheaper hardware without massive accuracy drops - Asynchronously build your caches — Pre-generate common responses before they’re requested or gracefully fail the first time a query comes and cache for the next time. 2. Guard Against Model Hallucinations: Sometimes, models express answers with such confidence that distinguishing fact from fiction becomes challenging, even for human reviewers. Fixes: - Use RAG - Just a fancy way of saying to provide your model the knowledge it requires in the prompt itself by querying some database based on semantic matches with the query. - Guardrails: Validate outputs using regex or cross-encoders to establish a clear decision boundary between the query and the LLM’s response. 3. The best LLM is often a discriminative model: You don’t always need a full LLM. Consider knowledge distillation: use a large LLM to label your data and then train a smaller, discriminative model that performs similarly at a much lower cost. 4. It's not about the model, it is about the data on which it is trained: A smaller LLM might struggle with specialized domain data—that’s normal. Fine-tune your model on your specific data set by starting with parameter-efficient methods (like LoRA or Adapters) and using synthetic data generation to bootstrap training. 5. Prompts are the new Features: Prompts are the new features in your system. Version them, run A/B tests, and continuously refine using online experiments. Consider bandit algorithms to automatically promote the best-performing variants. What do you think? Have I missed anything? I’d love to hear your “I survived LLM prod” stories in the comments!

A very hard pill to swallow to quite a few organizations: Business Process Automation does not equal Business Process Improvement. These are two different disciplines, and automation may be one of the steps/tools in the overall process improvement initiative. But automating a process does not improve it by default. In fact, automation must be done after the process has already been reviewed and already improved. Otherwise, the automation initiative will most likely fail to achieve its goals because: 📌 It is more time-consuming to automate an inefficient process, meaning it will take longer to implement a solution 📌 The more effort needed means it is also more expensive, effectively leading to lower (if any) ROI 📌 Automating inefficient processes AS-IS results in inefficient solutions that run slower and require more support, effectively boosting the total cost of ownership exponentially To put it simply: 💩 in ➡️ 💩 out. A review of the process before attempting to automate might save lots of time and money, even if it means an extra step and some extra investment up front. It will most likely lead to a better solution design that will be easier (and thus cheaper) to implement and maintain. In some scenarios, it may even lead to a case where the process becomes so efficient that further automation isn't even needed. It has happened to us in the past on numerous occasions. It may seem counterproductive for me to tell my clients to not automate something, effectively losing the income we could have gained from delivering the solution. But what it actually lead to was happier clients that would keep coming back for more and eventually showing up with a process that both is efficient and actually makes sense to automate. So, whenever considering automation, make sure that you review and improve the process first, and then automate. Not the other way around. And if you don't know how to, find someone who can help you and does not simply suggest automating AS-IS (that's usually a huge red flag).

𝗣𝗿𝗼𝗰𝘂𝗿𝗲𝗺𝗲𝗻𝘁 𝗶𝘀 𝗯𝗲𝘁𝘁𝗶𝗻𝗴 𝗼𝗻 𝘁𝗵𝗲 𝗱𝗶𝘀𝗿𝘂𝗽𝘁𝗶𝘃𝗲 𝗻𝗮𝘁𝘂𝗿𝗲 𝗼𝗳 𝗔𝗜 𝘄𝗵𝗶𝗹𝘀𝘁 this is also reality in Procurement: ▪️Supplier onboarding via paper forms and multiple portals ▪️Receiving PDF invoices to download, print and stamp ▪️Contract visibility and operationalisation gaps ▪️No tracking of savings vs benefit realisation ▪️Collection of wet signatures on contracts ▪️Outdated pricing and delivery records ▪️Complex delegation of authority rules ▪️Manual KPI calculations The benefits AI can unlock are manyfold. Some even dream of autonomous procurement down the road. Having helped companies through their Digital Procurement and AI discovery programs i can only stress the importance of not jumping ahead without a plan. There is continuous foundational work needed on process simplification and data quality as a prerequisite to harvest scaled AI value. AI is poised to give wings to Procurement enabling productivity and maturity in decision-making but will only do so if process and data runways are ready and cleared of the legacy junk... 🚩else it will just be another smart solution, bouncing more work back to humans. ❓Where do you see necessary pre-work for AI to take off ❓What results are you observing upon implementation of AI in Procurement #procurementtransformation #digitalprocurement #ai #intelligentautomation #fridayvibes

Last quarter, I worked with the MD of a heavy equipment manufacturer who believed AI would make status reports clearer and give leadership better visibility into project progress, but while the dashboards improved and the data looked sharper, the actual profit margins did not improve because delays were still being identified too late to prevent cost overruns. By the time problems appeared in reports, the financial impact had already occurred, and in 2026, with tighter compliance requirements and thinner operating buffers, that delay between issue and action is no longer affordable. What has truly changed is not reporting quality but execution speed, because AI systems can now reallocate resources, adjust schedules, and flag bottlenecks immediately instead of waiting for weekly or monthly review cycles; in plant upgrade programs and supplier transitions, I have seen problems addressed at the point of occurrence rather than after escalation. When corrective action happens closer to where the issue starts, delivery risk declines and cycle times shorten, since decisions are triggered by live data rather than by meetings or manual coordination. The main weakness I continue to see is governance, because many AI agents operate on fragmented data sources without clear ownership of decision rights, which leads teams to override outputs they do not trust and reintroduce manual controls that slow everything down, creating a false sense of stability where dashboards remain green but margin pressure builds quietly underneath. Two mistakes appear repeatedly. The first is treating AI as an advanced reporting layer, because manufacturing projects depend on operational control rather than visibility alone, and insight does not prevent delay unless the system is allowed to act within clearly defined boundaries. The second is deploying AI without defining who owns the decisions it influences, because manufacturing plants rely on accountability structures, and when escalation paths are unclear, agents can create conflicting actions that slow adoption and reduce confidence across teams. If you are beginning this journey, start by mapping a single workflow where approvals consistently delay progress, such as change requests during shutdown planning, and introduce AI only where decision rules are already stable and measurable, while avoiding areas that depend on negotiation or human judgment.  #AIInProjectManagement #AgenticAI #ExecutiveLeadership #FutureOfWork #OperationalExcellence0 #DecisionIntelligence #EnterpriseAI #ProjectGovernance #DigitalTransformation #AIForCEOs #BusinessExecution #AIStrategy

AI models like ChatGPT and Claude are powerful, but they aren’t perfect. They can sometimes produce inaccurate, biased, or misleading answers due to issues related to data quality, training methods, prompt handling, context management, and system deployment. These problems arise from the complex interaction between model design, user input, and infrastructure. Here are the main factors that explain why incorrect outputs occur: 1. Model Training Limitations AI relies on the data it is trained on. Gaps, outdated information, or insufficient coverage of niche topics lead to shallow reasoning, overfitting to common patterns, and poor handling of rare scenarios. 2. Bias & Hallucination Issues Models can reflect social biases or create “hallucinations,” which are confident but false details. This leads to made-up facts, skewed statistics, or misleading narratives. 3. External Integration & Tooling Issues When AI connects to APIs, tools, or data pipelines, miscommunication, outdated integrations, or parsing errors can result in incorrect outputs or failed workflows. 4. Prompt Engineering Mistakes Ambiguous, vague, or overloaded prompts confuse the model. Without clear, refined instructions, outputs may drift off-task or omit key details. 5. Context Window Constraints AI has a limited memory span. Long inputs can cause it to forget earlier details, compress context poorly, or misinterpret references, resulting in incomplete responses. 6. Lack of Domain Adaptation General-purpose models struggle in specialized fields. Without fine-tuning, they provide generic insights, misuse terminology, or overlook expert-level knowledge. 7. Infrastructure & Deployment Challenges Performance relies on reliable infrastructure. Problems with GPU allocation, latency, scaling, or compliance can lower accuracy and system stability. Wrong outputs don’t mean AI is "broken." They show the challenge of balancing data quality, engineering, context management, and infrastructure. Tackling these issues makes AI systems stronger, more dependable, and ready for businesses. #LLM

If you are building AI for a regulated industry, most of the popular tools disappear the moment compliance enters the conversation. The API keys need to be enterprise tier. The data cannot touch certain environments. The workflows need specific approvals baked in, not bolted on. You cannot use half the integrations that other companies use freely because they have not signed the right agreements. This is the reality that teams in healthcare, financial services, and government face every day. The tools that were designed to make AI accessible were not designed with these constraints in mind. What I find interesting is that this is not a technology limitation. The models work fine. The constraint is the entire system around the model. Who controls the data. Where it lives. What audit trail exists. Who approved what. Building AI for regulated industries is less about the AI and more about building trust into every layer of the system. If you are working in this space, the question is not "can we use AI?" It is "can we prove exactly what this AI did and why?"

Why We're Not Seeing AI ROI in 2026 — And Why That Won't Change This Year The numbers are in, and they're uncomfortable. Despite billions flowing into AI, we're not seeing the returns. And here's the uncomfortable truth: we won't see them for the remainder of 2026 either. The ROI Gap is Real and Growing PwC's 2026 CEO Survey reveals that 56% of CEOs report no measurable revenue increase or cost reduction from AI investments. Only 12% report achieving both. That's not a technology problem — that's a strategy and execution problem. Bain & Company's research shows that while 37% of organizations targeted cost reductions of 11-20%, nearly 40% landed below 10%. Yet 90% are increasing their budgets anyway. Gartner's latest data: 72% of AI projects in infrastructure and operations fail to fully meet ROI expectations. Only 28% succeed outright. The Problem Isn't the Technology Here's what the research keeps telling us but we keep ignoring: We're automating broken processes. Bain puts it bluntly — AI doesn't fix workflow debt; it locks it in, speeds it up, and makes it vastly more expensive to unwind. We're throwing AI at processes that were never optimized in the first place. Data is still the wall. 41% of organizations cite data access and integration as the #1 barrier to AI progress. After a decade of data modernization investments worth hundreds of billions globally, we still can't reliably get access to our own data. The "autonomous" agents aren't autonomous. Only 7% of companies are running fully autonomous agents in production. The dominant model requires human approval. Yet investment cases are built on full automation economics. The gap between the business case and reality is bankrupting CFOs. We're funding the next wave with savings that never arrived. 44% of companies plan to self-fund AI agent investments from "prior automation savings" — but those savings consistently came in below target. We're compounding risk, not managing it. Usage doesn't equal value. Most AI deployments are still doing low-complexity work — summarization, basic Q&A — that doesn't move the needle. What the 12% Are Doing Differently CEOs who report financial returns are 2-3x more likely to have embedded AI extensively across decision-making and demand generation. They haven't just bought licenses — they've rewired operations. They've treated AI as a CEO-level mandate, not an IT project. The Hard Truth The era of unmeasured experimentation is over. We spent 2024-2025 celebrating pilots and "AI strategies." Now the board wants auditable outcomes, not slide decks about potential. We need to stop asking "Where can we apply AI?" and start asking "If we were designing this process from scratch today, what would it look like?" Only then should the technology conversation begin. The window to get this right is narrowing. The organizations that figure this out in 2026 will have a compounding advantage. The rest will keep writing checks hoping the ROI eventuall…

Ever feel like your team is stuck in an endless loop of manual data entry? (Automation Tip Tuesday 👇) That’s exactly where one of our clients — an education consulting firm — found themselves. They were juggling a whole tech stack of tools that didn’t “talk”  to each other, creating inefficiencies and double work. We started with a look into their sales workflow. 🔹 Sales data lived in HubSpot, but once a deal closed, someone had to manually update Asana to track project progress. 🔹 Internal teams worked from one Asana board, but clients needed visibility into their own project timelines — cue more manual updates. 🔹 With so much repetitive data entry, valuable time was being wasted on low-impact admin work. Here’s what we did: 🔗 HubSpot → Asana automation: We created an integration that auto-generates project tasks in Asana when a deal reaches a certain stage in HubSpot. No more copy-pasting! 📢 Internal and client boards sync: Internal progress updates in Asana now automatically reflect on client-facing Asana projects, reducing the back-and-forth. Less busywork, more productivity. By eliminating duplicate data entry, the team saved 10+ hours per week — time now spent on strategy and client success. When your tools work together, your team can focus on what really matters. Where is your team losing time? Drop a comment below! ⬇️ -- Hi, I’m Nathan Weill, a business process automation expert. ⚡️ These tips I share every Tuesday are drawn from real-world projects we've worked on with our clients at Flow Digital. We help businesses unlock the power of automation with customized solutions so they can run better, faster and smarter — and we can help you too! #automationtiptuesday  #automation #workflow #efficiency

The AI Agent Reality Check: You Are NOT Behind If you thought 2025 was the year of mass AI Agent adoption, the latest data offers a dose of reality: the majority of the business world is still navigating the initial transition. 1. The Organizational Bottleneck According to the latest Deloitte Tech Trends 2026 report, adoption remains low: 30% of the surveyed organizations are exploring agentic options, with 38% piloting solutions and only 14% having solutions ready to deploy. The number of organizations actively using the systems in production is even lower, at 11%. Furthermore, 42% of organizations report they are still developing their agentic strategy road map, with 35% having no formal strategy at all Deloitte states the core issue for adoption as organizational: companies try to automate broken processes instead of fundamentally redesigning workflows. This is compounded by scaling, infrastructure, and security hurdles. 2. The Taboo Topic: The Technology is not ready While the tech and consultancy world is fast to point to the lack of organizational readiness there is also evidence that the technology may simply not be good enough yet for complex tasks (though few will openly admit this). For example a recent Microsoft stress test on a synthetic marketplace showed high fragility of even the most advanced AI models: • Complexity Crash: Agents performed poorly when faced with realistic competitive environments, complex searches, and numerous results. • Easy to Exploit: They were easily susceptible to manipulation, including fake credentials and prompt-injection attacks, and displayed systemic biases (like favoring the first option). The Takeaway: Current AI agents remain "brittle." Outside of narrow applications like coding assistance, customer service support or narrow office automation, they require close human supervision and are not yet ready for autonomous, high-stakes decisions in unpredictable real-world markets. Adoption will come, but the timeline depends less on breakthrough tech and more on solving deep integration challenges and building truly robust, ethical models. Deloitte Tech Trends 2026 https://lnkd.in/ePcq5h84 MS Marketplace learnings https://lnkd.in/epznYdGB

We know LLMs can substantially improve developer productivity. But the outcomes are not consistent. An extensive research review uncovers specific lessons on how best to use LLMs to amplify developer outcomes. 💡 Leverage LLMs for Improved Productivity. LLMs enable programmers to accomplish tasks faster, with studies reporting up to a 30% reduction in task completion times for routine coding activities. In one study, users completed 20% more tasks using LLM assistance compared to manual coding alone. However, these gains vary based on task complexity and user expertise; for complex tasks, time spent understanding LLM responses can offset productivity improvements. Tailored training can help users maximize these advantages. 🧠 Encourage Prompt Experimentation for Better Outputs. LLMs respond variably to phrasing and context, with studies showing that elaborated prompts led to 50% higher response accuracy compared to single-shot queries. For instance, users who refined prompts by breaking tasks into subtasks achieved superior outputs in 68% of cases. Organizations can build libraries of optimized prompts to standardize and enhance LLM usage across teams. 🔍 Balance LLM Use with Manual Effort. A hybrid approach—blending LLM responses with manual coding—was shown to improve solution quality in 75% of observed cases. For example, users often relied on LLMs to handle repetitive debugging tasks while manually reviewing complex algorithmic code. This strategy not only reduces cognitive load but also helps maintain the accuracy and reliability of final outputs. 📊 Tailor Metrics to Evaluate Human-AI Synergy. Metrics such as task completion rates, error counts, and code review times reveal the tangible impacts of LLMs. Studies found that LLM-assisted teams completed 25% more projects with 40% fewer errors compared to traditional methods. Pre- and post-test evaluations of users' learning showed a 30% improvement in conceptual understanding when LLMs were used effectively, highlighting the need for consistent performance benchmarking. 🚧 Mitigate Risks in LLM Use for Security. LLMs can inadvertently generate insecure code, with 20% of outputs in one study containing vulnerabilities like unchecked user inputs. However, when paired with automated code review tools, error rates dropped by 35%. To reduce risks, developers should combine LLMs with rigorous testing protocols and ensure their prompts explicitly address security considerations. 💡 Rethink Learning with LLMs. While LLMs improved learning outcomes in tasks requiring code comprehension by 32%, they sometimes hindered manual coding skill development, as seen in studies where post-LLM groups performed worse in syntax-based assessments. Educators can mitigate this by integrating LLMs into assignments that focus on problem-solving while requiring manual coding for foundational skills, ensuring balanced learning trajectories. Link to paper in comments.