AI Edutainment is currently a self-funded business. Our revenue comes from companies that failed in three of four AI implementations and can no longer afford McKinsey and its cousins. In this blog, I summarise the basics of AI implementation and map the AI landscape across industries and business functions. It took me a year to nail this picture.
Let’s start from the beginning.
Before implementing an AI or ML solution
Implementing AI in any company starts with clarity: leaders must define the problem they are trying to solve, the decision they want to improve, or the workflow they want to reimagine, before choosing models or vendors. Once the use case is clear, the next step is to review which data is needed, where it comes from, what systems capture it today (ERP, CRM, sensors, ticketing tools, learning platforms), and whether that data is accurate, representative, and complete enough to avoid “garbage in, garbage out.” This naturally leads to data governance: companies need clear ownership, access rules, retention policies, quality checks, and privacy safeguards, ideally anchored in recognised frameworks such as DAMA-DMBOK, NIST, or ISO 27001/27701, so that AI initiatives don’t create hidden compliance and security risks.
On top of that foundation sits vendor and solution management: organisations must vet AI providers for security, transparency, data handling, model update practices, and exit options, and ensure contracts reflect their governance standards rather than simply accepting “black box” services. Any new AI application has to blend with the company’s existing cybersecurity posture—integrating with SIEMs, identity systems, and monitoring tools—so that it strengthens defences instead of introducing new attack surfaces or blind spots. Finally, if the AI solution augments or reshapes human work, leaders must revisit people policies and practices: job design, training, performance metrics, escalation paths, and change management all need updating so that employees understand how to work with AI, where human judgment remains decisive, and how accountability is shared between people and machines.
The power of clean workflows
Choosing the “right” AI solution only makes sense after you understand and clean up the work it will touch. Workflow definition comes first: map the current process end-to-end, including who does what, in which system, with which inputs and outputs, where decisions are made, and how exceptions are handled; techniques such as process mapping, task mining, and process mining help reveal how work actually flows versus how it is documented on paper. Best practice is to run workshops with frontline experts, draw the real workflow (including ugly workarounds), identify decision points and success metrics, and only then decide where AI could reliably take over or assist specific steps.
Once workflows are mapped, you can see which data each step depends on and start serious data cleansing: auditing all relevant data sources, standardising formats, handling missing values, removing duplicates, harmonising inconsistent labels, and detecting outliers or obvious errors so that AI is trained and run on trustworthy inputs. Organisations that invest early in data quality, shared definitions, and basic observability of their data pipelines are better prepared to plug in future AI tools, because their systems expose clean, well-understood events and metrics instead of noisy, fragmented logs that confuse models and humans alike.
Once you understand the problem and have mapped the workflows you want to augment, transform, or enhance with AI, you can then explore the full range of options. Those options should be guided by the scale of the problem, your budget, and your internal expertise—not by whatever a consultant or vendor happens to be selling. AI is a powerful enabler, but it cannot “magic away” a poorly framed problem or a broken process.
AI Implementation across businesses and business functions: From off-the-shelf solutions to domain ontologies and digital twins
Pre-built/ off-the-shelf tools
Many of the most useful AIs for companies today come “off the shelf”: they are pre-built tools embedded in SaaS products that solve common problems without requiring a data science team or heavy customisation. These include things like AI writing assistants for email and content, chatbots for customer support, invoice and expense automation in finance, AI-powered search across documents, and predictive features built into CRM or analytics platforms. Because they are designed for generic business workflows, they can often be configured with a few settings, connected to your existing tools, and rolled out within days rather than months, giving teams quick wins while you are still building your broader AI strategy.
The trade-off is that off-the-shelf AI is less tailored to your unique processes or data, but for many functions—marketing, sales, HR, customer service, basic operations—it is “good enough” to remove friction and free people’s time at a fraction of the cost of a custom solution. A growing number of platforms also offer no-code ways to orchestrate these AIs—letting non-technical teams combine chatbots, summarisation, document classification, and simple predictions into lightweight workflows without writing code. For most organisations starting out, these productized AIs are the fastest way to learn what works, build internal literacy, and identify where, later, a deeper custom build would actually create a competitive advantage.
Customised digital solutions with embedded AI components
Beyond off-the-shelf tools, many companies rely on customised digital solutions where IoT, sensors, and embedded machine learning are tightly woven into core operations. These systems turn machines, vehicles, buildings, and products into smart endpoints: edge devices that collect data, run local models for anomaly detection or predictive maintenance, and act in real time without waiting for the cloud. Digitisation specialists and system integrators who once delivered “classic” IoT—dashboards, alerts, and remote monitoring—are now upgrading their offerings with AI components such as on-device analytics, pattern recognition, and autonomous control loops, often using edge AI platforms and TinyML techniques to fit models into constrained hardware.
In practice, this means the partners who design a factory’s sensor network or a logistics company’s tracking platform increasingly propose AI-enhanced modules: models that predict equipment failure from vibration data, optimise routing from live telemetry, or adjust building systems based on occupancy patterns, all embedded into the existing IoT architecture. The value shifts from “seeing what happened” on a dashboard to acting on what’s about to happen, with providers offering lifecycle services—model deployment, monitoring, and retraining across thousands of devices—to keep these embedded ML solutions accurate and secure over time.
Customised LLMs
Customised LLMs are emerging as specialist assistants across many industries, designed to sit inside existing tools and workflows rather than replace them. They help humans navigate complex systems, reconcile different technology standards, and interpret large volumes of data—especially in environments where demand is growing but headcount cannot. In practice, this means a single LLM-based interface can sit on top of heterogeneous systems, logs, and formats, answering questions, generating actions, and orchestrating workflows so that engineers, operators, and analysts spend less time “speaking many machine dialects” and more time making decisions that matter.
Let’s review customised LLMs on an example from an industry I know very well – mobile network operators.
Customised large language models are starting to sit “inside” network operations, and Aira’s RANGPT for Radio Access Network (RAN) management is a good example of what this looks like in practice. RANGPT is an LLM-based utility that lets mobile network operators query and control the RAN in natural language, turning complex KPI data and configuration parameters into conversational insights and suggested actions for engineers and operations teams.
Under the hood, the model is adapted to telecom language, RAN data structures, and operational workflows, so it can understand questions about coverage, interference, throughput, and energy consumption, analyse data in the operator’s data lake, and return explanations, charts, and concrete optimisation commands rather than generic text. Operators can chain queries and control instructions into automation routines, allowing the LLM to help diagnose performance issues, recommend parameter changes, generate code for new xApps/rApps, and even drive RAN automation scenarios that reduce opex and energy use while keeping humans in control for approval and oversight.
More broadly, telecom-specific LLMs follow a similar pattern: they are fine-tuned or adapted on standards documents, logs, alarms, and procedures, and then embedded into OSS/BSS and service management platforms, where they assist with troubleshooting, ticket triage, configuration planning, and intent-based optimisation. For operations teams, this means moving from navigating many siloed tools and scripts to working with a domain-trained assistant that understands their environment, speaks their jargon, and can both explain what is happening in the network and help orchestrate safe, auditable changes at scale.
Customised LLMs with domain ontologies/ AI agents as new employees
Customised LLMs become truly powerful when they are grounded in domain ontologies—structured maps of the entities, relationships, and workflows that define a specific field. These ontologies give the model a shared “mental model” of the business: what objects exist (accounts, opportunities, proof-of-concept projects), how they relate, and which actions or states matter, which in turn makes the AI’s reasoning more consistent, less hallucination-prone, and far better suited to act as an agent that can take multi-step actions rather than just chat. When you connect such an ontology-driven LLM to operational systems (CRM, ticketing, configuration tools) and give it tools and memory, you get agentic solutions that can understand context, plan tasks, call the right APIs, and keep track of state across an entire presales cycle, incident, or project.
Vivun is a concrete example of this approach in the world of presales. Its “Agent Intelligence” and SE Copilot are built on a proprietary domain model informed by tens of millions of presales activities and decades of expert knowledge, capturing how technical sellers qualify opportunities, run demos, track product gaps, and influence revenue. Instead of a generic chatbot, Vivun delivers an AI teammate that speaks the language of presales, aligns its outputs with the Vivun Intelligence System’s ontology, and can take on real work: drafting technical win stories, summarising discovery calls, highlighting product risks, and feeding structured intelligence back into product and revenue systems. I had the chance to explore this in depth with Vivun’s Co-Founder and Chief AI Officer, Joe Miller, in Episode 60 of AI Snacks.
Digital Twins
A digital twin is a virtual replica of a physical system that stays in sync with reality through live data, so teams can monitor, simulate, and optimise operations without touching the real asset. In an enterprise, that can mean a twin of a single machine, a production line, an entire factory, or parts of a supply chain, always fed by IoT sensors, control systems, and business data to mirror how things actually behave over time.
AI technologies make these twins useful rather than just visual. Machine learning models ingest real-time and historical data from the twin to detect anomalies, predict failures, simulate “what if” scenarios, and recommend or even trigger actions—for example, predicting when a critical pump is likely to fail, or how a layout change will affect throughput and energy use. This turns the digital twin into a decision laboratory: AI can run thousands of virtual experiments inside the twin, so humans can choose better interventions in the real world with lower risk and cost.
Despite the hype, true end-to-end enterprise digital twins—covering every process, every site, and the full supply chain in one coherent model—still don’t exist in practice, mostly because of data integration, interoperability, and governance challenges. What we do see are strong local implementations, such as factory-level twins that mirror a single plant or production area and deliver concrete value in planning, operations, and maintenance.
For example, BMW uses digital twin factories to simulate and optimise layouts, material flows, and production scenarios before making physical changes, improving efficiency and supporting sustainability goals. BASF’s Ludwigshafen site has long used a digital twin platform that connects CAD, BIM, ERP, and workforce systems to mirror day-to-day factory reality and support faster, better-aligned decisions for thousands of employees. Solutions like NavVis and Matterport provide spatial twins of factory floors that allow planners and managers to “walk” a site remotely, compare layouts, and plan new lines without travel or downtime, especially when linked to sensor data for live conditions. In many smart factories, AI-enabled digital twins are now central to predictive maintenance—continuously monitoring equipment behaviour, forecasting failures, and optimising maintenance timing and energy use, even if they only cover one plant or one cluster of assets rather than the entire enterprise.
Midsized companies are increasingly using digital twins as a pragmatic way to “bottle” operational know-how and keep plants running smoothly, even as experienced workers retire. In manufacturing, for example, firms build twins of critical lines or machines that combine 3D layouts, sensor data, and control logic, so maintenance and process engineers can see real-time performance, test parameter changes, and plan interventions without physically standing next to the equipment. This makes it possible for a smaller expert team to support multiple sites remotely: a senior technician can log into the twin, diagnose an issue, and guide on-site staff through a fix using shared visual and data context, rather than travelling or relying on vague phone descriptions.
Digital twins also help soften the blow of a “grey wave” of retirements by capturing institutional knowledge—how seasoned operators interpret sounds, vibration patterns, or subtle process deviations—as part of the twin, and making those insights available through dashboards, playbooks, and AI-driven recommendations. For newer employees, twins of factory floors, warehouses, or complex equipment can serve as immersive training grounds: they can “walk” a 3D replica of the site, practice procedures, and run through safety scenarios virtually before touching real machinery, speeding up onboarding and reducing risk. In this way, digital twins become both a remote maintenance cockpit and a living training simulator, helping midsized firms maintain resilience and quality even when headcount is tight and expert experience is scarce.