Start here
The wrong vocabulary buys the wrong tools
Most professional services leaders have used ChatGPT. Fewer can explain the difference between machine learning, generative AI, and an AI agent without reaching for a vendor’s slide deck. That gap matters, because the wrong vocabulary leads to the wrong tools, and the wrong tools waste months.
The villain here is AI hype. Vendor decks, conference keynotes, and LinkedIn feeds treat “AI” as a monolithic thing — one technology, one purchase decision, one transformation. It is not. AI is a broad field, and buying a tool without understanding which layer of that field you are addressing is how firms end up with expensive subscriptions, sceptical teams, and nothing to show the board. Before you brief a vendor, challenge a proposal, or commission a pilot, these distinctions are worth knowing.
How the terms nest
AI, machine learning, and LLMs
- Artificial intelligence is the umbrella — any system that can sense, reason, act, and adapt. It has been in commercial use for decades. The camera at a motorway toll booth reading your number plate is AI. The algorithm flagging suspicious insurance claims is AI. None of that required ChatGPT.
- Machine learning sits inside AI: algorithms that uncover patterns in data with minimal manual programming. Arthur Samuel coined the definition in 1959 — giving computers the ability to learn without being explicitly programmed. If you have historical data where you know what success looks like, such as won proposals versus lost ones, machine learning can build a model that spots the pattern in new cases.
- Generative AI is a subset designed to produce new content: text, images, audio, video, code. Large language models (LLMs) sit inside generative AI — prediction engines for text, trained on vast language data, capable of understanding context and nuance in ways earlier systems could not. They power ChatGPT, Claude, and similar tools. They are impressive, and they are also one specific type of AI, not a synonym for all of it.
Knowing where a tool fits in this hierarchy is the first step to picking the right one for a given task.
Where the returns are
What AI is good at, and what it is not
AI systems can sense (read and interpret inputs), reason (plan and evaluate options), act (produce outputs or trigger actions), and adapt (improve through feedback). Those four capabilities, applied to the right workflows, are where the productivity gains come from.
Firms that apply AI well to decision support and process automation routinely achieve productivity improvements of 70 to 80 percent on those specific workflows. That is not a claim about the whole business — it is about repetitive, pattern-based tasks where the input data is reasonably clean and the output is definable.
The practical question is not “where can AI help us?” but “which tasks in our business involve making consistent decisions based on patterns in data?” Customer segmentation, document routing, proposal drafting from templates, meeting summarisation, compliance flag detection — these are the high-return starting points.
Human intelligence still wins at ambiguity, creative judgement, ethics, and client relationships. AI handles the repeatable, volume-based work; your senior team applies judgement where it actually counts.
Available via an API call
Deep learning and generative AI, for your firm
Deep learning — neural networks with many layers — is the foundation beneath the models large technology firms provide. You do not build or manage this. What matters is that you have access to its outputs through foundation models from providers like Anthropic, OpenAI, and Google.
Practically, capabilities that once required significant infrastructure investment are now available via an API call: call-centre transcription, document classification, natural-language search across your knowledge base. You are not building the AI; you are connecting your workflows to it. Generative AI extends this further — image generation, code generation, draft documents, structured data extraction from unstructured text were multi-month projects five years ago. They are now accessible to a 50-person consultancy on a modest monthly budget.
Junior employees, not executives
What an AI agent really is
The term “AI agent” gets used loosely. In practice, an AI agent is an LLM equipped with tools — the ability to call APIs, search databases, execute code, and chain a series of tasks together to hit a goal. If an LLM is a knowledgeable individual at a desk waiting for questions, an agent is that same individual with access to your CRM, your document library, your calendar, and permission to take actions within defined limits.
Treat AI agents as junior employees. They have strong recall and follow a process reliably, but every session starts fresh — they carry no memory between interactions unless you build that infrastructure. They need clear instructions, defined access, and a human reviewing anything consequential. You would not put a junior hire in charge of a client deliverable without a review step, and you would not deploy an agent into a critical workflow without oversight.
This is not a reason to avoid them. Junior employees at low cost, available around the clock, completing defined tasks with consistent quality is a significant operational advantage. It is simply a reason to deploy them with the same care you apply to onboarding anyone new.
This is the KWA framework: Knowledge (your institutional knowledge, centralised and AI-accessible), Workflow (your processes standardised and documented), and Agents (AI operating within those workflows with appropriate oversight). Most firms that struggle with adoption are attempting the Agents layer without having built the Knowledge and Workflow foundations. Sequence matters.
Reliable output, not noise
Context, prompts, and grounding
Three concepts determine whether your deployments produce reliable results or expensive noise:
- Context is the working memory of an LLM. Everything the model knows during a session — your prompt, the conversation history, any documents you provide — lives in the context window. Managing it well reduces cost and improves accuracy. Providing irrelevant or excessive context is one of the most common reasons outputs miss the mark.
- Prompts are your instructions. A weak prompt produces a generic output; a well-structured one, with a clear goal, specified audience, defined constraints, and relevant examples, produces something useful. The discipline of writing and refining prompts systematically — prompt engineering, or context engineering — pays for itself quickly in reduced rework.
- Retrieval-Augmented Generation (RAG) connects an agent to your actual data rather than embedding everything at training time. When an agent needs information about a client, it calls your CRM; when it needs a precedent, it searches your knowledge base. This grounds outputs in your real data, reduces hallucinations, and means you only pay for the information you need per query.
The risk you cannot ignore
Hallucination is structural, not a bug
LLMs produce confident, well-structured, fluent text. They also sometimes fabricate facts, invent citations, and misstate figures — with equal confidence. This is not a bug that will be patched away; it is a structural characteristic of how these models work.
For a professional services firm, an unverified AI output used in a client deliverable, a compliance document, or a board report creates real risk — reputational, legal, and commercial. The mitigations are straightforward. Ground the model in verified data wherever possible (RAG addresses much of this). Build human review into any workflow where the output informs a decision or goes to a client. For critical processes, consider connecting two models so one checks the other’s output. The junior employee analogy applies again: you review junior work before it leaves the building.
Closing the gap
Customising AI for your business
Foundation models are trained on publicly available data. Your business data — client records, delivery templates, institutional knowledge, proprietary methodologies — is not in there. Two approaches close that gap:
- Fine-tuning trains a model further on your own data so it becomes familiar with your terminology, clients, and processes. It reduces the length and complexity of prompts needed and can significantly improve accuracy on specialised tasks. The analogy is the onboarding investment you make in a new hire: time and cost up front, productivity gain thereafter.
- RAG connects the model to live data at query time rather than embedding it during training. It is more flexible, typically less expensive per query, and keeps responses grounded in current information. For most professional services workflows, RAG is the right starting point.
Sequence over speed
Applying the 5 Steps for AI Leadership
The 5 Steps framework describes the sequence firms need to follow to make adoption stick:
- Align — a written AI position your team and clients can act on.
- Activate — your team actually using AI tools in daily work.
- Amplify — wins captured and spread firm-wide, not siloed with one person.
- Accelerate — moving quickly from idea to deployed workflow.
- Govern — approved tools, data-handling rules, and a risk register in place.
Most firms that have “tried AI and seen nothing stick” have skipped Step 1. There is no written position, so the team does not know what is approved, what the boundaries are, or why it matters to strategy. Without Align, everything else is individual experimentation that does not compound. The other common failure is jumping to Accelerate without Govern — deploying workflows at speed without governance is how you create the compliance conversation, not avoid it.
What this means for your firm
Right tool, right job, foundations first
AI is not one thing. It is a set of technologies at different maturity levels, with different appropriate applications, requiring different infrastructure to deploy reliably. The firms that pull ahead are not necessarily those with the largest AI budgets. They are the ones that pick the right tool for the right job, build the workflow foundations before deploying agents, and maintain human oversight where it matters.
The risk is not moving too slowly. The risk is spending twelve months on demos, vendor pilots, and unfocused experimentation while competitors quietly ship one workflow, then another, and quietly extend their capacity. Make AI work so your team can deliver.