There is a sentence I now hear almost every around me steering committee, vendor pitch, and board deck: "We need to put AI into everything." It sounds decisive. It sounds like leadership. It is usually the moment I start asking quieter questions — because "AI everything" is rarely a strategy. It's a symptom of a market that has collapsed three very different technologies into one marketing word.
At the frontier where humans, AI, and technology foundations converge, the first job isn't to chase the frontier — it's to know which side of it you're actually standing on. Most organisations chasing "AI everything" haven't done that basic sorting exercise. They're buying a narrative, not a capability.
Not everything is AI. Some of it is just good technology, finally applied properly.
Not everything is AI
Let's start with the uncomfortable truth behind that line: a meaningful share of what gets pitched as "AI transformation" is just good technology, applied properly, for the first time. Workflow orchestration. Better data pipelines. Event-driven architecture replacing batch jobs. Robotic process automation stitching together systems that should have been integrated a decade ago. None of that is intelligence. It's plumbing — and plumbing, done well, delivers enormous value. But calling it AI doesn't make the outcome bigger; it just makes the budget conversation more confusing, because now finance expects "AI-scale" returns from what is, honestly, a workflow re-engineering project.
I've sat across the table from teams proposing "AI-powered" loan processing that, on inspection, was a rules engine and a straight-through-processing workflow — genuinely valuable, genuinely overdue, and not AI in any meaningful sense. The danger isn't the mislabeling itself. It's that leadership starts benchmarking a deterministic automation project against the expectations set by generative AI hype, and then wonders why the "AI initiative" feels underwhelming when it hits exactly the ROI a good BPM project should hit.
The real split: ML, GenAI, and workflow technology
It helps to separate three things that keep getting mashed into one slide:
Machine learning (predictive/classical) — models trained on your own structured data to predict, score, or classify: credit risk scoring, fraud detection, next-best-offer, churn prediction, collections prioritisation. This is mature technology. Banks have run ML in production for over a decade. The frontier here isn't "can we do it" — it's explainability, model risk governance, and drift monitoring at scale. A fraud model that silently degrades because merchant behaviour shifted is a governance failure, not an AI failure.
Generative AI (LLM-based) — models that produce language, code, or synthesis from unstructured input: summarising a 40-page credit memo, drafting a first-pass KYC narrative, turning a customer complaint thread into a structured case summary, powering a copilot that helps a relationship manager draft a pitch. This is genuinely new capability. It doesn't predict a number — it compresses and generates language and reasoning at a cost and speed that wasn't previously possible.
Workflow and integration technology — orchestration engines, RPA, BPMN, event streaming, API gateways. This is the connective tissue. It's what makes ML and GenAI outputs actually reach a human, a system of record, or a customer, reliably and auditably.
The mistake I see repeatedly is treating these as interchangeable, or worse, treating the third category as if it were the first two. A lending origination platform that automates document routing and status updates is solving a real problem — but if it's pitched internally as "our AI transformation," you've spent your one shot at executive attention on something that will look thin next to what a genuine GenAI use case can do.
Where GenAI actually earns its place
The clearest GenAI wins I've seen in banking and fintech share a pattern: they sit at the unstructured edge of a process, not the decisioning core. A commercial lending team using an LLM to draft the first pass of a credit narrative from financials and covenants — still reviewed and owned by a human underwriter — cuts drafting time meaningfully without touching the actual credit decision. A contact centre summarising call transcripts into structured case notes removes an hour of admin per agent per day without ever deciding a customer outcome. In both cases, the AI is doing the thing language models are actually good at: compressing and restructuring unstructured information for a human who still holds accountability.
Contrast that with the failure pattern: organisations trying to make an LLM decide — approve a loan, set a credit limit, resolve a dispute — because it feels like the more ambitious use of the technology. It's the wrong tool for that job. Decisioning at that level wants a well-governed ML model or a deterministic rules engine, with GenAI, at most, explaining the decision in plain language afterward. The moment you ask a generative model to be the decision-maker instead of the communicator, you've traded a governable model risk framework for a probabilistic text generator with no calibrated confidence score — and in a regulated industry, that's not innovation, that's exposure.
GenAI vs. existing workflow technology
There's a second confusion worth naming: GenAI replacing perfectly good deterministic workflow technology. If a process is stable, rules-based, and high-volume — payment reconciliation, standard KYC checks against known fields, document classification into fixed categories — a well-built rules engine or classical ML classifier will outperform an LLM on cost, latency, consistency, and auditability every time. GenAI is not a universal solvent. Where the process is variable, the inputs are messy free text, and judgement or synthesis is genuinely required, that's where it belongs. Use it where structure doesn't exist yet, not where structure already works.
What "AI everything" should actually mean
I'm not arguing against ambition — I'm arguing for precision, because precision is what turns ambition into delivery. "AI everything" should mean: know, use-case by use-case, whether you need prediction, generation, or orchestration — and build the right thing for the right job, with the right governance wrapped around it. That's a harder sentence to put on a slide. It's also the only version of the sentence that survives contact with a steering committee asking where the return actually came from.
The organisations getting real value aren't the ones with the loudest "AI-first" language. They're the ones who quietly did the sorting exercise — decomposed their "AI transformation" into a predictive layer, a generative layer, and an orchestration layer — and funded each on its own merits, with its own success metric, instead of one undifferentiated AI budget line chasing one undifferentiated AI outcome.
That's the frontier worth navigating. Not "more AI." The right AI, in the right place, doing the thing it's actually good at.