Most AI programmes that succeed start the same way — with the right question. And most that fail skip it entirely.
The question is deceptively simple: where does AI actually create value, aligned to what this organisation is trying to achieve?
Instead, most companies jump straight to copilots, agents, model selection, and governance committees before anyone has agreed on what problem they're solving. Six months later, they have pilots in twelve business units, adoption in none, and a board asking what the CHF 3M bought.
The fix isn't a better vendor or a larger AI team. It's a clearer framework for where to invest and in what sequence.
Two axes, not one programme
At Auron Advisory we think about AI adoption on two axes. Understanding the distinction is the first step to sequencing correctly.
Empowers every employee using tools the company already licenses. The goal is broad productivity — not deep transformation.
Embeds AI into 2–3 high-impact core workflows where the organisation can build a durable edge.
The pattern I consistently see: organisations overinvest in Vertical AI experimentation before they've done Horizontal AI well — then wonder why adoption is flat. The sequence matters. Horizontal AI builds the muscle; Vertical AI builds the advantage.
Horizontal AI: activate what you already own
Most large organisations are already paying for Microsoft Copilot, Google Gemini, Claude, Salesforce Einstein, or SAP Joule. The fastest ROI in enterprise AI is not a new model — it's activating the licences you're already paying for.
The vehicle that makes this work at scale is a prompt library: a curated, co-created set of prompts that bridges the gap between a tool sitting in the toolbar and an employee actually using it. One prompt champion per function. Deployed in week one. Built with the teams, not for them.
Four prompt categories that work across every department immediately:
Beyond the general library, the highest-value prompts are function-specific. Finance needs variance narratives, board commentary, and audit responses. Procurement needs RFP generation and contract redlines. Customer service needs ticket responses and escalation summaries. Each function gets its own prompt set, built with that team's language and context — not copy-pasted from a vendor's template library.
Vertical AI: where competitive advantage is built
Horizontal AI is a productivity play. Vertical AI is a strategy play. It's where you embed AI deeply enough into a core workflow that the capability becomes a structural advantage — one that compounds over time and is difficult to replicate.
But Vertical AI only works if you choose the right use cases. And most organisations don't have a systematic way to make that choice. They pick the use case that was loudest in the last leadership offsite, or the one the vendor demoed most compellingly, or the one the CTO is personally excited about.
A better way is a structured three-step process:
The 5 scoring criteria
Not all AI use cases are equal. Before committing resources, every candidate goes through this scoring framework. The weights reflect the reality of delivering AI in a regulated financial services environment:
| Criterion | What we look at | Weight |
|---|---|---|
| Business value | Cost savings, FTE hours recovered, error reduction | 30% |
| Technical feasibility | Data quality, SAP/Salesforce integration, build vs buy | 25% |
| Regulatory risk | FINMA, EU AI Act, Swiss nDSG — eliminative gate | 25% |
| Organisational readiness | Named sponsor, team capacity, change appetite | 10% |
| Time to value | Weeks to pilot, dependencies, quick win potential | 10% |
The regulatory criterion is an eliminative gate, not just a score. A use case that scores 100 on every other dimension but creates material FINMA or EU AI Act exposure does not make the shortlist. Building that discipline into the scoring process — rather than treating compliance as a later-stage concern — is one of the highest-leverage things a programme can do in weeks 1–3.
"Organisations often overinvest in experimentation and underinvest in adoption discipline. The framework only works when someone is accountable for making it land."
The framework is not complicated. Most organisations already have the data they need to run it. What they're missing is the discipline to do it before the pressure to show something — anything — becomes overwhelming.
Start with the right question. Sequence the axes correctly. Score use cases on what matters. The delivery work becomes significantly more tractable when the first three weeks are spent this way rather than in model selection workshops.