Article by James Flint
A couple of weeks ago I gave a presentation to members of the Jamaican AI Association (JAIA). Unfortunately, I didn’t get to go to Jamaica to do this; I gave the presentation via a Zoom call from wintry London. I blame the internet.
In one of my slides (see picture) I distinguished between horizontal and vertical deployment of AI. A horizontal deployment is relatively straightforward: it’s giving your employees access to (or permission to access) a software-as-a-service (SaaS) AI system like ChatGPT or Copilot, or a coding tool like Cursor which performs individual actions as instructed which close human oversight, even when it acts as an agent (for example by searching the web for information).
A vertical deployment is a deeper integration of a more focussed AI toolset – which could be a remote service connected by API, or might sit on the company’s own private cloud – which fully or partially automates a workflow that was previously carried out by humans or non-AI software, and does it largely unsupervised, activity that more closely corresponds with the term “agentic AI”. This might involve monitoring digital infrastructure to detect behavioural anomalies, or summarising research trends, or assessing supplier risk profiles, or reviewing incoming CVs for a job application, or forecasting customer demand... any one of a range of workflow tasks that were hitherto too complex to be handled by traditional process automation, but are within the possible scope of AI.
The distinction between the two types is not rigid, and some services – like a customer service chatbot, for example – might happily sit at the junction of both categories. But I think it’s a useful distinction nonetheless, as a vertical deployment is much, much more challenging than a horizontal one.
A horizontal deployment, while not without its challenges, is still relatively straightforward. Rollout is essentially a matter of doing supplier due diligence and appropriate compliance work, giving employees access to the service, and putting – human – guardrails or guidelines in place, which are then enforced (or at least encouraged) through policies and training.
To get a vertical deployment working, however, you are likely to have to make significant changes to your digital infrastructure that allow you to transform your company’s data resources – traditionally seen as a risk factor – into an asset than can be leveraged by the agentic system.
As accurate data is essential for an AI to make reliable predictions and analyses, and as these decisions are now going to impact automated decision-making processes, maintaining up-to-date data is crucial for generating actionable insights and enhancing operational efficiency. This may in turn mean overhauling your data governance provision so that it becomes a proactive function, monitoring and managing data at every stage in its collection and use ensuring its integrity, availability and confidentiality.
This is a lot to think about before you even get onto the system implementation and security, and goes a long way to explaining why Deloitte’s 2025 Emerging Technology Trends study found that while 30% of the 500 US organised surveyed have been exploring agentic options, only 11% are actively using them in production.
As the report explains, data isn’t always interoperable between systems, which limits the ability to blend data from different sources - 48% of surveyed companies identified the searchability of data as a challenge, when a systematic approach to ingesting, organising and storing data is required if you’re to make it available to modern AI tools.
"You have to have the investments in your core systems, enterprise software, legacy systems, SAS, to have services to consume and be able to actually get any kind of work done,” explains Deloitte CTO Bill Briggs, in an interview with ZDNet, “because, at the end of the day, [the AI systems are] still calling the same order systems, pricing systems, finance systems, HR systems, behind the scenes, and most organizations haven't spent to have the hygiene to have them ready to participate."
On top of that, organizations often fail to create the proper governance and oversight mechanisms for the agentic systems to operate autonomously, as traditional IT governance doesn't account for AI agents' ability to make their own decisions.
"You've got this layer on top, which is the orchestration/agent ops,” Briggs says. “[The question is] how do we instrument, measure, put in controls, and thresholds, so if we got it right, the meter wouldn't be spinning out of control, kind of like we saw with the early days of cloud adoption?"
The point is, that if you want to use vertically-integrated AI to transform your organisation’s data from a passive risk into an active asset, you need to start from the data architecture and the data governance and work outwards to the AI, rather than the other way around.
If you are looking and either horizontal or vertical deployment of AI and would like to understand how best to create the data architecture and data governance framework for safe deployment, we would be delighted to help. Click here to get in touch.
