Insight: Artificial Intelligence and Automation
Getting it working is not the hard part. Deciding what it should do is.
Artificial intelligence and automation are two branches of the same tree, and both depend on understanding the process underneath before anything is introduced. The term itself covers several very different capabilities, and choosing between them is not possible until the intended result is stated. Nearly every stalled adoption I have seen failed at that point, not at the technology.
Illustrative rather than tied to a named client. Access and permissions are covered in more depth in the data governance insight, which is a prerequisite for most of this work.
The Distinction
Two branches, one foundation
Both rest on the same underlying language models. What separates them is whether the system is producing something for a person to use, or acting on its own to complete a task.
Branch One
Generative
It produces; a person decides
Content generation: text, images, video, code. You ask, it responds. The quality of the response depends on two things: what the model was trained on, and how well the request was framed. That framing is a skill, and it is the difference between a useful answer and a generic one.
Branch Two
Agentic
It acts; a person defines the boundaries
Where autonomy becomes possible. An agentic system uses the same foundation to break a requirement into steps and carry them out, moving information between systems, completing a process, or responding to an event. It needs specific direction about the result it is working toward.
What both need
Resources
Intelligence is bounded by access
A model working against the open web is bounded by its training and what it can find there. One working against your business is bounded by what you have given it: the templates, the records, the historical material. Access determines usefulness more than the model does.
The question that is rarely asked first
Requests for help adopting artificial intelligence arrive constantly, and almost none of them arrive with an intended result attached. It has become the phrase of the decade, so it is assumed to simply work once introduced. It does not. To generate content, the generative branch is the answer. To perform tasks, the agentic branch is. Both require a defined outcome and a defined scope, and neither can be selected before those exist.
The Method
What every engagement needs before it starts
These six questions apply whichever branch is being used. Answered properly, the implementation is usually straightforward. Left unanswered, no amount of capability rescues it.
Every engagement needs a defined outcome.
Without one, the capability has nothing to aim at, and the project becomes an experiment nobody can judge.
In Practice
Two requests, two entirely different builds
Both are reasonable asks. They share almost nothing in terms of what has to be established beforehand.
“Generate a quote for a customer”
A generative requirement. Which means: is there a repository of pricing and product information it can draw on, and is that information current? Is there a template it should build against, so output is consistent with everything else the business sends? What must never be varied, such as terms, disclaimers or approval thresholds? And who checks the result before it reaches the customer? Without the repository and the template, the model will produce something plausible and wrong.
“Move this data between systems automatically”
An agentic requirement, and a different set of questions entirely. What must it reach, and with what permissions? Where in the environment does it run? What are the defined inputs and outputs, and what does a successful run look like against a failed one? What happens when the source is unavailable or the data is malformed? Here the constraint is rarely the intelligence. It is the access and the error handling.
Where this meets the rest of the estate
An agentic system operates with the permissions of whoever runs it, and reads faster and further than any person would. That makes access correctness a precondition rather than a later consideration. A permission nobody noticed for five years becomes significant when something can traverse everything reachable in seconds. For most organisations, sound data governance is the work that makes artificial intelligence adoption safe rather than merely possible.
The Point
Start with the result, not the technology
Introducing artificial intelligence into a business is not especially difficult. The platforms are capable, the integration paths exist, and the cost of entry is lower than it has ever been. What is difficult is stating precisely what should be different once it is working, and that is a business question rather than a technical one.
Answer it properly and the rest follows: the branch, the resources, the access, the cost and the return all become determinable. Skip it, and the result is a capable system producing output nobody asked for. The outcome is the work. Everything after it is execution.