Agentic AI in the enterprise: where to start
Data and limits first, then the agent: the order I suggest.
Many conversations about AI in a company start with “we want an agent.” Building the agent is usually the easiest part of the project. Whether it works depends on something else: reliable data and clear rules about what it can do on its own.
This note describes the order I follow when working with agentic tools in analytics and development teams.
Automation, assistant and agent
It helps to separate three things that sometimes get the same name:
- Classic automation: fixed rules over predictable inputs, such as a script that consolidates five spreadsheets and produces a report. It is cheap, deterministic and easy to audit. If that solves the problem, an agent is unnecessary.
- Assistant (chat): answers questions, drafts and summarizes. It is useful, but it does not run processes or own an outcome.
- Agent: receives a goal, plans the steps, uses tools, checks what it did and corrects it, with defined permissions and a record of every action. For example: it calculates a benefit, generates the document, sends it for approval and flags anything that does not add up.
The problem with starting from the agent
The most common failure I have seen is building or buying the agent first and then finding out that the data it needs sits in seven files with three different versions. An agent working on dirty data repeats the same mistake many times, very fast.
That is why I suggest building in this order:
Governed data platform
A single source of truth, DEV, QA and PROD environments, and defined access.
Guardrails and traceability
What the agent can do, what it must escalate to a person and what gets logged.
Agents
Narrow use cases, with a person in the loop.
Layer 1 · The platform
Before an agent touches anything, you need to know where the master data lives, who can read what and how a change moves from development to production without breaking what already works. In my experience that is solved with platforms such as Microsoft Fabric or Azure, with orchestrated pipelines and separate DEV, QA and PROD environments. With that in place, the agent works on reliable data and can be switched off without affecting anything else.
Layer 2 · The guardrails
Before writing code, put in writing which actions the agent can run on its own, which need a person’s approval and what evidence each step leaves. An agent that generates benefit letters can draft and pre-calculate on its own, but a person approves sending it. If you cannot reconstruct why the agent did something, it is not ready for production yet.
Layer 3 · The first use case
A good first use case meets four criteria:
- Repetitive and frequent: the savings add up every week.
- Known rules: there is a procedure a person can explain.
- Bounded, visible cost of error: if it fails, you notice quickly and fixing it is cheap.
- Human in the loop: someone approves the outcome while trust builds.
An example of this kind of flow: a retirement plan process that combines a benefits calculator, the automatic generation of the documents and an approval app with one screen per role (manager, employee and labor attorney). Each piece is simple. What changes is that the whole process no longer depends on emails and shared folders.
Readiness checklist
| Question | If the answer is no… |
|---|---|
| Is there a single, governed source for the data the agent will use? | Start with the platform. |
| Can you write on one page what it can do alone and what it must escalate? | The guardrails are missing, and finding them in production is expensive. |
| Does the target process have an owner and a documented procedure? | The agent will inherit the ambiguity of the process. |
| Is there a way to measure before and after, in hours or money? | It will be hard to justify the investment and to detect regressions. |
| Can the team operate it without the person who built it? | Include documentation and handover in the scope. |
What to build first
I almost always recommend the same thing: a single case, end to end, in six weeks or less. That includes the minimum platform for that case, the guardrails in writing, the agent with a person approving and a before-and-after measurement. That first case leaves the infrastructure in place, and the second and third cost considerably less.