Which business process should you automate first—and does it need AI?

The best first process to automate is a frequent, well-understood task with a clear trigger, a defined result, accessible data and a person who can handle exceptions. Start with a narrow workflow you can measure against its current manual version. Use ordinary rules and integrations when the decision can be written down. Add AI only where unstructured information or variable language makes those rules insufficient—and keep human review for consequential actions.

This is a selection method, not a promise that a particular tool will save a fixed number of hours. NIST's AI risk guidance on mapping a use case asks teams to define the intended use, consider non-AI alternatives and set oversight responsibilities. Its management guidance asks whether an AI system is appropriate for the task.

Start with the work, not the tool

Write down three tasks your team repeats every week. For each one, capture where it starts, who performs it, which systems they open, what they decide, where the result goes and what happens when information is missing. A candidate is easier to automate when these steps are consistent and the output can be checked.

Consider a shared inbox. If staff copy a form submission into a CRM, assign an owner and send a standard acknowledgement, the handoff may be handled through workflow automation and system integration. If messages arrive in free text and someone must interpret the request before routing it, an AI-assisted classification step may help. In either version, sensitive or uncertain messages should have a review path. These are illustrative designs, not MTREXA client results.

Compare candidate processes with five questions

Use the questions below to shortlist a pilot. They are a practical editorial checklist, not an external certification or a universal scoring formula.

1. How often does it happen? A recurring task creates enough examples to understand the normal path and measure whether the change helps.
2. Are the inputs and outcome clear? Identify the source of truth, required fields, expected output and the point at which the task is complete.
3. Can the systems connect? Check whether the CRM, email, database or other tools have usable APIs, exports or approved integration methods. Do not assume access merely because a product has an API.
4. How many exceptions occur? Note missing data, duplicate records, unusual requests and cases that need judgment. A process with many undocumented exceptions needs discovery before automation.
5. What happens if it is wrong? A mistaken internal notification is different from an incorrect invoice, disclosure of personal data or message to a customer. Set permissions, approval steps and a way to undo or correct errors according to the stakes.

NIST recommends selecting tests and metrics for the actual context, documenting acceptable performance limits and comparing behavior before and after deployment in its measurement guidance.

When is a rules-based workflow enough?

Choose a rules-based workflow when the inputs are structured and the decision can be expressed as stable conditions: “If the form is complete, create a CRM task; otherwise request the missing field.” This is usually easier to explain, test and audit than asking a model to infer a known rule.

Consider AI when the task depends on interpreting variable text or documents, drafting a response from approved information, or comparing several pieces of context. Even then, let deterministic software handle identity, permissions, calculations and final actions where it can. An AI agent may be useful when a bounded task needs several tools and decisions, but its authority, allowed data, stopping conditions and human approvals must be defined before it operates.

AI is not automatically the better solution. NIST explicitly asks teams to examine non-AI alternatives and to weigh an AI system's risks against its benefits before proceeding.

Run a small pilot with a real baseline

Pick one process, one team and a limited set of representative cases. Record the current time per case, error or rework rate, exception rate and time to resolution. Then test the proposed workflow on normal cases and difficult ones: incomplete forms, duplicate records, unusual wording and unavailable systems.

Decide in advance what counts as a pass. For an AI-assisted step, review output quality on a defined sample, track when a person overrules it and log errors that could affect customers or records. Keep a manual route available while the pilot is evaluated. The decision to expand should follow observed results and the cost of operating and reviewing the system, not a demo that handles only easy examples. NIST's measurement guidance calls for fit-for-purpose tests, documented limits and ongoing review.

What affects the cost and scope?

The largest unknowns are usually the number of systems, quality and access of the source data, volume of exceptions, approvals, security requirements and the amount of testing needed. A simple routing rule and a multi-system agent are different projects. Model usage, monitoring and human review can also add continuing costs when AI is involved.

Before requesting a proposal, bring a short process map, sample inputs with sensitive data removed, the systems involved, a process owner and one measurable goal. MTREXA's AI opportunity assessment is a starting point for describing the workflow; the contact page is available if you want to discuss a specific integration or approval boundary. A reliable estimate depends on reviewing those details.

A sensible first choice

If two candidates look equally valuable, begin with the one that has cleaner inputs, a clearer owner and a safer way to recover from mistakes. Prove that the handoff works, measure the result and document the exceptions. Then decide whether to extend the workflow, introduce an AI-assisted step or leave the human process in place.

Sources

• NIST AI Risk Management Framework Playbook, Map — intended purpose, context, alternatives and human oversight.
• NIST AI Risk Management Framework Playbook, Manage — whether AI is appropriate and how risks and benefits are weighed.
• NIST AI Risk Management Framework Playbook, Measure — fit-for-purpose metrics, testing and pre/post-deployment assessment.