“Can we add AI?” is not a product requirement.

It is the beginning of a diagnosis.

Sometimes AI removes hours of repetitive work.

Sometimes a form, rule, search index, or better template solves the problem with less cost and risk.

Before automating a workflow, I use this checklist.

1. Is the pain frequent and expensive?

A workflow used twice a year is rarely the first automation target.

I look for:

• repeated volume;
• meaningful time spent;
• avoidable delay;
• costly inconsistency;
• people doing copy-and-paste judgment;
• a result the business can value.

The problem should exist before the model arrives.

2. Can we describe the input?

AI performs poorly when the system cannot identify what it is receiving.

I inventory:

• source;
• format;
• language;
• missing fields;
• size;
• sensitivity;
• ownership;
• freshness;
• known bad examples.

“It reads our documents” is too vague.

Which documents? Whose documents? Which version? With which access?

3. Is the task judgment or deterministic logic?

Use deterministic code for deterministic rules.

Examples:

• tax calculation;
• permission checks;
• invoice totals;
• eligibility thresholds;
• status transitions;
• required-field validation.

AI may help extract, summarize, classify, draft, or rank when ambiguity is part of the task.

It should not replace a rule that can be written and tested exactly.

4. What does a wrong answer cost?

Error tolerance decides the architecture.

Low-cost example:

• draft a subject line that a human reviews.

Higher-cost example:

• approve financing;
• send a binding quote;
• change account access;
• publish health or legal guidance;
• delete business data.

As cost rises, require stronger evidence, narrower scope, human approval, and deterministic protection.

5. Where does the human belong?

“Human in the loop” is not a button labeled Approve.

The reviewer needs:

• enough context;
• clear uncertainty;
• the original source;
• editable output;
• an easy reject path;
• no pressure to accept everything;
• feedback that improves the workflow.

If review takes longer than doing the task, the automation design is wrong.

6. What data leaves the system?

Before calling a model provider:

• classify the data;
• minimize the payload;
• remove unnecessary identifiers;
• confirm provider retention settings;
• check region and contract requirements;
• keep secrets server-side;
• log safely;
• define deletion.

Do not paste customer data into a provider because the prototype was convenient.

7. What happens when AI is unavailable?

Providers time out. Rate limits change. Output fails validation.

Define:

• timeout;
• retry;
• queue;
• fallback model or deterministic path;
• manual continuation;
• user-visible status;
• no-duplicate behavior.

“Try again” is not enough when the action already caused a side effect.

8. How will output be constrained?

Depending on the task:

• structured output schema;
• allowed values;
• retrieval from approved sources;
• maximum input and output size;
• prohibited actions;
• post-generation validation;
• tool permission boundaries;
• citation back to source records.

The model proposes.

The application enforces.

9. What is the baseline?

Measure the workflow before automation:

• time per case;
• error/rework rate;
• queue delay;
• completion rate;
• human touches;
• customer outcome;
• cost.

Then compare.

Without a baseline, “AI saved time” is a feeling.

10. Can we stop safely?

The business needs:

• feature flag or controlled rollout;
• model and prompt version;
• audit trail;
• rollback;
• stored-output policy;
• owner;
• review date.

An AI feature is an operating system, not a one-time integration.

My decision

I automate when:

• the pain is real and repeated;
• the input is governable;
• ambiguity suits the model;
• error cost is controlled;
• human responsibility is clear;
• privacy is acceptable;
• fallback exists;
• improvement can be measured.

If those conditions fail, I improve the workflow first.

For production safeguards, read responsible AI features for SaaS. For scoping the first release, read how I cut a SaaS MVP.

If you have a manual workflow, send me one real input, the current steps, and the cost of a wrong result.