What is the best first process to automate with AI?

There is no universal answer, but the strongest first candidates run daily or weekly, have a predictable input and output, stay internal or pass a reviewer, and can be checked in seconds. Call notes into a CRM, enquiry triage and report assembly are common examples. Long-form expert content and anything customers see unreviewed are poor first choices.

Should we automate the task that takes the most time?

Not necessarily. The most time-consuming tasks are often the least repeatable and the hardest to check, such as proposal writing. A smaller task that runs every day with a simple review step will give you clearer evidence and fewer surprises, which is what you need to justify the next project.

How long should it take to know if our first AI automation is working?

For a workflow that runs daily, a few weeks of real use is usually enough to see error patterns, review time and time saved. That is one reason to favour frequent workflows first. Decide the success and stop criteria before you start so the review is a decision, not a debate.

Do we need clean data before starting with AI?

You need the data for the chosen workflow to be reachable, consistent and current, not your whole CRM to be perfect. If the candidate depends on data that is scattered or contradictory, fixing that may be the first piece of work, or a sign to pick a different first workflow.

How to Select the First AI Workflow to Automate

The short answer

Pick the workflow that will give you clean evidence fastest with the least chance of embarrassing anyone. That usually means something that happens daily or weekly, follows the same shape every time, produces output a person can check in seconds, and stays inside the building (or passes a reviewer before it leaves).

If you do not yet have a list of candidates, start with AI automation opportunities for established marketing teams, which covers where AI tends to earn its keep and how to build a longlist. This article picks up from there: you have three or four finalists, they all look reasonable, and you need to choose one.

The first workflow has two jobs

The obvious job is to save time. The less obvious job is to teach your organization how to run AI in production: who owns an automation, how output gets reviewed, what happens when it breaks, how you measure whether it worked. Every later project inherits those habits.

That changes the selection logic. A workflow that saves a little time but runs every day, with a visible owner and a simple review step, teaches more than a workflow that might save a lot of time once a quarter. You want enough cycles in the first few weeks to see the failure modes, fix them and build trust in the numbers.

It also means the first project should be finishable. A workflow that needs three systems connected, a cleaned-up CRM and a new approval policy before it can run is a multi-phase initiative, not a first project. Save it for when you know what you are doing.

Four tests for your finalists

Repeatability

Does the work arrive often, in a predictable form, and can someone describe a good result in a sentence or two? Repeatable work has a template, even if nobody wrote it down. If three people do the task three different ways, the first job is agreeing on one way, and that is a process project before it is an AI project.

Risk

Ask who sees the output (staff, clients, the public), whether personal or regulated information is involved, and whether a mistake can be undone. An internal summary with a wrong detail costs a correction. A wrong answer in a public reply, a misquoted price or a confused French translation of a program term costs trust. For a first workflow, choose something where the worst realistic mistake is cheap and visible.

Quality controls

This is the test most teams skip, and it decides more than the others. For each finalist, describe how you would catch a bad output. Is there a simple rule (required fields, allowed values, a length limit)? Is there a person who already knows the right answer and can check in seconds? Can you compare the output to a source document? If the only way to check the output is for an expert to redo the work, the workflow has no cheap control, and it should not go first. Building controls is part of the cost of the project, so a workflow that needs elaborate ones is more expensive than it looks.

Payback

Payback here means two things: whether the time saved, net of review, exceeds the cost of building and running it, and how quickly you will know. A daily workflow produces evidence in a couple of weeks. A monthly one takes a quarter to show three data points. Speed to evidence matters as much as size of prize for a first project, because the decision to fund the second one will be made on what you can show.

A worked example

An illustrative case. Say a 60-person engineering consultancy in Moncton has three finalists from its workflow audit:

First drafts of proposal responses for public sector RFPs, pulling from past submissions.
Call notes into the CRM. Eight business developers each spend a chunk of time after every client call writing up notes, next steps and contact updates, and many calls never get logged at all.
Social posts repurposed from project announcements, drafted for LinkedIn in English and French.

Run them through the four tests.

Proposals look like the biggest prize, since they consume senior hours. But RFPs arrive irregularly, each one is different, the content is full of commitments (scope, staffing, schedules) and the only person who can check a draft is a senior engineer who would rather write it. High risk, expensive controls, slow evidence. Not first. AI may still help later as a research assistant that finds relevant past answers, rather than a drafter.

Call notes happen daily. The input (a transcript or a few rough bullets) has a predictable shape, and so does the output (summary, next steps, updated fields). The output stays internal. Best of all, the checker is the person who was on the call: they can confirm a summary in seconds because they know what was said. Payback evidence arrives within two weeks, and better CRM data is a benefit sales leadership will notice.

Social posts are repeatable and the source material is approved, but the output is public and bilingual, and it needs a fluent French reviewer every time. It is a reasonable second or third project once the review habit exists.

The consultancy picks call notes. Not because it saves the most time on paper, but because it scores well on all four tests and will produce honest evidence quickly. The selection memo names the owner (the business development lead), the control (each developer approves their own summary before it writes to the CRM) and the stop condition (if summaries need heavy editing after two weeks, pause and fix the prompt or the input).

Estimate payback honestly

A payback estimate for a first workflow does not need a spreadsheet model. It needs five honest inputs:

Time per task today, measured by watching the work, not by asking a manager to guess.
Frequency, from real counts over a few weeks.
Review time after automation. Approving a draft is work. If checking takes nearly as long as doing, the saving is imaginary.
Exceptions. The inputs the automation cannot handle still land on a desk, and they tend to be the awkward ones.
Running cost: usage, platform fees, monitoring and someone's time to maintain it.

Take the baseline before you build anything. Without it, you will be arguing about impressions in a month. Measuring ROI from AI marketing workflows covers the baseline and the measures in detail, and what AI marketing automation costs to maintain lists the running costs that business cases tend to forget.

Disqualifiers

Some features knock a candidate off the first-project list regardless of how attractive it looks:

Nobody can describe a good output. If the team cannot agree what right looks like, nobody can check the AI's version.
The knowledge lives in one head. If only one person understands the work and they are too busy to help design it, the automation will drift and nobody will notice.
It needs far more data than the task requires. A tool that wants the whole CRM to draft one follow-up email is a data protection problem waiting to happen.
It is regulated or advice-like. Legal, clinical, financial or employment content can be automated later with qualified review, but not as the learning project.
It does not run right now. Atlantic Canadian businesses know this one. A workflow tied to the summer tourism season or the fall registration rush cannot be piloted in the off-season with real inputs. Either time the project to the season or pick something that runs year-round.
The main reason is that a vendor suggested it. Vendors suggest the workflow their product handles, which is not the same as the one your team needs.

Write a one-page selection memo

Before anything is built, write one page that a sceptical executive could read in two minutes:

The workflow, in one sentence, and why it beat the other finalists.
The owner, and the person who approves any output that leaves the team.
The baseline: time, volume, error rate.
The controls: rules, reviewer, sampling. Where people stay involved is covered in human review in AI marketing workflows.
What success looks like in four to six weeks, and what would make you stop.
The next candidate, so the organization knows this is a sequence, not a one-off.

That page then becomes the brief for a short, bounded trial. How to pilot an AI marketing workflow before scaling covers how to run it so it ends in a decision rather than drifting.

When we run our AI and automation work, this selection happens in the first week, with the people who do the work in the room. It is the least glamorous part of the sprint and the part that most often determines whether the thing is still running a year later.