The most common AI automation project we see starts with a tool. Someone has watched a demo, an agent that answers emails or writes proposals, and the question becomes “where could we use this?” That is backwards. The tool is the cheapest part of the decision. The expensive part is choosing the wrong process, automating it well, and discovering that nobody’s day got shorter.
Start with the queue. Every business has work that piles up: enquiries waiting for a reply, quotes waiting to be drafted, records waiting to be updated, reports waiting to be assembled. The queue is where the hours go, and it is where the first automation should go.
Four tests for a first automation
Score each candidate process against these. The one that passes all four is the first project. The ones that fail the third are the ones to leave alone.
1. Volume: does it happen often enough to matter?
Automation has a fixed cost and a per-use saving. A task that takes twenty minutes and happens twice a month is not worth a build, however annoying. A task that takes four minutes and happens forty times a day is. Count it before you guess. Most businesses overestimate how often the interesting work happens and underestimate the boring work.
2. Judgement: can the rules be written down?
Take the task and try to write the instructions a competent new starter would need. If you can, and the exceptions fit on the same page, it is a candidate for rules-based automation, which is cheaper and more predictable than an AI agent. If the instructions keep saying “it depends” and the dependency is on reading something (an email, a document, a form with free text), that is where an AI agent earns its place. If the dependency is on knowing the customer, the history or the politics, it is not a candidate yet.
3. Cost of error: what happens when it gets it wrong?
It will get it wrong sometimes. The question is what that costs. A mis-categorised enquiry that a human catches in the morning review costs a minute. A wrong quote sent to a customer costs a relationship. A wrong figure filed with a regulator costs more than the automation ever saved.
The design answer is to put the automation in front of a human where the cost of error is high, and let it act alone where the cost is low. The first project should be one where a mistake is cheap and visible.
4. Data: is the information actually available?
An agent that triages enquiries needs to read the enquiries. One that drafts quotes needs the price list, the rules and the past quotes. If that information lives in someone’s head, or in six spreadsheets that disagree, the automation project is really a data-tidying project with an automation at the end. That is still worth doing. It just takes longer, and you should know that going in.
The usual first wins
Across the businesses we work with, the same handful of processes pass all four tests more often than anything else.
- Inbound enquiry triage. Reading each enquiry, classifying it, extracting the details that matter (location, budget, timeframe, service) and routing it to the right person with a summary. High volume, clear rules, cheap errors, and the data is right there.
- First-response drafting. A reply drafted in the business’s voice for a human to check and send. The human stays in the loop; the blank page disappears.
- Quote and document preparation. Pulling the right template, filling the known fields, flagging the ones that need a decision.
- CRM hygiene. Logging calls and emails, updating stages, chasing the follow-up nobody scheduled. Unglamorous, and consistently the highest-value item on the list.
- Reporting. Assembling the weekly numbers from the systems that hold them, so the Monday meeting starts with the figures rather than a search for them.
What to leave alone
- Anything where the person doing it is the reason customers stay. Relationship work does not automate; it gets worse.
- Anything with a regulator on the other end and no human check in the middle.
- Anything you cannot yet describe. If the process changes every time depending on who does it, standardise it first.
- Anything that happens rarely enough that a checklist would do.
Agent or workflow?
A worked example
Take a design-and-build firm that receives around thirty enquiries a week through its website, phone and email. Each one is read by the office manager, who works out what it is (a loft, an extension, a full renovation, a supplier pitching, spam), asks for the missing details, and passes the real ones to the right surveyor with a note. Twenty minutes each, most of it re-typing information the enquirer already gave.
Against the four tests: thirty a week is enough volume to matter. The rules can be written down, but the input is free text, so this is agent territory for the reading step. A mis-routed enquiry is caught at the morning check, so the cost of error is low. And the data, the enquiries themselves, is already arriving in an inbox.
The automation reads each enquiry, classifies it, extracts postcode, project type, budget band and timeframe, drafts a reply asking for whatever is missing, and files a summary against the right surveyor for the office manager to confirm. The office manager still sends the reply. Twenty minutes becomes three. Nobody built a chatbot, nobody replaced a person, and the number that changed is one the business already tracked: how long enquiries wait for a first response.
Not every automation needs AI. A workflow tool with clear rules is cheaper, faster and easier to trust when the inputs are structured. An agent is the right choice when the input is language: email, documents, transcripts, free-text forms. Most good first projects are a workflow with an agent doing one specific step inside it, reading the unstructured part, rather than an agent running the whole thing.
What a first project looks like
Scoped to one process, with a human in the loop wherever the cost of error is not trivial, and measured against a baseline you recorded before you started. That last part gets skipped constantly. If you do not know how many enquiries came in last month and how long each took to handle, you will not know whether the automation worked, and neither will the agency.
How long it takes depends far more on your systems than on the AI. One workflow against tools with clean APIs and a clear internal owner moves quickly; the same workflow against a legacy system with no API and three people who disagree about the process does not. Scope the first one narrowly enough to reach production and produce evidence.
A good first project changes one number that somebody already cares about. Everything else follows from that.
Start with the queue, not the tool.
Michail leads strategy, development, and growth at Molo Agency. He works closely with clients on SEO, web design, automation, and custom software projects, with a strong focus on building systems that generate real business growth.