AI Automation for Business Operations: Which Three Processes to Start With

"We want to use AI but don't know where to start" is now the most common sentence I hear from founders. The second most common is "we tried, it didn't stick". The two are the same problem. Companies start with the most impressive use case instead of the most boring one, and impressive use cases fail quietly when nobody owns the process underneath them.

This is how I choose the first processes to automate in a company of ten to a hundred and fifty people, and what the work looks like once they are chosen.

The rule: automate a process that already works

AI compounds the value of clean systems and exposes the weakness of broken ones. If a process is undefined — two people do it two different ways, the inputs live in someone's inbox, nobody can say what "done" means — automating it produces a faster mess. So the first filter is brutal: only processes that already have an owner, a defined input and a defined output qualify. Everything else goes on the list for later, after the process is fixed by hand.

This eliminates most of the exciting candidates and leaves the ones that actually pay.

Three questions that rank the rest

1. How many times a week does it happen? A task done fifty times a week with a small saving each time beats a task done monthly with a large saving. Frequency compounds; drama does not.
2. How much of it is judgement, and how much is transfer? Copying data between systems, summarising, classifying, drafting a first version — that is where current AI and workflow tools are reliable. Deciding, negotiating, approving — that stays with people, at least for now.
3. What breaks if it is wrong? Start where a mistake is cheap and visible. Internal summaries, first drafts, lead routing. Not invoicing, not customer-facing promises, not anything that touches money without a human check.

The three processes that usually win

Across the companies I have done this in, the same three keep coming out on top. Not because they are glamorous, but because they score well on all three questions.

1. Lead intake and qualification

A form or an inbox receives an enquiry. Someone reads it, decides if it is real, enriches it, creates a CRM record, assigns it and writes the first reply. That is four transfers and one judgement. Automating the transfers — parsing the enquiry, enriching from public data, scoring against a written definition of a good lead, creating the CRM record, drafting the reply for a human to send — typically removes most of the manual time and, more importantly, removes the two-day delay that was losing the warm leads.

2. Reporting and the weekly numbers

If your weekly numbers are assembled by hand from several systems, this is the highest-return automation you will do, and it is not really an "AI" project. It is integration: shop, CRM, ad accounts, finance and delivery data flowing into one place on a schedule. The AI layer, if any, writes the commentary — what moved and the likely reason — for a human to check. I have seen this take weekly reporting from days to hours, and it changes the management meeting from arguing about numbers to deciding on them.

3. Internal handoffs between marketing, sales and delivery

A deal closes and someone has to tell delivery. A project finishes and someone has to tell finance. A customer complains and someone has to tell the account owner. These handoffs are where work stalls in almost every company. Automating them is mostly about triggers and templates: when this status changes, create that task, notify that person, with this summary. It is unglamorous and it removes an astonishing amount of friction.

What the stack looks like

For companies this size I keep it deliberately small:

• A CRM as the system of record — usually HubSpot, sometimes Salesforce. If the CRM is not trusted, fix that first; automation built on a distrusted CRM inherits the distrust.
• A workflow layer — n8n or Make — to move data and trigger actions between systems without waiting for engineering.
• A language model — Claude or OpenAI — called from the workflow layer for the judgement-light steps: summarise, classify, draft, extract.
• A human checkpoint before anything leaves the company or touches money.

The whole thing can be built by one capable operator with a workflow tool and a clear process definition. What it cannot survive is an undefined process or an absent owner.

What not to automate yet

• Anything customer-facing that makes a promise: quotes, delivery dates, refunds. Draft, yes; send, no.
• Any process you are about to change. Automate after the redesign, not before.
• Anything where the input quality is bad. A model given a messy CRM produces confident nonsense.
• "Agents" that run multi-step tasks unsupervised. The tooling is improving fast; the failure modes are still expensive and hard to see.

How to run the first project

Pick one of the three above. Write the process down on one page: trigger, inputs, steps, output, owner, what "wrong" looks like. Build the automation in two weeks, with the owner watching. Run it in parallel with the manual process for another two weeks and compare. Only then switch over. Measure one number — hours saved per week, or lead response time, or reporting cycle — and put it in the weekly review so the win is visible and the next project has a budget.

That is the whole method. It is slower than buying an "AI platform" and faster than every project that starts with the platform. For the CRM-specific version, see HubSpot + n8n: automating the handoffs nobody owns.