AI and automation for CEOs: what to automate first

The short answer: if you run a company of 20 to 200 people, automation and AI do one thing for you before anything else. They take the work that only lands on your desk because no system exists for it, and they give it a system. The weekly numbers, the approvals queue, the inbox, the CRM that stopped being updated in March, the paperwork around every new hire. Start with the one that costs you the most hours a week, judge vendors by what you own when they leave, and keep anything that depends on your judgment alone for now.

Last updated: September 9, 2026

What does AI and automation actually do for a CEO?

It takes you out of the loop on work that only reaches you because there's no other route for it.

At 5 people you knew everything because you did everything. At 40, a surprising amount of the company still routes through you, only now it's by habit rather than by need. A supplier waits for your OK on a 3,000 shekel order. A salesperson pings you on WhatsApp to ask what happened with a proposal. Your bookkeeper emails a spreadsheet on Sunday night and you build the weekly picture yourself on Monday morning.

That's plumbing, and plumbing is exactly what automation is good at.

There's a second thing AI adds on top, and it's newer. Rule-based automation (a form arrives, a row is created, a message is sent) has existed for 15 years. What changed in the last 2 years is that a model like Claude can read an email, a contract or a scanned invoice and tell you what's in it. So the steps that used to need a person reading something can now sit inside the same pipeline. That's the part most guides skip past, and it's the part that opens up processes you'd have called "needs a human" until recently.

In plain terms: fewer things wait on you, and you find out about problems from a system on Tuesday morning instead of from a customer on Thursday afternoon.

What should a CEO automate first?

The first thing to automate is the process that eats the most of your hours and has a clear rule behind it. In a 20 to 200 person company that's almost always one of 5 things.

Pick one. The mistake we see most is picking 3 at once, because each of them looks small on paper. They're small to describe. They aren't small to get right, and the second one goes faster once your team has lived with the first.

If you want a number to argue with, the ROI calculator will give you one from your own hours and wage costs.

The weekly numbers

Most CEOs of mid-sized companies get their numbers late, and they get them in a layout that was convenient for whoever built the spreadsheet.

Here's the version that works. Sales pipeline from the CRM, cash position from the bank, receivables from the accounting system (Priority, SAP Business One, Hashavshevet, whatever you run), open tickets or projects from wherever your delivery team lives. An n8n workflow pulls each one on a schedule, a short Python step does the arithmetic Excel was doing, and the result lands as one page, same layout every week, before you've had coffee.

The AI part is optional and small. A model can write the 4-line summary at the top ("cash down 8% on the month, driven by 2 large invoices slipping into next month; pipeline up; 3 tickets over SLA") so you read the exception rather than hunting for it.

What this is worth depends on who builds the report today. If it's you, it's your Monday morning back. If it's your finance lead, it's a few hours of their week and, more importantly, numbers you trust because no one re-keyed them.

The approvals queue

The approvals queue is the fastest win in most companies and the one CEOs are least willing to give up.

The logic is simple. You already have thresholds in your head: purchases under a certain amount, discounts under a certain percentage, leave requests under a certain length. You approve those every time. You just approve them slowly, because they arrive by WhatsApp between other things.

Put the threshold in a system. Anything under it goes through and gets logged. Anything over it comes to you with the context attached: who asked, what for, what they got last time, and what the budget line looks like. You answer with one tap.

Two warnings. First, write the thresholds down before the build starts, because the act of writing them down usually shows they were never consistent. Second, keep an audit trail on the automatic approvals. The control stays with you. It moves from the moment of each decision to the rule that governs all of them.

The inbox

An inbox that only you can process is a bottleneck with your name on it.

The automation here has 3 layers, and you can stop after any of them. Layer 1 is classification: an AI step reads each incoming email and tags it (customer, supplier, internal, candidate, spam that got through). Layer 2 is routing: the customer ones go to the account owner, the supplier ones to finance, the candidate ones to whoever runs hiring, and only the ones addressed to you stay with you. Layer 3 is drafting: for the emails that do need you, a draft reply is waiting, written from your previous replies to similar mails, and you edit rather than compose.

Layer 3 is where CEOs get nervous, and they should. A drafted reply should never go out without a human reading it. The draft saves the 10 minutes of composing; the reading is still yours.

We run ours on n8n and Claude, and the honest summary after a year is that the classification and routing did more for us than the drafting. Fewer things reach the top. That's the win.

The CRM that never gets updated

A CRM that salespeople update once a month is a diary, and you can't run a pipeline review on a diary.

Every CRM asks the salesperson to stop selling and type, and the salesperson makes the rational choice. So the fix is to take the typing out. Meetings from the calendar create activity records. Emails to a contact attach to the deal. A WhatsApp message from a customer moves the stage. A proposal sent from your document tool sets the value.

This is CRM automation in the plain sense: the CRM fills itself from where the work already happens. Once it does, two things become possible. Your pipeline number on Monday is real. And you can stop asking "what's happening with X" in the team chat, because the answer is on the deal.

If your team has a CRM and hates it, read this before replacing it. Migration takes months. Fixing the update problem takes weeks, and if the update problem survives the migration you'll have paid twice.

Hiring and onboarding paperwork

Every new hire in an Israeli company triggers the same list: employment agreement, Form 101, pension and study-fund enrolment, an email account, the right Slack channels, a laptop, access to 4 systems, and a first-week calendar. Someone assembles that list from memory every time, and something is always forgotten.

The automation is a checklist that runs itself. One trigger (the signed offer letter) creates the accounts, sends the forms, opens a ticket for IT, and chases the new hire for whatever they haven't returned by day 3. The document-processing piece reads the returned forms and files them where the accountant expects to find them.

This one is rarely the first project, because hiring happens a few times a year and the weekly numbers happen every week. It's a good second or third project, and it's the one new employees notice.

What should a CEO leave alone?

Leave alone anything where the value is your judgment, anything that happens fewer than 10 times a year, anything your team can't yet describe as a process, and anything that speaks to customers in your name without a person reading it first.

Judgment first. Deciding which customer to part with, or whether to open a second office: there's no rule to encode, and an AI summary of the inputs is the most you should want. Automating the decision itself is a category error.

Frequency second. A process that runs 8 times a year won't repay the build. Write a good checklist and move on.

Describability is the one that catches people. If 2 people on your team describe the same process differently, you have 2 processes, and automating either one will annoy the person who runs the other. Get the description agreed first. Our own project process spends its first phase on exactly that, and it's the phase that gets skipped when someone's in a hurry.

And the customers. An AI agent answering support questions at 2am is a fine project once you've watched it draft for a month. It's a bad first project.

How should a CEO judge an automation vendor?

Judge a vendor by what you own when they leave, and by whether they can tell you what to leave alone.

Here's the list we'd use if we were on the buying side.

The tools question matters less than it looks. n8n, Make, a Python service, a model from Anthropic or from OpenAI: these are choices a good vendor makes per process, and they should be able to explain each one in a sentence. What matters more is whether your team can read the thing after the vendor is gone. A workflow your operations manager can open and change is worth more than a cleverer one only the builder understands. We wrote up how we work so you can hold us to the same list.

What does the first 90 days look like?

The first 90 days should produce one process running in production that a named person on your team can change without calling anyone.

Weeks 1 to 2 are mapping: someone sits with the people who actually run the process, writes it down, and finds the 3 exceptions everyone forgot to mention. Weeks 3 to 6 are the build, with your team looking at it weekly along the way. Weeks 7 to 8 are running it alongside the old way, because the old way catches what the new way missed. Weeks 9 to 12 are handover and the decision about what's next.

If a vendor promises 5 processes in 90 days, ask which of those steps they're skipping. It's usually the parallel run, and the parallel run is where you learn whether to trust the thing.

The wider case for starting small, with the basics of what counts as a process worth automating, is in our guide to business automation. If you're further along and thinking about where AI fits across the whole company, implementing AI in a business is the longer read.

Frequently asked questions

What should a CEO of a small company automate first?

The process that costs the most of your own hours and has a clear rule behind it. For most CEOs of 20 to 200 person companies that's the weekly numbers report or the approvals queue. Both have a known input, a known output and a known rule, which is what makes a first project succeed. Pick one, not three.

Do I need a technical person on my team to run automation?

You need a person who owns the process and has the authority to change it, and that person doesn't have to be technical. Tools like n8n are readable by an operations manager after a short session. What you do need is a vendor who builds for that reader.

Is AI safe to use on our company email and documents?

It can be, and the safety depends on where the data goes. Ask which model is used, where it runs, whether your data is used to train anything, and whether the workflow itself runs on your own server. A self-hosted orchestration layer with a model accessed through a business API is the common answer, and your vendor should be able to draw it for you.

How long until we see results?

A first process typically runs in production within 6 to 8 weeks, including the mapping and a parallel run alongside the old way. The hours it saves show up immediately; the trust in it takes another month, which is why the parallel run matters.

Should we replace our CRM before automating?

Usually not. If the CRM's problem is that no one updates it, that problem follows you to the new CRM. Fix the update problem with automation first, and only migrate if the software itself can't hold your process.

What should we never automate?

Decisions that depend on your judgment, processes that run fewer than about 10 times a year, and anything that speaks to customers in your name without a person reading it first. Those can come later, if at all.

Next step

If you're a CEO with a list of things that shouldn't need you and do, that list is the whole brief. Talk to us and bring it. The first conversation is about which one to start with and why, and it takes an hour.