I run four companies and all four have AI working in production. Not as a demo: answering email, qualifying leads, getting ahead of admin work. And in all four the order was the same, because the reverse order does not work.
The three things AI does well today
After two years putting it into real businesses, what survives production comes down to three families of task. Everything else is still a pretty demo.
1. Read, classify and route
A model reads an email, a form or a WhatsApp message and decides what it is and where it goes: quote, complaint, supplier, noise. It has the best ratio of cost to weight-off-your-shoulders, because it is boring, it happens a hundred times a day, and the mistake is caught immediately.
2. Draft it, never send it
AI writes the first version of the reply, the proposal or the report. A person looks at it and lets it go. The saving is not in writing faster: it is in not starting from zero, which is where the half hour disappears.
3. Watch and raise a hand
A quiet process that watches the numbers, the conversations or the calendar and speaks up when something drifts: a client with no reply for 20 days, a quote about to expire, an expense that does not add up. It is the one nobody sells and the one that saves the most money.
The three it cannot, whatever the model
- Decide with data it does not have. A model does not know that client pays late if that lives in your salesperson’s head instead of a system. AI does not fix missing data: it makes it obvious.
- Carry the responsibility. When the answer is wrong and the client is angry, a person answers. If you have not decided who, you automated the problem, not the work.
- Supply the business judgement. What margin you accept, which client you turn down, when you drop the price. That is not delegated to a model, it is delegated to a rule you write — and then yes, AI applies it without fail.
Where to start: the most repeated, least important process
It sounds backwards, but it is the only entry point that cannot get expensive. Take the process that meets all three conditions: it happens often, it eats your time, and if it goes wrong nothing serious happens. That is where you learn how the model behaves with your data and your people, with the risk contained.
What you must not do is start with the critical process because “that is where it shows”. It is also where the failure shows, and the first visible failure kills internal adoption for a year.
| If the process… | Then… |
|---|---|
| Repeats often and is verifiable | Automate it, with or without AI |
| Repeats often and needs judgement | AI proposes, a person approves |
| Happens rarely and is costly if wrong | Leave it with a person |
| Changes every month | Fix it before automating anything |
The most expensive mistake: buying before looking
The usual sequence is inverted: the tool gets bought, and then someone looks for a use for it. That is how you pay twelve months of licence to discover in month three that the process that hurt was a different one.
What works is one afternoon of inventory. The processes of a small company fit on one sheet of paper: how many times a month each one happens, how long it takes, and what happens if it goes wrong. With that sheet in front of you the decision takes twenty minutes, and it is almost never the one you assumed. I go into it in business process automation: what to automate first.
How to know whether it worked
Before switching anything on, write down two numbers: how long that process takes per month today, and how often it goes wrong. Without them, AI will look either magical or useless depending on the day, because you will have nothing to compare against.
And when AI starts acting on its own
There is a moment when the model stops suggesting and starts doing: it sends the email, changes the client’s status, moves the invoice. That changes the conversation, because we are no longer talking about a tool but about how much autonomy you granted it.
That jump matters enough that I gave it a unit of measurement. The Vergara Agentic Autonomy Scale is six levels, V0 to V6, for saying precisely how much an agent can do without asking. It works for demanding it from a vendor and for not fooling yourself about your own.
Frequently asked questions
How long before an AI project shows results?
If the first process is well chosen, between two and six weeks. A project that needs six months to give its first signal is almost always cut wrong: it needs splitting.
Do I need a lot of data to use AI in my company?
For the three tasks that work today — classifying, drafting and watching — you do not need historical data: you need present data in the right place. The problem is almost never the amount, it is that it is spread across five tools and three people’s heads.
Is AI going to replace my team?
Not your team; the tasks your team does reluctantly and without judgement, yes. In practice the same team handles more volume without new hires, and that is only good if you knew what you wanted them to do with the time freed up.
Shall we do this with your business on the table?
One session: you arrive with a problem and leave with a plan. What to delegate, what to automate and what an AI can already do.
Business process automation: what to automate first (and what not)
The one-afternoon method for deciding which process to automate, which to delegate, which to hand to an AI and which to leave alone for now.
8 min → CRM and AICRM with AI for business: what actually changes
A CRM with AI is not a chat window bolted on the side. The four things that actually change, the five questions for a vendor, and how not to migrate duplicates.
9 min → FundamentalsAI and neural intelligence: how much brain is in a model
Where the word “neural” comes from, what a neural network actually does, and when the brain metaphor starts costing you money in business decisions.
10 min →