Agentic AI for Business — A Buyer's Guide
Agentic AI
for business.
A buyer's guide.
What it actually is, where it pays for itself first, what it needs from you, and the questions that separate a real system from a demo.
Francesco Pettinelli
Why this exists
I built this for my own company first.
I run an operating business. Everything in this guide was built to solve problems I was losing money on — quotes going out late, calls going to voicemail, a follow-up nobody made, a margin nobody noticed slipping until the job was done.
Then I sold the same system to other people, because the problems turned out to be identical whether you pour concrete, file tax returns or ship pallets. Every business that quotes, schedules and invoices leaks in the same four places.
That is the only credential that matters here. I am not writing this from a product roadmap. I am writing it from the receiving end, where a bad tool costs you a real client.
What you will get out of this
- A straight definition of agentic AI, without the vendor fog.
- A ranking of where it pays first, so you don't start in the wrong place.
- Twelve questions to ask anyone selling you AI, and the answers that should worry you.
- A thirty-day pilot designed to prove it or kill it fast, with your own numbers.
One warning before you spend anything
Most AI disappointment is not a technology problem. It is a data problem. If nobody in your company can say what a job costs, who the client is and what was promised, no agent can either. The good news: fixing that is worth doing regardless of whether you ever buy a thing.
The definition
Chatbot, automation, agent.
Three different things sold under one word. The difference is not how clever they sound. It is what happens when reality does not match the plan.
| Chatbot | Automation | Agent | |
|---|---|---|---|
| What you give it | A question | A rule you wrote | A goal |
| What it does | Replies | Repeats the same steps | Works out the steps itself |
| Unexpected input | Apologises | Breaks | Adapts and reroutes |
| What it remembers | The current chat | Nothing | Everything about your business |
| Who it talks to | You | Nobody | The rest of the team |
| Who signs off | Not applicable | Nobody | You do |
The test that settles it
Give the thing a situation nobody wrote a rule for. A client calls at 9pm asking to change a scope you quoted six weeks ago, and mentions they need it before a date you have never discussed. A chatbot will apologise. An automation will not even notice. An agent will pull the original quote, work out what the change does to the price and the schedule, draft the revised version, flag the date conflict, and put the whole thing in front of you to approve.
That is the entire difference, and it is the reason the word matters commercially. You are not buying a smarter interface. You are buying something that finishes work.
The part vendors skip
One agent is a toy. The memory is the product.
You can buy a single agent that answers your phone. It will be impressive for a week, and then it will annoy you, because it knows nothing about what happens after the call.
The value shows up when the agents share one memory. Reception writes down what the caller wanted. Quoting reads it without being told. Purchasing sees what was sold. Scheduling sees what was ordered. Finance sees what it cost. Nobody re-explains anything, because there is nothing to re-explain — it is all in the same place.
Where your money actually leaks
- The handoff. Every time work passes between two people, context gets dropped. Most businesses have five to nine handoffs per job.
- The gap. The hours between a customer asking and anyone responding. This is where deals die quietly.
- The re-explanation. Time spent telling somebody something the business already knew.
- The forgotten thing. The follow-up nobody made, the invoice nobody chased, the client nobody called back after two years.
Shared memory attacks all four at once. That is why a team of agents is worth more than the sum of its parts, and why buying one agent at a time rarely gets anyone excited.
A question worth sitting with
If your best employee left tomorrow, how much of what they know leaves with them? Whatever that number is, that is the size of the memory problem you already have. AI did not create it. It just makes it fixable.
Sequencing
Where it pays first.
Not everything is worth automating on day one. Start where the payback is fastest and the risk is lowest, then work down. This is the order I would use in almost any business.
1. The response gap
Answering enquiries fast, every time, at any hour. It is the shortest path to money because it converts demand you are already paying to generate. If you do nothing else, do this.
2. Quoting and proposals
Getting priced work out the door in minutes instead of days. Fast, consistent, and priced off real costs rather than whoever is doing it at 11pm.
3. Follow-up
The unglamorous one that quietly outperforms everything else. Most businesses have more money sitting in unfollowed quotes than in their entire new-lead pipeline.
4. Scheduling and dispatch
Fewer wasted trips, fewer collisions, fewer angry calls about a window nobody kept.
5. Collections
Polite, relentless, and unemotional in a way humans find hard. Nothing improves cash flow faster.
6. Margin visibility
Knowing which jobs make money while you can still do something about it, instead of at year end.
What I would not start with
Content and social posting. It is the easiest thing to demo and the slowest thing to pay you back. If a vendor leads with it, they are showing you what is easy for them, not what is valuable for you.
Your side of the deal
What it needs from you.
This is the part nobody puts in the sales deck. An agentic system is only as good as what you load into it, and four things have to come from your side.
Your data, in one place
Customers, services, prices, suppliers, people, and history. It does not have to be beautiful. It has to exist and be roughly true. Half a day of cleanup here is worth more than any feature.
How you actually work
Not the process on the wall — the real one, including the exceptions. Every business has rules that live in one person's head. Those rules are the difference between an agent that sounds like you and one that sounds like a call centre.
One decision-maker
Somebody who can say yes on the spot. Committee-run rollouts stall, not because the technology fails, but because nobody is allowed to approve the first thing that goes out.
An approval habit
For the first few weeks you review what it produces before it sends. Ten minutes a day. That review is not overhead — it is the training. What you correct in week one is what it gets right forever after.
The honest timeline
Week one it needs supervision. By week three you will be approving in batches without reading closely. By week six you will notice you stopped checking some categories entirely. Anyone promising you day-one autonomy is selling you a demo.
Due diligence
Twelve questions to ask any vendor.
Ask these in order. You will learn more from the first four than from any demo.
01Show me it handling something you did not prepare for. Hand them a real situation from your business, live.
02Where does the memory live, and can I export it? If the answer is vague, you are renting your own data.
03What happens when it is wrong? There must be an approval step and an audit trail. Both.
04Do the agents share context, or are they separate tools? Separate tools with one login is not a team.
05Who else has my client list? Get the sub-processor list in writing.
06What does it cost when I double in size? Per-seat pricing punishes growth. Know the curve.
07What is the exit? How do I leave, what do I take, how long does it take.
08Who does the setup, and how long? "Self-serve" usually means "you do the hard part".
09Show me a customer in my industry. Then ask to speak to them without the vendor present.
10What can it not do? An honest answer here is the strongest signal you will get.
11What happens when the model changes? Everything underneath moves every few months. Who absorbs that?
12Who do I call at 7am when it sends something wrong? Get a name, not a portal.
Warning signs
Seven red flags.
And one green flag worth more than all of them
The vendor tells you not to buy something. When somebody talks you out of the expensive module because your business is not ready for it, they are optimising for you still being a customer in two years. That is the person to buy from.
The pilot
Thirty days to prove it or kill it.
Do not roll it out everywhere. Pick the response gap, run it for a month, and measure two numbers you already understand.
Kill criteria, written in advance
Decide now what result would make you walk away, and write it down before you start. A pilot without a kill criterion is not a pilot — it is a purchase with extra steps.
Control
You keep the keys.
The single most common fear I hear is not "will it work". It is "what will it say to my client without me knowing". That fear is correct, and the answer is architectural, not reassuring words.
The approval queue
Nothing reaches a customer until a human releases it. Quotes, emails, texts, contracts — all of it drafts into one queue. You approve, edit or reject. Over time you widen what is allowed to go straight out, category by category, at your pace.
The audit trail
Every action an agent takes should be logged with what it did, why, and what it read to decide. If you cannot reconstruct a decision after the fact, you cannot manage the system — and one day you will need to explain something to a client.
Your data stays yours
Ask where it is stored, who can read it, and how you get it out. Insist on export before you sign, not after you are unhappy. A system that holds your customer history hostage is worse than no system at all.
The rule I run my own business by
It can do the work. It does not get to decide what leaves the building. That line has never moved, and everything else is a matter of taste.
One page
The checklist.
Print this. Take it into every conversation.
- I know my current response time and quote conversion rate.
- My customer, service and price data exists somewhere I can point to.
- I have written down the rules that live in one person's head.
- One named person can approve decisions without a meeting.
- I have picked ONE place to start, not five.
- I have written my kill criteria before spending anything.
- The vendor demoed on my data, not theirs.
- There is an approval step between the agent and my customers.
- There is an audit trail I can read.
- I know how to export everything and leave.
- I know what it costs at twice my current size.
- I have a name and a phone number for when it goes wrong.
If you want to see one running
Twenty minutes, a screen share, real jobs and real numbers — not a demo dataset. If it is not right for your business I will tell you on the call rather than sell you something you will switch off in a month.
Your next hiredoesn't sleep.
Twenty minutes on video. We look at how you run today, show you the system live on a real contracting business, and tell you honestly whether we can move the needle.