Questions to ask before you sign an AI proposal

Our service desk runs on agents. A password reset, a mailbox permission request, a license assignment, a new starter's account: an agent reads the ticket, checks the system, does the work and then closes it. When the ticket needs judgment, an engineer takes over with the history already assembled. Overnight, our analysts read one correlated account of what happened across every device and mailbox instead of forty separate alerts. In finance, a tool compares what we buy against what we invoice and flags the difference.

None of this began as a product. It began because we run a service business across four continents and wanted to know which parts of it could safely be given to software. The answer took longer to find than the tools took to install.

Some context, if this is your first encounter with us. Lenet has been an IT services company since 2005. We work with around fifty companies across the Americas, Europe, Africa and Asia. For some of them we are the entire IT department, while for others we sit alongside an in-house team, taking the routine tickets and the night cover. AI adoption is the newest thing we sell but the oldest thing we practise.

The claim that most IT providers cannot make

Every IT company sells AI now but very few have automated their own work. A provider can resell you a license, configure a tenant and run a demo on sample data. But that is not the hard part. The hard part arrives in six weeks, when the agent answers a question wrongly and nobody can explain which system gave it the bad number. A provider who has never lived through that week cannot tell you where your process will break. They can only tell you the model is improving.

We know where it breaks because we broke it here first, on our own tickets and our own invoices, with our own clients waiting.

Where AI runs inside LENET

The service desk

Agents resolve routine requests end to end. Engineers keep the work that needs a person, and they get it with context attached rather than a one-line complaint. The point was never to remove the engineers, but to stop spending engineering hours on endless password resets.

Security operations

Agents correlate alerts across endpoints, identities and mail, then produce a single readable account of an incident. An analyst reads one narrative and makes one decision. Alert fatigue is the real cause of missed breaches, and this is the part of our stack we would defend hardest.

Integrations

Our agents read client systems through secure, documented connections. Nobody exports a spreadsheet. This is unglamorous work and it is the difference between a second AI project that moves faster than the first and a second project that starts from zero.

Finance review

An agent reconciles purchasing against invoicing continuously and surfaces the gaps. We now run the same analysis on client technology spending, which is how a cost review stops being an annual argument and becomes a monthly fact.

What our own first agents taught us

We built our first agent against a process nobody owned. It worked in testing and drifted within a month, because the underlying steps kept changing and there was no one to tell us they had. We rebuilt it after assigning an owner.

We also over-permissioned early. An agent inherits the access of whoever configured it, and our first build could read more than it had any business reading. Nothing came of it. It was still a security incident sitting quietly in our own environment. We found it ourselves rather than being told about it.

AI begins with your data, not your model

Which model you choose is a late decision and a reversible one. Where your data lives, what condition it is in and which systems disagree with each other: those are the decisions that determine whether anything you build survives contact with a real week of work.

So we assess before we pilot. That sequence came out of running it on ourselves, not from a framework.

Three questions worth answering before you sign anything, with us or with anyone:

1. Which process will this improve, and who owns it today?
2. Where does the data it depends on live, and is it accurate?
3. Who maintains the result a year from now, and at what cost?

If any of the three has no answer yet, you can start from there.

Ask us the same questions

One rule governs what we recommend to a client: it has to run inside our own business first, on our own tickets or our own books. That makes the questions above fair to put to us, and it means the answers already exist rather than being drafted for the meeting.

If you want a written view of where your business should begin, the AI readiness assessment takes five minutes and comes back with a short recommendation.