Secure AI agents. The agent sits on your data, so who built it matters.
An AI agent reads your records and acts inside your systems. That is the whole point of one, and it is also the whole risk. We are a Central New York managed IT and security firm that has protected local businesses since 2005, on site across all of Upstate New York and at the other locations our multi site clients run, and we build agents on that same ground. Call (315) 682-6372.
New to this? Start with one agent and see how it is locked down.
- over 21 years of managed IT and security in Central New York
- Least privilege by default, agreed in writing before go-live
- A person signs off on anything risky or hard to undo
A secure AI agent is one that can only reach the data its job requires, hands anything sensitive to a person before acting, and leaves a record of what it did. Express IT Solutions sets those rules before the agent touches live work: least-privilege access, your data handled to your privacy and compliance requirements and kept out of public models, human sign-off on anything risky, and one accountable local team. Fayetteville, New York, since 2005. Call (315) 682-6372.
Seven rules we set before the agent does any real work.
Data governance is part of the build here, not a box we tick at the end. Every one of these gets decided with you and written down before go-live.
-
Least-privilege access
The agent gets the narrowest permissions that let it finish its job, and nothing beyond. If it does not need your payroll system, it cannot see your payroll system.
-
A defined data boundary
We agree exactly which data the agent may read and what happens to it, handled to your privacy and compliance requirements and kept out of public models.
-
Human approval gates
Anything sensitive or hard to undo waits for a person to say yes. Sending money, deleting records, and messaging a customer about something delicate are not the agent’s call.
-
A record of what it did
Actions get logged so you can go back and see what happened and why. When something looks wrong, you are reading a record, not guessing.
-
Its own identity, and an off switch
The agent gets its own credentials rather than borrowing a staff login, so it can be reviewed, restricted, or switched off without disrupting anybody’s account.
-
Somebody watching it
We monitor how the agent behaves after go-live the same way we monitor the rest of your systems, and we tune it when your business changes.
-
A written way to retire it
Stopping for good gets planned before go-live rather than improvised later: what happens to the agent’s credentials, who keeps the logs, and how the work moves back to people. A clean exit is part of what drives AI agent cost, so it belongs in the scope from day one and not in a difficult conversation two years later.
An AI shop has to learn your network. We already run it.
Most firms make you choose: an IT company that keeps the lights on, or a separate AI shop that has never touched your systems. We are both. For a lot of our clients, the network an agent runs on is one we already look after through our managed IT services in Syracuse and Central New York, built, patched, and watched for years.
- We know what is on your network because in many cases we put it there. Nobody has to discover your setup from scratch to scope an agent safely.
- Security is not a separate vendor. The people writing the agent’s access rules are the people who manage your accounts and firewall.
- We will say no. If a workflow should not be automated yet, or the network needs work first, you will hear that instead of a proposal.
- One accountable team. IT, security, and AI under one roof, so nobody gets to point at the other vendor when something goes wrong.
2005
Running networks for Central New York from an office in Fayetteville, and running the agents that sit on them.
Three ways to get an agent, and what each does about security
These are the honest structural differences between the options, not a claim about any particular company.
| On security | Off-the-shelf tool | An MSP that builds agents | An AI-only agency |
|---|---|---|---|
| Knows your existing network | No, it is a product | Already manages it | Has to learn it first |
| Who sets the access rules | You do, in its settings | The team that runs your accounts | The builder, then handed to you |
| Can fix the network if it is not ready | No | Yes, same team | Not their remit |
| Who you call when it misbehaves | A support queue | A local number, real techs | Your account manager |
| Ongoing monitoring after go-live | Whatever the product offers | Part of the service | Usually a separate retainer |
| Best when | The task is generic and low risk | The agent touches real business data | You need deep custom AI work and have IT covered |
Scroll the table sideways for every column
This compares the three routes structurally. Any specific provider may differ, and a good AI-only agency paired with a solid internal IT team is a perfectly sensible setup.
Is my business data safe with an AI agent?
It depends entirely on how the agent is set up, which is the honest answer and the reason this page exists. An agent is only as safe as the permissions it is given, the boundary drawn around your data, and the actions it is allowed to take without asking. Set up carelessly, it is a login with no judgment. Set up properly, it reaches only what its job needs, hands anything sensitive to a person, and leaves a record. The technology is not the risk. The configuration is.
| The risk | What actually causes it | What we set up instead |
|---|---|---|
| Sees too much | Given a broad admin login for convenience | Its own identity, narrowest permissions that work |
| Acts too freely | No approval step on damaging actions | Human sign-off on anything risky or hard to undo |
| Data goes somewhere | Nobody defined the boundary up front | An agreed boundary, kept out of public models |
| Nobody notices | No logging and no one watching it | Logged actions and ongoing monitoring |
Scroll the table sideways for every column
The technology is not the risk. The configuration is.
What is least-privilege access for an AI agent?
Least-privilege access means the agent is given the narrowest set of permissions that still let it finish its job, and nothing more. It is the same principle we apply to staff accounts, applied to software that acts on its own. In practice it means the agent gets its own identity rather than borrowing a person’s login, its permissions are scoped to the specific systems and records its workflow touches, and anything outside that scope is denied by default rather than allowed by oversight.
- Its own identity. Not a shared or borrowed staff login, so it can be audited and revoked cleanly.
- Scoped to the workflow. A support agent reads the support inbox. It does not get the finance folder as well.
- Denied by default. Anything not explicitly needed is off, rather than on until somebody notices.
- Reviewed as things change. When the workflow or your business changes, the permissions get looked at again.
Where does our data go when an agent processes it?
That is exactly the question to ask any vendor before you sign anything, and you should get a specific answer rather than a reassurance. On the agents we build, we agree the data boundary with you before go-live: which systems the agent may read, what happens to that data while it works, and where it is kept afterwards. Your data is handled to your privacy and compliance requirements and kept out of public models. If a workflow involves data that cannot leave a particular system, we design around that or we tell you it is not a good fit.
If your business has specific regulatory obligations, bring them to the scoping conversation. We will tell you plainly what we can and cannot support rather than promising coverage we have not confirmed.
The security setup that comes with every agent we build
This is not an add-on tier. It is how we build, and it gets written down and agreed with you before anything goes live.
We would rather turn down the project than rush the security.
Sometimes the right answer is that a workflow should not be automated yet, or that the network needs attention before an agent goes anywhere near it. We will say so. That has cost us work, and it is still the right way to run this.
- Security first, not security later. The access rules get set before go-live, not after something goes wrong.
- No overselling the technology. Agents are good at some things and bad at others, and we will tell you which is which.
- Your rules win. If your compliance obligations rule something out, we design around it or we say no.
- A real local number, (315) 682-6372, answered by the people who did the work.
What people ask before letting an agent near their data
Will our data be used to train a public AI model?
Not on the agents we build. We agree the data boundary with you before go-live and your data is handled to your privacy and compliance requirements and kept out of public models. If you want that written into the scope document, ask and we will put it there.
What stops an agent from doing something expensive by mistake?
Approval gates. Anything sensitive, financial, or hard to undo waits for a person to say yes before it happens. We agree that list with you during scoping, so the boundary is a decision you made rather than a default somebody else picked.
Can we turn it off?
Yes, and that is deliberate. The agent runs on its own identity rather than a staff login, so it can be suspended or revoked on its own without disrupting anybody’s account. You get a documented off switch and a named person on our side to call.
We are in a regulated industry. Can you work with that?
Bring your specific obligations to the scoping conversation and we will tell you plainly what we can and cannot support. We set the agent up to your privacy and compliance requirements, and where something rules an approach out, we design around it or we tell you it is not a fit. We will not promise coverage we have not confirmed.
Do you have to be our IT provider for this?
No, though it does make the security work faster when we already manage the network. If your day to day IT is handled elsewhere we spend longer at the scoping stage understanding your setup, and we will be honest if we find something that needs fixing before an agent should touch it.
How do we know what the agent actually did?
Actions are logged, so you can go back and see what happened and why rather than guessing. We also monitor how the agent behaves after go-live the same way we monitor the rest of your systems, and we tune it when your business changes.
TELL US THE SYSTEMS
Tell us what the agent would need to touch.
Name the systems and the data. We will walk you through how it would be locked down, and where we think the real risk sits.
- It goes to our office in Fayetteville, not a call centre.
- We will tell you straight if an agent is the wrong tool for the job.
- No revenue question, no budget question, no qualification hoops.
Or just call
That line reaches a person who knows the work. If you want an answer now instead of later today, calling is faster, and we would rather say that than pretend the form is.
The form is the better route if you would rather write it down than say it out loud, or you are somewhere you cannot take a call.
(315) 682-6372Tell us what the agent would need to touch.
Bring us the workflow and the systems behind it. We will walk you through exactly how it would be locked down, what a person would still sign off on, and where we think the real risk is. Free, no obligation, no hard sell.
Free AI security review, or email help@expresssupport.com. If we think a workflow should not be automated yet, we will say so.
On site across all of Upstate New York, worked from one office in Fayetteville. The towns we are in constantly are Syracuse, Fayetteville, Baldwinsville, Cazenovia, Chittenango, Lafayette, Jamesville, Liverpool, Manlius, Minoa and Tully. Clients with more than one office get the same team at every location, including the ones we fly to.
Express IT Solutions · Secure AI agents for Central New York business · Last updated August 12, 2026