AI governance and agent deployment consulting
AI is joining your workforce. Somebody has to decide what it's allowed to touch.
We help organizations put copilots and AI agents to work without handing them the keys: access control first, then deployment, with a written record of who approved what.
How we approach it
Every copilot and every agent you turn on is a new worker with credentials. It never shows up in HR, and it inherits whatever the account behind it can reach. In most environments we look at, that account can reach far more than anyone meant it to. The second workforce is already here.
So we start with access, not use cases. Before anything is deployed we map what the agent would run as, which data that identity can read, and which systems it could change. Microsoft 365 Copilot is the common example. It shows people whatever their permissions already allow, so years of quiet oversharing in SharePoint and Teams become search results on day one. It's cheaper to fix before the rollout.
Then we write two lists. The first is what an agent may do alone: read, summarize, draft, open a ticket. The second is what needs a person: anything that sends money, deletes data, changes someone's access, or speaks to a customer in your name. Every action lands on one list or the other, and every item on the second list has a named approver. Actions move from the second list to the first one at a time, as they prove themselves under monitoring. Two lists. That's the governance. Everything after it is enforcement.
Enforcement is ordinary identity work done properly. Each agent gets its own identity, never a shared admin account. Service principals use certificates, not secrets pasted into config files. Conditional Access and narrow scopes limit what each one can reach, and the logs show what it did on whose authority. When something goes wrong, you can answer the question that matters: who said yes to that?
The tools that enforce it
A policy document doesn't stop an agent. The AI security and control platforms now on the market do, and we partner with several of them. We know what each kind of tool is good at, and where it fits in your organization:
- Prompt redaction: sensitive data is masked or removed before a prompt leaves your environment.
- Agent visibility: an inventory of the agents and AI tools in use, sanctioned or not, what each one connects to and what it did.
- Runtime control: which agents may call which tools and data, enforced while they run rather than written down and hoped for.
- Execution stops: a risky action can be held for approval or stopped mid-run, before it reaches a production system.
- Audit: a record of prompts, actions and approvals that your security team and auditors can use.
We match the tool to your environment, budget and licensing, and we'll tell you when what you already own is enough.
What you get
- An inventory of the AI already in use, sanctioned or not, and the identities behind it
- An access review of the data your first copilot or agent would see, with the oversharing called out
- The two lists, written for your environment, with a named approver for every action that needs a person
- An identity design for agents: separate accounts, certificate-based service principals, scoped permissions and logging
- A controls design: redaction, agent monitoring and stop points, matched to the tools you have or the ones worth adding
- A staged rollout plan for Copilot or your chosen agent platform
- A short AI use policy your staff can read in five minutes
What a first engagement looks like
- 1
Discovery
Shaped to what you need. When AI will work inside your systems, we start with the access review, because that's where the surprises are. When the job is simpler, such as us running AI to compile data and deliver it to you or your systems, the discovery is lighter and may not need access to your environment at all.
- 2
Fixed price
You know the price before we start. We set it from the size and scope of the discovery, so it fits your organization rather than a one-size package. If we find something that would change the scope, we tell you first, and nothing extra happens without your OK.
- 3
Written read-out
A written document: what we found, what we'd do, what it costs, and what we'd leave alone.
- 4
You decide, in writing
Nothing moves to a build until you say yes in writing. Taking the read-out and doing the work yourselves is a fine outcome too.
Every engagement is led by a senior architect, and we bring in specialists from our bench when the work calls for them. After the first job, some clients keep us on as a fractional CIO or CTO for the decisions that keep coming.
Questions CIOs ask
Should we turn on Microsoft 365 Copilot for everyone at once?
No. Start with a small group whose data access you've reviewed. Copilot respects existing permissions, so the risk is the permissions you didn't know you'd granted. Fix the oversharing, roll out in stages, and watch the first group before you widen it.
Our people already use ChatGPT and other tools on their own. Do we block them?
Blocking usually moves the use to personal phones, where you can't see it at all. Give people a sanctioned tool that's good enough, say plainly what data can't go anywhere else, and make the approved path the easy one.
Who should own AI governance: IT, security or legal?
One named person, with IT, security and legal at the table. Somebody has to sign the two lists, and it should be someone who can say no to a business unit.
Can an AI agent have write access to production systems?
Yes, and that's where much of the efficiency and the room to scale comes from. What makes it safe is method. Each agent gets its own identity, scoped to the systems and actions it needs. It starts read-only, then runs alongside your people while its proposed changes are checked against what they actually did. Write access is granted one action at a time as each one proves out. Runtime controls can hold or stop an action before it lands, and every change is logged so it can be traced and rolled back. Actions that move money, change access or delete data keep a named approver for as long as you want one.
An example
Production AI at the industrial edge
As VP of Research, Development & Innovation at a rail-technology company, our founder architected AI that runs in production rather than in a demo: trackside inspection portals that image trains at speed and cut inspection from three to four hours to minutes.
On the high-speed passenger portal, trains pass at up to 125 mph. More than 100 GB/s of image data arrived from multiple acquisition clients. With RAW-to-JPEG conversion running on NVIDIA nvJPEG and NVIDIA AI Enterprise on VMware vSphere, storage sustained roughly 120 GB/s of writes, and read tests reached 200 GB/s. On the freight portal, where trains typically pass at up to 70 mph, the system averaged about 60 GB/s of image writes on top of the processing load.
He also designed a scalable inferencing gateway on NVIDIA Triton Inference Server and Kubernetes. It fit more models onto the existing gateway servers and gave the platform an optional path to redundancy and scale as inferencing load grew.
The lesson we carry into governance work: an AI system is only as trustworthy as the controls on what it can see and change.
See the press coverage on the portfolioWhere this applies
Start with what you want AI to do.
A 30-minute AI discovery call on Microsoft Teams, times in ET. Bring the tool you're being asked to turn on.