Skip to main content Skip to footer

Interested?

Get in touch

Article AI agents series: part one

Before the model: the foundations that decide whether an AI agent works

The model is the easy part

You’ll know the AI conversation tends to start with the model. Which one, how big, how much it costs per thousand tokens. It’s a reasonable place to start. It’s rarely where things go wrong.

In our experience, it’s the data underneath that decides whether an agent works. An agent can only be as good as what it reads, and only as safe as what it’s allowed to see. Get that right and the model choice becomes a much smaller decision. Get it wrong and no model will save you.

This paper sets out the foundations we put in place before we built our own agents, across data, infrastructure and cyber security. It isn’t a framework. It’s what we did, what we learnt and what we’d do first if we were starting again.

Four questions to answer before you build

Before we write a prompt or pick a model, we ask four questions about the data an agent will depend on. If we can’t answer all four, the agent isn’t ready to build.

01

Where does it live?

Not where it should live, but where it actually sits today: which systems, which SharePoint sites, which spreadsheets on which desktops.

02

Who owns it?

Someone has to be accountable for its accuracy, and someone has to say yes when an agent wants to use it. Often that’s two different people, and often neither knows it yet.

03

Does it join up?

The same customer, asset or incident usually appears in several systems under several names. If the agent can’t tell they’re the same thing, it’ll count them twice, or miss the connection entirely.

04

Is it safe to expose?

An agent can be configured to run under different security contexts, including the end user, the agent author, or an account created explicitly for that agent. If the agent operates with permissions greater than those of the end user, it may expose data the user would not otherwise be authorised to access, effectively allowing the user to inherit the agent’s permissions.

None of these are AI questions. They’re questions most organisations have been putting off for years. AI just makes the cost of not answering them visible, and fast.

A small example from our own operations

45%

of duplicate vulnerability findings removed by one AI agent in our security operations, merged into a single prioritised remediation plan.

In our security operations, an AI agent now merges overlapping vulnerability reports into one prioritised remediation plan. It removes 45% of duplicate findings.

The problem it solves is a familiar one. Different scanning tools report on the same estate. They describe the same weakness in different ways, name the same server differently and rate the same risk on different scales. Read side by side, they look like far more work than there really is, and the things that matter most get lost in the noise.

The model was the simple part. The work was getting the data consistent: agreeing how assets are identified, mapping each tool’s findings to a common description, and deciding which source wins when two disagree. Once that was in place, the agent could do in minutes what used to take careful manual cross-checking.

That’s the lesson we’d pass on. Getting the data consistent was most of the work, and most of the benefit. Even without the agent, a single trusted view of what needs fixing is worth having. The agent just makes it repeatable.

Foundation 1: data

We start by mapping the data landscape for one use case at a time, not the whole organisation. A full data strategy is valuable, but it shouldn’t be a gate every agent has to wait behind.

For each use case, our data team covers four things:

  • Mapping. Which sources the agent needs, where they sit and how often they change.
  • Ownership. A named owner for each source, who signs off on its use and stays accountable for its quality.
  • Quality and lineage. How accurate and complete each source is, and where each field comes from, so you can explain an agent’s answer when someone asks.
  • Joining up. Common identifiers and agreed rules for matching records across systems, so the agent sees one customer, one asset or one incident, not five.

The joining-up step is where most of the effort goes, and where most of the value comes from. It’s also the step that’s easiest to skip in a pilot, because a pilot usually runs on a clean, hand-picked sample. Production doesn’t.

Foundation 2: infrastructure

An agent has to run somewhere reliable, close enough to its data and at a cost that still makes sense at scale. A pilot can hide all three problems. A live service can’t.

We look at three things before an agent goes anywhere near production:

  • Where it’s hosted. Public cloud, private or hybrid, decided by where the data lives, what data is allowed to leave that location and what the workload needs. Moving large data sets to suit a model is slow and expensive; it’s usually better to bring the work to the data.
  • Capacity sized for the workload. AI workloads behave differently from the systems most estates were built for. We size for real volumes and peaks, not demo traffic, so cost and performance don’t surprise anyone in month three.
  • A landing zone built for it. A consistent, governed environment with networking, identity, logging and cost controls already in place, so each new agent starts from a known baseline rather than a one-off build.

Most of this is familiar infrastructure work. What changes with AI is the pace: once the first agent proves its worth, the second and third follow quickly. A sound foundation is what lets them follow without starting from scratch.

Foundation 3: cyber security

“Every agent is a new identity with access to data.”

The single most useful way we’ve found to think about AI security

Every agent is a new identity with access to data. That’s the single most useful way we’ve found to think about AI security, because it turns a new problem into one your security team already knows how to handle.

So we treat each agent the way we’d treat a new colleague with privileged access:

  • Identity and access. Each agent gets its own identity, with the least access it needs to do its job, reviewed when its job changes.
  • Data protection. Sensitive content is labelled and protected before an agent can reach it. In a Microsoft 365 estate, that means getting Purview labels and SharePoint permissions right first, because oversharing that’s tolerable for people becomes a real risk when an agent can search everything it’s allowed to see in seconds.
  • Monitoring. Agents are watched like any other identity, with their activity logged and unusual behaviour flagged to a security team who can act on it.

We run our own agents inside the same controls, watched by the same security operations team. It’s why we’re comfortable letting them work on live data.

Where to start

You don’t need every foundation in place across the whole organisation before you build anything. You need them in place for the one agent you’re about to build. Pick a use case with real volume and a clear before-and-after measure, then check it against this list:

Ready-to-build checklist

✓  We know every source the agent will read, and where each one sits today

✓  Each source has a named owner who has agreed to its use

✓  We’ve tested whether records match up across sources, on real data, not a sample

✓  Sensitive content is labelled, and permissions reflect who should really see it

✓  The agent has its own identity, with the least access it needs

✓  We know where it’ll run, and what it’ll cost at full volume

✓  Someone will monitor it once it’s live, and knows what to do if it misbehaves

If you can tick every box, the model is the next decision, and it’s a much easier one. If you can’t, that’s where the work is.

How we can help

Our data, infrastructure and cyber security specialists work in the same team as our data scientists and Microsoft specialists. That means the foundations and the agent get designed together, not handed between departments.

If you’ve got an agent that’s stuck between pilot and production, there’s a good chance the answer sits in one of these three foundations. We’d be glad to look at it with you, and tell you honestly what it would take.

Talk to our team Explore AI at ITPS

Next in this series: part two Speed with a person on every action that matters How we govern agents once they’re running, from approval gates to the agent register. Read part two

Our vendors, suppliers, and partners