Your company is already using AI. Nobody has looked at how.

The question stopped being whether to adopt AI. It is now whether anyone in your organization can tell you what is already running, what it can reach, and who approved it.

Most of the AI in a mid-market business did not arrive through a procurement process. It arrived inside tools you already pay for, or through a browser tab someone opened to get a deadline met. It is doing real work, on real data, and in most organizations no one owns it.

The gap is measured, not theoretical

IBM's 2025 Cost of a Data Breach Report studied 600 organizations between March 2024 and February 2025. What it found about ungoverned AI:

One in five organizations reported a breach caused by shadow AI — AI nobody sanctioned. Only 37% have any policy to manage AI use or detect it.

13% reported a breach of their own AI models or applications. Of those, 97% had no AI access controls in place. A further 8% could not say whether they had been compromised at all.

63% of breached organizations either had no AI governance policy or were still writing one. Among those that did have one, only 34% ran regular audits for unsanctioned AI.

Organizations with heavy shadow AI usage carried an average of $670,000 in additional breach cost. Those incidents also exposed personal data 65% of the time against a 53% average, and intellectual property 40% of the time against 33%.

The access surface is growing faster than the policy

Palo Alto Networks' 2026 Identity Security Landscape found organizations now manage 109 machine identities for every human one, with AI agent growth expected at 85% over the following twelve months. More than half of the organizations surveyed said they cannot consistently enforce least-privilege access for service accounts across cloud, SaaS and on-premises systems.

Every agent is an identity with permissions. Most were provisioned by a team trying to ship something, not by whoever is accountable for access.

What the engagement covers

An inventory of what is actually running. Sanctioned tools, unsanctioned tools, and the AI features switched on inside software you already license and may not have noticed.

A map of what each system can reach. Which data stores, which records, which downstream systems — and whether that access was ever reviewed.

A contract and data-processing review of the AI vendors already in your estate. What they are permitted to do with your data, what they retain, and whether anything is being used to train a model you do not control.

A non-human identity and agent access review. Service accounts, API keys and agent credentials, assessed against least privilege rather than against convenience.

A governance assessment against the framework that applies to you. NIST AI RMF, the EU AI Act, and whatever your sector regulator expects — not a generic maturity score.

A remediation sequence ordered by exposure. What to shut off this week, what to bring under policy this quarter, and what can wait.

What this is not

It is not an AI strategy deck. It is not a model build or a data-platform implementation. It is not a tooling recommendation with a rebate attached — MALA does not sell AI software, and the governance review is deliberately separated from any decision about what you buy next.

If the honest answer is that your exposure is small and your controls are adequate, that is the answer you will get.

Start with the briefings

Two MALA guides cover this ground in depth, and both are free:

Shadow AI: What Your Board Doesn't Know Is Already Costing You — on the AI already in the business, and how it gets there.

You Now Have More AI Agents Than Employees. None of Them Have a Manager. — on agentic AI, non-human identity and what happens when agents inherit access.

Where this connects

Governance findings rarely stay in one lane. Agent and service-account exposure routes into Identity & Access Management Advisory. Control gaps route into Cybersecurity Advisory. Framework and regulatory obligations route into Governance, Risk & Compliance Advisory. Vendor terms that need renegotiating route into Technology Contract Negotiation.

If the answer turns out to be about where AI workloads should physically run, that is AI Infrastructure Advisory — a different question from this one, and a different engagement.

Questions

We already have an AI policy. Isn't that the job done?
A written policy and an enforced one are different things. IBM found that among organizations that did have an AI governance policy, only 34% audited for unsanctioned AI. The policy is the easy half; knowing whether anyone is following it is the half that shows up in a breach report.

How is this different from your Cybersecurity Advisory engagement?
Cybersecurity Advisory assesses your security posture broadly. This engagement is narrower and deeper: it is specifically about AI systems, the data they touch, and the identities they run as. Findings frequently hand off between the two, and we will tell you when the right answer is the other engagement.

Who leads it?
Advisors who have run governance programs inside operating companies, working with MALA's identity and security practices. Not a model-building team, and not a reseller with an AI platform to place.

Do you recommend specific AI tools?
Not as part of this engagement. Separating the governance review from any purchase decision is what makes the review worth having. If a tooling decision follows, it runs as its own engagement with the sourcing disclosure that applies to all of them.

How long does it take?
The inventory and access mapping typically take a few weeks, depending on how many systems are in scope and how much of the estate is undocumented. The regulatory assessment depends on which frameworks apply to you.

What does it cost?
There is no direct fee to you. MALA is compensated by the provider you ultimately select, and never by a vendor whose product we assessed but you did not choose — the full explanation is here, including how to verify it.

Sources: IBM Cost of a Data Breach Report 2025 (600 organizations, March 2024 – February 2025); Palo Alto Networks 2026 Identity Security Landscape.