In the modern enterprise, “AI-everywhere” has turned into a mess of technical debt, runaway costs, and security blind spots. If you’re already juggling PLM, ERP, and MES systems competing for budget, the unchecked spread of AI subscriptions is a real threat to whatever governance model you’ve built.
This was a lesson learned through challenging circumstances while consolidating an AI stack across a vendor estate that had grown to 120 separate tools, part of a broader ~£5M technology budget covering everything from PLM and ERP to infrastructure and SaaS. The conclusion I kept coming back to: AI governance without vendor discipline is really just a policy document nobody follows. By the end of the exercise, we’d brought the annual AI subscription run-rate down from £50k to £20k without losing any real capability. Here’s what that process taught me about where governance actually breaks.Shadow AI creeps in department by department.
When procurement is decentralized, every team ends up treating AI as its own innovation project. Engineering picks up a niche LLM tool for CAD work. HR pilots a resume screener. Marketing signs up for a generative copywriting tool. None of these looks expensive on its own. Add them up across 120 vendors, though, and you’re looking at £50k a year in AI subscriptions alone that nobody actually signed off on as a whole.
The shadow AI cost trap
The part that worried me more than the cost was what happens without a gatekeeper: a lot of these tools are sending data to third-party APIs that were never vetted against the compliance bar you’d expect for a safety-critical or IP-heavy environment-GDPR, ISO 27001, whatever your baseline is.
Governance keeps missing the actual risk.
Most of the public conversation around AI governance-NIST’s AI RMF, the EU AI Act-gets framed around fairness and bias, and that’s part of what they cover, but both frameworks go considerably wider than that headline. In my experience, the operational risk inside a large enterprise is vendor management. Governance fails in practice because we keep treating AI tools as ordinary SaaS subscriptions instead of as integrated systems touching real data.
What actually worked was tying governance directly to procurement. No tool gets deployed unless it clears a vendor risk assessment against our internal security baseline-I think of this as the audit-first rule, and it sounds obvious until you see how many tools sail through without it otherwise. The second piece was harder to enforce but mattered just as much: Do you really need five tools that are all wrapping the same underlying model API?What the audit actually found.
When we went through the full vendor landscape, the numbers were rougher than I expected. Something like 60% of our AI spend (which lines up with that £50k-to-£20k annual cut) was tied to redundant or overlapping subscriptions, a fair amount of it shadow IT that nobody in procurement even knew existed.
The approach was straightforward, if a bit unglamorous:
Action
Impact
Subscription cleanup
Identified and cut 40+ overlapping “AI assistant” tools, most used by fewer than 5 people
Centralized identity (MFA/SSO)
Required every AI vendor to integrate with corporate SSO; the ones that couldn’t were dropped
Policy enforcement
Blocked unauthorized AI browser extensions and shadow APIs at the network layer
The 4-question governance testWhere I Landed
I don’t think the goal here is to slow innovation down-it’s to make it accountable. Moving from a subscription-first habit to something closer to an architecture-first one is what actually matures an AI program, in my opinion.
If you don’t control the vendor, you don’t really control the data, and if you don’t control the data, the governance framework on paper isn’t doing much. The conclusion I keep coming back to after this project: discipline is the part that decides whether the rest of it means anything.
If you want a starting point, pull the last 90 days of invoices and go through them one by one. Any tool you can’t explain in terms of your actual product lifecycle or Digital Thread is one you’re paying for as risk.
Source link
Leave a Reply