Insights
Most AI projects fail in production. Our research on why points somewhere most leaders haven't looked.
New Sinch research surveying 2,527 senior decision-makers reveals something most leaders haven't seen yet: more governance maturity correlates with more rollbacks, not fewer. The real answer is somewhere else.
The AI isn't the problem. The architecture around it is.
A new piece of research from Sinch surveyed 2,527 senior decision-makers across ten countries about how enterprise AI is actually going.
The headline finding: 74% of enterprises have rolled back or shut down an AI agent after it went into production.
Three quarters of all enterprise AI deployments. Failing.
That's a stat worth sitting with for a moment.
But the part that should make any leader thinking about AI sit up straight is hiding in the second tier of the data. Of the organisations with the most mature governance frameworks, the rollback rate isn't lower. It's higher: 81%.
In other words, the better your governance, the more often your AI is being pulled out of production.
The Sinch interpretation
The Sinch report is generous about this. Their explanation: "Higher rollback rates reflect better monitoring and control, not weaker performance." Mature organisations don't fail more, they just see the failures sooner.
That's true. And it's also missing the harder, more important point.
The AI was failing the whole time. The governance framework just told you about it.
A compliance framework is a scoreboard. It tells you whether something is going wrong. It doesn't stop the thing from going wrong. (I wrote about that here, if you're curious.)
The guardrail tax
The Sinch report names this directly, though in their language. 84% of AI engineering teams are now spending at least half their time building safety infrastructure. Sinch calls it the guardrail tax.
It's the cost of every business trying to solve the operational layer themselves, from scratch, alongside the customer-facing AI they actually want to ship.
That's a huge bill. Half your AI engineering capacity, spent reinventing locks, keys, prompt inspection and audit logs, when those things already exist in your Microsoft 365 or Google Workspace tenant and could be plugged into instead of rebuilt.
What we built ORCA Opti for
We did the operational layer once, properly. Identity, access control, prompt inspection, AI Guardian, sovereign hosting and audit logging all sit inside the same product, on the security foundations of Microsoft 365 and now Google Workspace.
Then we did the part most leaders haven't been told about. We made it so the same Safe Zone that protects how your team uses AI also protects the AI agents and chatbots your team builds on top of it.
That second part is where the design gets interesting. Most AI security products on the market today protect one thing: either the chat surface your team types into, or the agents you've deployed for customers. Almost none protect both, on both platforms, with the same identity, the same data-sovereignty boundary, and the same audit trail. The strongest dedicated guardrail products are API-call priced, so the bill scales with usage. The closest integrated rival is Microsoft-only and per-user priced.
ORCA Opti is none of those things. It's per-organisation. It covers Microsoft 365 and Google Workspace. And the Safe Zone protects the AI you use AND the AI you build, all the way through to the audit log.
The Safe Zone, in plain language
If you build an internal chatbot with Opti Assist to answer policy questions, ORCA AI Guardian inspects every prompt before it reaches the model, blocks data leakage in real time, logs every interaction, and includes red-team testing against known attack vectors. None of that requires you to buy three separate products.
If your finance team uses Opti Assist directly to draft commentary against this month's numbers, the same Safe Zone applies. Identity stays in your tenant. Data stays in your country. Audit stays in the system of record that compliance can actually pull a report from.
That's what we mean when we say groundbreaking. Not the individual pieces, which are competitive in their own right. The combination, in one product, on the platforms your business already runs on. (Kat's written about why building on something you control matters if you want the strategic version.)
If you're investing in compliance in 2026
The Sinch report says 76% of organisations are increasing investment in trust, security and compliance this year, against 98% increasing AI investment overall.
The question worth asking is this: where does that investment actually deliver protection?
In a framework document that tells you when something has already gone wrong?
In five separate point products that you wire together yourself?
Or in an operational layer that stops the failure happening in the first place, that protects the AI you use and the AI you build, with the safety infrastructure already there?
One we can sell you, working today, on the platforms your business already runs on. The other you'll spend half your engineering capacity on, year after year, hoping eventually it's enough.
Source: The AI Production Paradox, Sinch, May 2026 — survey of 2,527 senior decision-makers across ten countries.