Most companies can't tell you what their AI bought them.
We can. Fifty years inside service operations and more than 400 operational assessments. We know where AI pays, where it doesn't, and what has to be true before it works at all.
The gap
The money is going in. The return isn't coming out.
Bain surveyed 951 companies across nine industries. Among those that measured their results:
saw cost savings of 10% or less
got past 30%, and most had targeted double what they achieved
are funding the next round of AI out of savings that never arrived
The technology worked. The value didn't.
The diagnosis
It isn't the model.
McKinsey tested 25 organizational attributes against EBIT impact from generative AI. Redesigning workflows had the largest effect of any of them: larger than tooling, larger than talent, larger than spend.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and names the causes: escalating cost, unclear business value, inadequate controls. Model capability is not on the list.
These are operating failures. The wrong process was automated, the scope was wrong, or the metric being optimized was the wrong one.
What the market is saying
The AI industry knows this, and is solving it with half the answer.
Look at what they're hiring. Postings for forward deployed engineers (engineers embedded inside the customer's operation to make the software actually work) have grown several hundred percent in a year. The largest AI companies compete for them at $300,000 to well over $500,000 a head, and in 2026 two of them stood up dedicated deployment ventures. Nobody spends that because they suspect deployment might be difficult. They spend it because they know it is.
But every one of those hires is an engineer. Put a brilliant engineer inside a support operation and they can integrate anything, automate anything, instrument anything. What they can't do is know that the escalation path is the bottleneck rather than handle time. Or that the quality program is measuring the wrong thing. Or that the metric they were asked to optimize is the one that will make the operation worse.
An engineer can't build the right thing if nobody has told them what right looks like.
Our position
PSP is the link between what AI can do and what it actually does.
We've spent fifty years inside service operations and completed more than 400 operational assessments across 30 countries. That is the knowledge an AI build has to be aimed with.
- We know what best-in-class looks like, function by function, because we've measured against it hundreds of times.
- We know the processes that separate a well-run operation from one that merely looks busy.
- We know which metrics matter, and which ones move without anything actually improving.
- We know where the opportunities usually sit, which is why we can assess an operation in weeks rather than months.
Build once. Pay across the operation.
AI gets bought as tools. It should be built by function. Fix forecasting once, properly, and every team that forecasts gets better: support, customer success, implementations and back office, from the same build.
The right order
Assess the operation. Fix what has to be fixed. Then aim the engineering at the function that matters, against a metric that means something, measured from a baseline that existed before the build.
Before you build AI on your operation, find out what it's built on.
A short, principal-led review of where your operation stands against best-in-class, and where AI will pay.
Get in touch