Companies abandon AI stacks, 40% projects cancelled. Seek specialist AI firms to deliver results.

Most agentic programmes stall while the company is still assembling the stack. The ones that ship tend to buy a finished workflow from a firm that already runs it, at a lower total cost than standing up the same stack, hiring, and pilot cycle in-house.

40% of the projects get cancelled before they become a result

Gartner's 25 June 2025 note is a cancellation forecast, and the causes are ordinary. Cost runs away. Nobody can say what the business got. The risk controls are incomplete. Verma, a senior director analyst at Gartner, said most agentic projects right now are proofs of concept driven by hype and aimed at the wrong job. Integrating an agent into a system the company already runs is messy, and it often breaks the workflow you were hoping to speed up.

The vendor count is the part a buyer should memorize. Thousands of companies now sell an "agent." Gartner thinks about 130 of them are actually that. The label has been stuck on assistants, old automation, and chat windows. A stack diagram with six logos and nobody accountable for a named output is a diagram purchase, and those are the programmes that struggle to survive the next budget cycle.

I have sat in the meeting where a pilot with no named output loses the next budget. Gartner is telling you that meeting arrives by the end of 2027, and a large share of these projects will not survive it.

Specialists reach deployment about twice as often

MIT's NANDA group wrote The GenAI Divide from interviews, a staff survey, and a few hundred public deployments. The line I use is in the report itself. External partnerships, with tools that can be fitted to the work, reached deployment about 67% of the time. Tools the company built and maintained itself reached deployment about 33% of the time. Partnership pilots were about twice as likely to arrive fully deployed, and employees actually used the outside tools more often. Fortune's 18 August 2025 piece carried the same split to CFOs: buy from a specialist, or build it in the building, and the build loses.

The report is careful. The 67 and the 33 are what people reported, and other factors could be mixed in. The direction was consistent across the interviews. Companies that insist on owning every model, every prompt, and every integration are paying for control they will not finish. Companies that hire a firm whose only job is the workflow get to a result while the internal team is still drawing the architecture.

That is the specialist firm in the headline. You are buying the output of an AI workflow from people who already run it, priced against the cost of standing up the same stack, the same hiring, and the same failed pilot inside your own walls.

Forward-deployed engineers are how the specialist does the work

A forward-deployed engineer is an engineer who embeds inside the customer's operation. They sit with the people who already do the job, learn the real handoffs, connect the tools into that workflow, and leave software the operator uses on a Tuesday. The labs started hiring them when it became obvious the miss was in deployment. Palantir ran this model for years, because a platform that never enters the building does not change the building. The job descriptions put the split at roughly 40% code and 60% customer. An AI vendor that never meets the person who does the work is selling a slide deck, and Gartner's 40% is the forecast for programmes that never leave the slide.

I sell this service, so I will state it as a commercial fact and you can price it against the alternative. Kitsune puts model-side people and forward-deployed engineers on the workflow. The deliverable is the running workflow and the result the operator already needed. Keeping a bench of generalists to own an internal AI stack is the build path MIT watched reach deployment about half as often as the partnership path.

Pakistan already knows this shape, from transformations that were not called AI

I ran media and digital at Nestlé Pakistan from 2018 to 2023. The savings that stuck were PKR 6 billion on television and PKR 3 billion on digital, and they stuck because we changed the buy, the data, and the people in the process. Real efficiency came from rewriting the operating workflow and partnering on outputs, with specialists where we did not have the muscle. A pitch house, a data partner, a measurement study. The output had a number. The committee could argue with the number. It could not fund a permanent internal science project and call it transformation.

That is the read I bring to a Karachi or Lahore board that has been told to in-house an agent platform. You already know how a transformation dies. It dies as a tool the centre built and the line will not open. The 67% path is the same path as hiring the firm that has done the workflow in someone else's building, then putting a forward-deployed engineer next to your operator until the result shows up in the weekly numbers without ceremony. A cancelled project is what you get when the programme was measured by the stack it assembled rather than by the output it delivered.

By the end of 2027 a lot of those stack programmes will be written off. The firms left standing will be the ones who sold a result and stayed in the room until it showed up in the numbers.