The model keeps getting better and the results keep stalling. Gartner says the reason most projects get cancelled is management, not capability. The fix is a name against the work.
By Jack Northall
A new frontier model ships most weeks now. Each one is cheaper and more capable than the last. If raw capability were the blocker, results would be climbing with it. They are not.
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027. The culprit it names is escalating cost, unclear business value, and weak controls. In plain terms: management failure, not machine failure.
That should change what you worry about. The model is almost always good enough. The question is who owns the work it touches.
The algorithm and the data plumbing are roughly a tenth of what makes AI work in a business, and they are the part vendors sell. They arrive nearly finished. You can have a capable model running against your CRM in an afternoon. This is the easy tenth, and almost everyone spends their whole budget here.
Who runs the process the AI plugs into. Who checks its output. Who is accountable when it is wrong. Who gets the team to actually use it. That is about seven-tenths of the job, it cannot be bought in a licence, and it is exactly the part that goes unassigned. A project without a named owner for this half is a demo waiting to be cancelled.
Ownership sounds like a job title. In practice it is three answerable questions. A real person has to answer all three before you switch the AI on. If they cannot, you have a pilot, not a deployment.
The AI plugs into a workflow: lead routing, forecasting, follow-up, record updates. Someone has to own that workflow end to end, know how it is meant to run, and decide what good looks like. AI dropped onto an unowned process automates the confusion faster.
A model is confident whether it is right or wrong. Someone has to review what it produces before it counts, catch the misses, and feed the corrections back. That review loop is what earns the right to trust the AI in front of a board.
When a wrong number reaches a client or a partner, accountability cannot sit with software. A named person has to own the outcome, which means the record has to show what the AI did and why. No trail, no accountability, no defence.
A director at a mid-market recruitment firm told us his entire sales process lives in his head. The CRM gets used occasionally and erratically. Fields filled in inconsistently, gaps papered over by memory, the whole pipeline held together by one person who happens to know what all of it means.
Hand an unsupervised AI write access to that CRM and here is exactly what you get: the same gaps, replicated at speed, with no one who can explain why a field changed. The AI inherits the ownership problem. Fix who owns the process first, and the same model suddenly has something solid to stand on.
We build the 70% on purpose. AI suggests and reviews, a person stays in the loop for anything that writes a change to your systems, and every action is logged so there is always a record of what happened and why. No field in your CRM has two writers: if the AI owns a value, a human is not silently fighting it, and if a human owns it, the AI does not overwrite them. Ownership is explicit, so accountability is too. You can read more about how that works on how we work.
The ownership gap is also why most AI fails: an unowned process, not a weak model. Internal AI builds reach production about 22% of the time. With an implementation partner who assigns owners and builds the oversight first, that rises to 67%. Same models. Ownership is the only variable that changed.
Ownership is the only variable that separates a pilot from production.