the evidence
Why enterprise AI transformations fail
MIT's NANDA research put a number on the AI version of this in 2025: 95% of enterprise generative-AI pilots produced no measurable P&L impact. The number was treated as news about AI. It is not. BCG studied roughly 900 digital transformations and found only 30% met or exceeded their target value. McKinsey found only 16% both improved performance and sustained the improvement, which is where the widely quoted 84% failure rate comes from.
The rate has held steady across cloud migration, big data, machine learning and now AI. The technology keeps changing and the failure rate does not, which is the first clue that the technology was never the binding constraint. The failures are not random either. They cluster into five modes, and only one of the five is about capability.
- 01
Weak executive sponsorship and leadership misalignment
When the top team is not aligned and visibly sponsoring the change, the organization receives mixed priorities instead of one. Decisions that need a single person take four weeks and three committees. Business units optimise locally, which is the rational response to being told everything matters equally.
This one is rarely fixed later. Sponsorship that has to be won halfway through a programme arrives after the budget is committed and after the people affected have worked out that nobody senior is going to defend the change.
The tell
The named sponsor has delegated the steering committee to someone two levels down. - 02
Unclear case for change and fuzzy value drivers
People cannot say why they are doing it, what success means, or where the value is supposed to come from. Nearly half of the value loss in a transformation happens during target-setting and planning — before execution begins, according to McKinsey's work on the subject. The programme is already losing before anyone builds anything.
Vague targets survive because they are comfortable. A number nobody can miss is also a number nobody can hit, and it postpones the argument about priorities to a point where the argument is far more expensive.
The tell
Nobody in the room can state the single number the programme is supposed to move. - 03
Poor change management
The classic version: the new org, process, or system was rolled out, and adoption did not happen. Prosci's research puts a multiplier on it — initiatives with excellent change management are up to seven times more likely to succeed. A multiplier that large means this is the main event, not a workstream running alongside it.
Resistance in a large organization rarely arrives as a stated objection. It arrives as process: another review, another round of requirements, an audit that has to close first. What is being protected is usually authority or expertise rather than the design you are being asked to defend.
The tell
Training is scheduled after the build, and the change plan is a communications plan. - 04
Weak program governance and no tracking discipline
Enterprise transformations run many workstreams that depend on each other. Without governance, execution drifts and value leaks quietly, because no workstream ever announces that it has stopped delivering the thing it was funded for. It reports that it is on schedule.
The failure is measurement, not discipline. When status is reported as activity — milestones hit, systems delivered, people trained — nothing in the reporting can distinguish a programme creating value from one consuming budget on time.
The tell
Steering reviews report milestones rather than movement in the business metric. - 05
Insufficient resources and capability gaps
The organization does not have the specialised skills the transformation requires, so it buys them. The capability leaves when the engagement ends. What remains is a system nobody inside can extend, documentation written by people who have moved on, and a dependency that has to be repurchased for the next phase.
This is the mode that makes the other four permanent. An organization that never builds internal capability has to start the same argument, with the same outsiders, every time something needs to change.
The tell
The people who understand how the new system works do not work for you.
What addresses each mode
Forward Deployed is organised around these five. Weak sponsorship and an unclear case for change are handled before a contract exists, through customer selection — refusing work where leadership is not aligned and the value driver is not named. Change management is treated as the main event and taught as a practised skill rather than a communications plan. Governance is enforced by the architecture instead of by discipline, so tracking is a property of the system rather than a meeting. And capability is injected directly into the organization by an engineer who owns the problem end to end, rather than rented and withdrawn.
The book uses Palantir as the case study because it is the clearest available example of a company that made this repeatable and profitable. It is an independent analysis, not authorized or sponsored by Palantir.
Sources
MIT NANDA initiative (2025) — 95% of enterprise generative-AI pilots delivered no measurable P&L impact, from 150 leader interviews, a survey of 350 employees and 300 public deployments. BCG, Flipping the Odds of Digital Transformation Success (2020) — 30% of roughly 900 transformations met or exceeded target value. McKinsey, Unlocking Success in Digital Transformations (2018) — 16% both improved performance and sustained it. McKinsey, The science behind transformations — value loss concentrated in target-setting and planning. Prosci, Best Practices in Change Management (12th edition) — excellent change management up to 7× more likely to succeed.
Forward Deployed maps each of the five failure modes to the part of the operating model that addresses it — customer selection, deployment discipline, the forward-deployed engineer, and the platform architecture underneath all of it.