Why AI transformation fails: nobody knows how the work actually happens
AI projects stall because companies automate processes nobody has described. The fix is discovery before tooling: find the real process, then build.
Most AI transformation fails because a company tries to automate a process it has never described. The manual describes one process, the shared inbox runs another, and the model gets pointed at the first. The fix is not a better model. It is finding out how the work happens before deciding what to build.
The failure rate is now a matter of record
IBM's survey of 2,000 CEOs (IBM Institute for Business Value, 2025) found that only 25 percent of AI initiatives had delivered the expected return. In the same survey, 64 percent of CEOs said the risk of falling behind drives investment in some technologies before they understand the value.
BCG's survey of 1,000 executives (BCG, 2024) put the share of companies yet to show tangible value from AI at 74 percent. MIT NANDA's 2025 study, as reported by Fortune (2025), found that about 5 percent of pilots achieved rapid revenue acceleration. The rest stalled with little or no measurable effect on the P&L.
The work that matters was never written down
Every company has two process documents. The first is the SOP folder. The second lives in the heads of the people doing the work: the discount that is officially approved by the sales director and actually approved by whoever answers Slack first, the spreadsheet that is the real inventory system.
A YouGov survey of 1,001 US employees at companies of 200 or more people (Panopto, 2018) found that 42 percent of institutional knowledge was unique to the individual holding it, and that the average employee spent 5.3 hours a week waiting for help or information from colleagues. The figure is from 2018 and we have not found a newer one. Nobody measures the second document.
That second document is what an AI project needs. The first is what it usually gets.
The technology is the small part of the problem
PwC's predictions for 2026 (PwC, 2025) put it plainly: technology delivers about 20 percent of an initiative's value, and the other 80 percent comes from redesigning work. BCG's prescription, drawn from the 26 percent it classes as AI leaders, is about 70 percent of effort into people and processes and the rest into technology and algorithms.
The MIT researchers, in Fortune's account, describe the cause as a learning gap: generic tools do not learn from or adapt to how a company's workflows run. You cannot redesign 80 percent of something you have not seen, and a tool cannot adapt to a workflow nobody has written down.
Automating an undescribed process gives you a faster version of the confusion
When the build starts from the manual, the same pattern repeats across functions. The rows below are illustrative, not results from any company.
| Process | The manual says | What tends to happen instead | What a build from the manual would do |
|---|---|---|---|
| Invoice approval | Two-step approval by amount | Finance calls the buyer because the PO is often wrong | Approve faster against the wrong PO |
| Customer onboarding | Form, KYC check, account | Compliance skips fields for known clients on request | Reject the clients the team used to wave through |
| Discount approval | Sales director above a threshold | Whoever answers first in the deal channel | Enforce a threshold nobody used and stall deals |
| Stock reorder | ERP suggests, planner confirms | Planner keeps a private sheet because ERP lead times are stale | Reorder on stale lead times |
Builds like these do not fail loudly. They pass the pilot, because the pilot tests the manual's process, and then they get quietly worked around.
Our view
Our view, and it cuts against how most AI programs are set up: the first workstream should be discovery, and the tooling decision should wait for it. Most programs start with a vendor demo and a steering committee, and the process gets inferred from the org chart.
Discovery means hearing the people who do the work, all of them, privately, and keeping the contradictions between teams instead of averaging them into a tidy diagram. The contradictions are where the money is. We compared this approach with process mining in What is AI process discovery?. Transcript was built for this: hundreds of private conversations at once, matched against each other, turned into a world model of how the company actually runs, with each problem valued in the company's own hours and cost per hour.
Questions people ask next
We have documented processes. Does this still apply?
Almost always. The documentation describes the intended process. Ask three people on the same team to describe one handoff and compare the answers.
Does process mining solve it?
Partly. It reconstructs flows from system logs and is precise about what touched a system. The phone call that unblocked a shipment and the email chain that is the real approval are not in the log.
How long does discovery take before we can build anything?
Each person spends five to ten minutes a day during the interview period, and everyone is interviewed in parallel, so the elapsed time depends on how quickly people answer.
Will people describe their workarounds honestly?
Only if their manager cannot read what they said, and that has to be enforced in the database rather than promised in a policy. Remove the privacy and you get the official process restated.
Lightbloom AI starts by finding out how the company actually works, then builds on what it finds. Transcript runs that discovery: /transcript.
References
- IBM Institute for Business Value, IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles (2025), https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles
- BCG, AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value (2024), https://www.bcg.com/press/24october2024-ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value
- Fortune, MIT report: 95% of generative AI pilots at companies are failing, reporting on MIT NANDA, The GenAI Divide: State of AI in Business 2025 (2025), https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- Panopto, Inefficient Knowledge Sharing Costs Large Businesses $47 Million Per Year (2018), https://www.prnewswire.com/news-releases/inefficient-knowledge-sharing-costs-large-businesses-47-million-per-year-300681971.html
- PwC, 2026 AI Business Predictions (2025), https://www.pwc.com/us/en/tech-effect/ai-analytics/ai-predictions.html
Keep reading.
All notes- How do you capture tribal knowledge before someone leaves? Tribal knowledge is how the company actually runs, held by the people who run it. Why handovers fail, what works, and why it is the first layer of a world model. Read →
- Why AI agents need a world model of your company AI agents stall inside companies because nobody has described how the work runs. What a company world model gives an agent, and why to build it before you deploy. Read →
- What is a company world model? A company world model is a living picture of how a company really works: every process as it runs, what hurts most, and who everything depends on. Why AI agents need it. Read →