Someone on the board asks the CTO a simple question: are we ready for AI? The CTO says yes. The company has cloud infrastructure, a data warehouse, a few dashboards, so the answer feels true. Three months later the pilot fails quietly. The model worked in testing and fell over in production, and nobody can say exactly why. I have watched this play out across more than 200 portfolio companies, and the honest answer to “are we ready for AI” almost never lives in the model. It lives in the data foundation underneath it. Five signals tell you the real answer before you spend a dollar.
Sign 1: Your data lives in silos that nobody fully owns
The first sign is fragmentation. Customer data sits in the CRM, revenue lives in the billing system, operational numbers hide in spreadsheets, and the BI tool pulls from a fourth place that disagrees with all three. There is no canonical source, so every important number has a few slightly different versions depending on who you ask and which system they checked.
The AI symptom is predictable. A model trained on dirty, inconsistent inputs produces unreliable outputs, and you cannot trust an answer you cannot reconcile. “Garbage in” is not a slogan here. It is the daily reality of an agent acting on data that three teams define three different ways.
The test is simple, and it tends to embarrass people. Can any data engineer on your team describe the lineage of your five most important metrics without opening four different systems? If tracing “revenue” or “active customer” turns into a scavenger hunt across tools, your data is not ready to train anything. Fix ownership and stand up a canonical source first. AI built on contested numbers inherits every disagreement buried underneath it.
Sign 2: You’re still moving data with scripts, not pipelines
The second sign is how data actually moves through your company. If the answer is a pile of manual ETL, Excel-based transfers, and cron jobs nobody monitors, you have a stack held together by individual heroics. It works right up until the one person who understands the Tuesday load goes on vacation.
The AI symptom shows up the moment a source format changes. Inference pipelines fail silently, the model keeps returning answers, and the answers are quietly wrong because an input shifted and nothing caught it. Manual movement has no alerting, no retries, no contract for what the data should look like when it arrives. That is survivable for a weekly report. It is fatal for an agent making decisions every few minutes.
The test: count the ad-hoc data movement scripts your team maintains, then ask who gets paged when one breaks. If the number is high and the answer is nobody, you do not have pipelines. You have a backlog of silent failures waiting for a model to amplify them.
Sign 3: You can’t explain where your numbers come from
The third sign is governance, or the absence of it. No data catalog, no lineage tracking, no documentation of how a number gets produced. The institutional knowledge lives in a few people’s heads, and when they leave, it walks out the door with them.
The AI symptom is that you cannot validate model outputs because you cannot trace the inputs. When a model produces a confident, wrong answer, and they all eventually do, you need to follow that answer back to its source data and find the break. Without lineage you are guessing.
The test I use: if a model produces a wrong number, can a data engineer trace the error back to its source data in under 30 minutes? If the honest answer is hours, or no, then you cannot operate AI safely, because you cannot debug it. Governance is not bureaucracy here. It is the difference between an AI system you can fix and one you have to switch off and apologize for.
Sign 4: AI/ML work lives in notebooks, not production systems
The fourth sign is where your AI work actually runs. A Jupyter notebook on a data scientist’s laptop is a pilot. Production is a monitored, versioned, retrainable service that someone owns and gets paged for. These are completely different things, and the gap between them is where most AI initiatives die.
The AI symptom is the line every executive has heard: the model works great in testing, but nobody can run it at scale. The notebook had clean sample data, a human in the loop, and no uptime requirement. Production has none of those luxuries. Recent MIT research found that roughly 95% of enterprise generative AI pilots never deliver measurable returns, and the failure rarely traces to the model. It traces to the missing path from notebook to production.
The test: what percentage of your ML experiments have actually deployed to a monitored production system? If the answer is close to zero, you are running a research project, not an AI capability, no matter how good the demos looked.
Sign 5: Your team is adding data scientists but skipping data engineers
The fifth sign is who you are hiring. Companies under pressure to “do AI” hire data scientists, because that is the title that sounds like AI. Then they skip the data engineers who build the infrastructure to serve, monitor, and retrain what those scientists produce. The result is a team that can model but cannot ship.
Data scientists end up spending the large majority of their time on data preparation rather than modeling, because the plumbing they need does not exist, so they build it by hand, badly, between experiments. A widely cited survey put data preparation and cleaning at roughly 80% of the data-science workload. That is expensive talent doing janitorial work the infrastructure should be handling.
The test is a ratio: for every data scientist you employ, how many data engineers do you have? If you have five scientists and one engineer, your models will keep getting built and keep failing to deploy, because nobody owns the path to production. At this stage the bridge role matters more than the modeling role.
What to do next
None of these signs means “do not do AI.” Each one means “fix the foundation first.” That distinction matters, because the wrong reaction to a failed pilot is to go buy a better model, and the better model will fail the same way for the same reason. The right reaction is to close the specific gap the signs exposed: a canonical source, real pipelines, lineage you can trace, a production path, and the engineers to build and own it.
If you want a complete diagnostic instead of five gut checks, read AI Readiness Assessment: A Framework for Enterprise Data Teams, which scores all five areas in depth. And if you want a fast read on where you stand right now, take the Blueprint assessment. It is free, returns an instant score benchmarked against more than 200 PE-backed companies, and tells you which gap to close first.
Take the free Blueprint AI Readiness Assessment
FAQ
How quickly can you fix these issues?
It depends on the sign. Hiring a data engineer or standing up monitoring takes a quarter. Consolidating siloed data and building real governance across a company takes two to three quarters of focused work. None of it requires a frontier model, and most of it pays for itself before any AI ships, because the same foundation makes ordinary reporting faster and more trustworthy.
Are all five signs disqualifying?
No. One or two signs mean you have specific gaps to close before you run a scoped pilot. Four or five signs mean you are not ready to invest in AI yet, and the honest move is to fix the foundation before you spend a dollar on models.