Skip to main content

If Your Data and AI Initiatives Keep Stalling, We Should Talk

When AI pilots and data projects keep getting stuck, the cause is usually something specific in how data is defined, owned, and accessed. A focused diagnostic can find it and tell you what to fix first.

By Marco Porracin
Model: Claude Opus 5.5 (On Cursor)
#data strategy#AI readiness#data governance#analytics engineering#self-service analytics

If your company keeps starting data and AI projects that never quite make it into day-to-day work, we should probably talk.

You likely have the people for it. Engineers who know the systems, analysts who can answer hard questions, leaders with a list of things they'd like AI to do. You may also have a warehouse, a BI tool, more dashboards than anyone uses, and a few pilots running.

And still, things get stuck in the same places. A use case waits weeks for someone to pull the data. A dashboard needs a manual reconciliation before anyone will present it. An AI prototype works on someone's laptop and never gets deployed. Most questions still go through the same three or four people.

It's easy to read that as an execution problem and push the team to go faster. In our experience the cause is usually more specific: some part of the foundation the project depends on isn't there yet, and nobody has said out loud which part.

Owning the Tools Doesn't Mean You're Ready

The first instinct when assessing readiness is to list the stack. Cloud warehouse, transformation tool, catalog, vector database, access to an LLM.

Those things help, but plenty of companies have all of them and still can't produce a revenue number two teams agree on, rebuild last month's model inputs, or let the sales team look up customer data without opening a ticket.

What matters is whether you can turn data into decisions and products over and over. That depends on a few things that aren't software:

  • Someone has picked the decisions and workflows that are worth improving.
  • The data those workflows need can be found, accessed, and joined.
  • Key entities and metrics have written definitions people can look up.
  • When data is missing, late, or wrong, it's clear whose job it is to fix it.
  • Access rules are designed up front instead of negotiated per request.
  • There's a known way to take an experiment and run it in production.
  • Business users can work with approved data without filing a request every time.

If one of these is missing, a new tool usually won't help. It just adds another system to maintain.

Everyone Sees a Different Piece

These gaps show up differently depending on where you sit. Leadership sees a roadmap that keeps slipping. Business teams see reports they can't open or don't believe. The data team sees a request queue that never shrinks. Engineers get vague requirements and brittle integrations. Security gets asked to grant broad access to sensitive systems for a tool they've never heard of.

Each group fixes the problem in front of it. Leadership pushes for speed. Business teams build spreadsheets. Analysts redefine metrics inside individual reports. Engineers add one more export. Security tightens access. The next AI pilot starts outside the usual process, because the usual process is what slowed down the last one.

None of these fixes are unreasonable, but they pile up around the real constraint instead of removing it. After a while people learn the workarounds: which dashboard needs a manual adjustment, which analyst understands a certain table, which spreadsheet has the number finance accepts. The business keeps running, and the underlying problem gets harder to see.

That's why reading the backlog won't tell you much. The team is busy. The trouble is that the way work flows turns every request into a custom job.

AI Makes Weak Foundations Obvious

AI still needs good data, and it tends to expose bad data faster than a dashboard does.

A model can summarize a document or classify a ticket in seconds. Putting that into production still requires the right context, permissions, a way to evaluate the output, monitoring, and a place in an actual workflow.

Some examples. If a customer has a different ID in each system, an assistant can't build a reliable view of that customer. If policy documents have no owner or version, retrieval will happily return last year's policy. If nobody agrees on what counts as an active account, an AI-written explanation of churn can sound perfectly reasonable and be computed on the wrong set of customers.

Often the model is fine and the problem is the data and processes around it.

You don't need a big platform before you try AI. But the first production use case should be picked with your current foundation in mind, and the gaps that specific use case needs closed should be written down. Part of the job is deciding what has to be fixed now and what can wait.

What a Readiness Diagnostic Should Look At

Some people call this a readiness audit. Whatever the name, a maturity score plus a reference architecture diagram doesn't count. The useful version follows real decisions and workflows through the company, checks the assumptions behind them, and ends with a list of things to do in order.

In practice that means looking at these together:

  1. Outcomes and decisions: Which decisions, customer experiences, or internal processes should get better? Who uses the result, and how will you know it worked?
  2. Sources and flows: Where does the data come from, and how does it move today? Which manual files, exports, and undocumented integrations are part of the real process?
  3. Definitions and quality: Which entities and metrics matter most? Where are they defined? Can you test freshness, completeness, consistency, and history?
  4. Architecture and operations: Can the platform handle the volume, latency, reliability, and cost you need? How are changes deployed, monitored, and rolled back?
  5. Governance and access: Who should see what? How are sensitive fields protected? Can access grow without people emailing files around?
  6. Team and ownership: Who owns each source, model, data product, and business definition? Where do handoffs slow things down?
  7. AI delivery: How do use cases get picked, evaluated, integrated, monitored, and reviewed? What happens when the model or its context is wrong?

The output should show where the constraints are, how they connect, and what order to tackle them in. It should also say what's already working. Most internal teams have made plenty of good calls, and those are worth keeping. The aim is to remove friction that doesn't need to be there and spend where it actually speeds up delivery.

What You Should Get Out of It

A diagnostic is only worth it if it changes what you do next.

That means a short list of prioritized actions tied to specific outcomes. Some are technical, like building a reliable customer model or adding tests to a pipeline the business depends on. Others are organizational, like naming an owner for revenue definitions or setting up a review step for AI use cases.

Order matters a lot here. Build an assistant before you've sorted out access to its source material and you get a nicer interface on top of an unreliable process. Roll out a catalog before anyone owns the data and you've documented the confusion. Open raw database access in the name of self-service and you add risk without making anything easier to understand.

A sequence that tends to work:

  1. Pick one important workflow or decision that is stuck today.
  2. Map the data, definitions, people, and systems it touches.
  3. Figure out the smallest foundation needed to make it reliable.
  4. Deliver it with quality checks, access rules, and ownership included.
  5. Reuse that foundation for the next area instead of starting over.

The business sees a result early, and the platform gets better underneath it.

More People Should Be Able to Use the Data

Long term, you want more people using reliable data on their own, within clear limits, so the data team isn't the help desk for every question.

Nobody outside the data team needs raw tables. They need to find the data that fits their role, understand what it means, and trust that the important definitions are the same for everyone.

If every question still needs a specialist, access hasn't been solved. We wrote about this in If You Have to Ask Engineering for Every Report, We Should Talk: engineering should build a dependable path to answers instead of acting as the company's reporting API.

So a diagnostic shouldn't stop at whether the pipelines run. It should check whether access can grow:

  • Can a business user find an approved dataset?
  • Can an analyst see its grain and where it comes from?
  • Does the owner find out when quality drops?
  • Can security apply policies without reviewing every report by hand?
  • Can a new use case reuse definitions and controls that already exist?

If the answers are yes, you're in better shape for AI, and the company also gets better at learning from its own data.

Why Bring In Someone From Outside

The people closest to the systems know things nobody else does: why certain exceptions exist, which migrations are halfway done, which "simple" change will break production.

They also work inside the current setup every day, which makes some of its problems hard to notice.

An outside team can trace dependencies across business, data, engineering, and security without having a stake in any particular tool or backlog. It can surface work nobody tracks, compare how things are supposed to work with how they actually do, and give leadership and the delivery team the same picture.

What it shouldn't do is hand over a document the internal team then has to figure out how to implement. Recommendations need evidence, a realistic scope, stated tradeoffs, and a plan the team can take over. When it goes well, leadership knows what has to change, the technical team has a sequence it can defend, and business users know which access improves first.

Start With a Diagnosis

Stalled data and AI projects don't automatically mean you need a new platform, a bigger team, or a company-wide transformation program. They mean you should find out what is stopping good ideas from turning into things people use.

Sometimes the answer is agreeing on one business definition. Sometimes it's fixing a data flow, moving an ownership boundary, adding governance, changing how work gets delivered, or narrowing the scope of an AI use case. Usually it's two or three of these, done in the right order.

If your roadmap keeps moving but access to reliable data doesn't, talk with Blueprintdata. We'll find the constraints, pick the smallest changes that matter, and leave you with a plan your team can run with.

RSS Feed

Prefer RSS? Subscribe to our RSS feed to get updates directly in your feed reader.

Subscribe to RSS

Want to talk through a similar data problem?

If this is close to the kind of work your team needs, request a conversation and tell us what you are trying to solve.

Request a Conversation