Every other conversation we have with a new client starts the same way: we want to add AI. Generative models, predictive analytics, recommendation engines — the ambition is real. But in most cases, the platform underneath is not ready. Not even close.
The gap between 'we want AI' and 'our systems can support AI' is where most projects stall or fail outright. Before writing a single prompt or training a single model, there is a brutal but necessary audit that has to happen. We run this for every engagement, and it saves months of wasted effort.
Data Quality: The Foundation Nobody Wants to Talk About
AI is only as good as the data feeding it. That sounds obvious, but the state of most enterprise data is genuinely alarming. We consistently find duplicated records, inconsistent schemas, missing fields, and data that hasn't been validated in years.
The audit starts here. We profile every data source the AI system will touch. We check completeness rates, freshness, schema consistency across environments, and whether there is a single source of truth or five conflicting ones. If your customer table has three different date formats and a 15% null rate on key fields, no amount of prompt engineering will save you.
The standard we apply: every data source must have documented ownership, a defined schema, freshness SLAs, and automated quality checks before it feeds any AI pipeline. If it doesn't, we fix that first.
API Surface Area: Can Your Systems Actually Talk?
Most AI use cases require real-time or near-real-time access to business data. That means APIs. And most legacy platforms either have no API layer, a brittle SOAP interface from 2009, or internal-only endpoints that were never designed for the load AI will put on them.
We map the full API surface: what exists, what's documented, what's authenticated properly, what can handle concurrent requests without falling over. We check rate limits, response times, and error handling. A retrieval-augmented generation system hammering an endpoint that times out after 200 concurrent calls is a production incident waiting to happen.
What we look for specifically: RESTful or GraphQL APIs with proper versioning, OAuth2 or API key authentication, sub-500ms response times on critical paths, and graceful degradation under load. If any of this is missing, we build it before the AI layer goes on top.
Infrastructure: GPUs, Latency Budgets, and Cost Reality
Running inference — even with hosted models — has real infrastructure implications. We assess whether the current setup can handle the latency, throughput, and cost of AI workloads.
For teams using external APIs like OpenAI or Anthropic, we model the cost at realistic scale. A proof of concept processing 50 requests a day looks cheap. That same system handling 50,000 requests a day with 8K-token contexts can blow through budgets fast. We build cost projections tied to actual usage patterns, not demo volumes.
For on-premise or self-hosted models, the conversation shifts to GPU availability, VRAM requirements, and whether the ops team can actually manage model serving infrastructure. Most teams cannot, and that's not a criticism — it's a reason to choose hosted inference and invest the saved effort elsewhere.
Team Readiness: The Human Factor
This is the part of the audit that creates the most discomfort, but it matters more than any technical checkbox. We assess whether the team has the skills and structure to build, maintain, and govern AI systems in production.
Key questions: Does anyone on the team understand prompt engineering beyond copying examples from blog posts? Is there a process for evaluating model outputs systematically? Who owns the AI system when it produces wrong answers — and it will produce wrong answers? Is there a feedback loop from users back to the team?
We are not looking for a team of ML PhDs. We are looking for engineering discipline applied to a new domain: version-controlled prompts, automated evaluation suites, observability on model performance, and clear escalation paths when things go sideways.
The Audit Output: A Roadmap, Not a Report
The deliverable from this process is not a 60-page PDF that collects dust. It is a prioritized, time-boxed roadmap with three lanes: what you fix now, what you build in parallel, and what the AI system looks like once the foundation is solid.
In about half of our engagements, the roadmap pushes AI deployment back by 8-12 weeks while we shore up data quality and API infrastructure. Clients are sometimes frustrated by this. But the alternative — bolting a sophisticated AI layer onto a shaky platform — is how you end up with an expensive system that produces unreliable results and erodes trust in the technology before it ever gets a fair shot.
Do the audit. Fix the foundation. Then build the AI. The order matters.