AI-ready data: What Mid‑Market CEOs Must Audit Before Funding an AI Project

When I saw the recent Show HN: A Comprehensive AI Data Quality Evaluation Tool post on Hacker News, my first thought was how many CEOs I’ve spoken to would try to buy that tool without first checking if their data could ever be *AI‑ready*. The excitement around generative models often masks a simple truth: if the data feeding the model is noisy, incomplete, or mis‑aligned with business outcomes, the model will deliver noise, not insight. In my experience as an operator‑advisor, the moment a mid‑market company tries to bolt AI onto an existing data stack is when the first cracks appear.

In this essay I’ll walk through what *AI‑ready data* really looks like for a company with a few hundred employees, the audit questions you should ask before green‑lighting any AI spend, and a concrete five‑day plan to start fixing the gaps you discover. This is not a trend piece; it’s a checklist you can run on Monday and a set of actions you can execute by Friday.

The moment we hit a wall

A recent client – a regional logistics firm with $80 M in revenue – wanted to use a large language model to automate freight‑booking emails. They hired a vendor, fed the model a dump of their CRM export, and expected a 30 % reduction in manual effort within weeks. The model churned out garbled responses, missed key contract clauses, and occasionally suggested shipping routes that violated internal compliance rules. The vendor blamed “data quality” but the CEO had already signed a $250 k contract.

What went wrong?

1. No single source of truth – the CRM export was a mix of manually entered notes, legacy spreadsheet rows, and a few API pulls. Duplicate customer records and missing fields were the norm.
2. No data governance – there was no documented definition of what a “valid address” meant, nor any process for keeping it up to date.
3. Business‑logic drift – the rules that determine a valid freight quote lived in a Word document that hadn’t been updated in two years.

The model was fed data that was technically *present* but practically *useless* for the task at hand. The result: a costly pilot that delivered nothing but a fresh set of excuses.

Defining AI‑ready data beyond the hype

For a mid‑market firm, *AI‑ready data* is not a massive data lake or a petabyte‑scale warehouse. It is a disciplined, business‑aligned data set that meets three concrete criteria:

1. Completeness for the target use case – every record required to answer the specific question the AI model will address must be present. If you are automating contract review, every contract clause, amendment, and expiration date must be captured.
2. Consistency and correctness – data values follow a single, documented schema. Dates use ISO‑8601, monetary values are stored in a base currency, and identifiers are unique across systems.
3. Contextual relevance – the data includes the metadata that tells the model *why* a record matters. For example, a shipment record should carry the business unit, priority tier, and any regulatory flags.

These criteria are deliberately narrow because they keep the scope manageable. You do not need to ingest every log file or sensor stream; you need the *right* data, clean and well‑documented, for the *right* problem.

Audit checklist for CEOs before funding an AI project

Below is a list of questions you can ask your CFO, CTO, and data owners in a single 30‑minute meeting. Treat the answers as a go/no‑go gate rather than a formality.

If you cannot answer more than half of these confidently, pause the AI spend and allocate resources to data remediation first.

Monday‑to‑Friday: practical steps to get there

Below is a five‑day sprint you can run with a small cross‑functional team (product manager, data engineer, and a business analyst). The goal is to produce a *data readiness report* that either clears the path for AI or surfaces the remediation work needed.

Day 1 – Scope & baseline

1. Define the use case in a one‑sentence statement and capture the KPI.
2. Identify the source systems and pull a 10 % random sample of records.
3. Run a quick data‑quality script (even a spreadsheet pivot) to surface missing values, duplicate keys, and format mismatches.

Day 2 – Ownership & governance

1. Assign data stewards for each source system.
2. Document field definitions in a shared Confluence page – include format, allowed values, and business meaning.
3. Create a data‑issue backlog in Jira/Asana with severity tags.

Day 3 – remediation pilot

1. Pick the top‑5 data issues (e.g., missing address line, duplicate customer IDs).
2. Write simple scripts (Python/pandas or SQL) to clean the sample data.
3. Validate the cleaned sample against the KPI baseline – does the cleaned data improve the model’s confidence? (Even a rough test with a small LLM prompt is enough.)

Day 4 – governance workflow

1. Build a lightweight data‑quality dashboard using Metabase or an internal BI tool – show completeness % and duplicate count.
2. Define a change‑request process: any new field or rule must go through the steward and be logged.
3. Schedule a weekly sync with the AI vendor to share the dashboard and get feedback on data needs.

Day 5 – decision gate

1. Compile the data readiness report – include the KPI baseline, remediation cost estimate, and a risk matrix.
2. Present to the executive sponsor (you) and decide: proceed with AI spend, extend the remediation sprint, or reconsider the use case.

By the end of the week you will have a concrete answer to the audit questions above, and you’ll have demonstrated to your board that AI spend is disciplined, not speculative.

FAQ

What if my company already has a data lake?

A data lake can be a source of truth, but for mid‑market firms it often becomes a dumping ground. Focus on the *cataloged* and *governed* subsets that directly support the AI use case.

How much budget should I allocate to data remediation?

There is no one‑size‑fit number, but a good rule of thumb is to keep remediation costs under 30 % of the total AI project budget. If remediation exceeds that, revisit the scope.

Do I need a full‑time data scientist to run this audit?

No. A data engineer with basic SQL/Python skills and a business analyst who understands the KPI can execute the five‑day sprint.

How often should I repeat this audit?

Treat it as a living document. Re‑audit whenever you add a new data source, change the AI use case, or after a major system upgrade.

Can I outsource the data‑quality work?

You can, but ensure the vendor provides transparent logs and that a trusted internal steward retains final sign‑off.

Next steps

If the audit reveals gaps, the fastest way to close them is to start with the five‑day sprint outlined above. If you already have a clean data set, the next step is to prototype the model on a small, controlled batch and measure the KPI improvement before scaling.

I’m happy to walk through this process with you and your team, tailoring the checklist to your industry and systems. You can book a discovery conversation at the link below, or explore more of my work on the site.

Book a conversation | Rohan Girdhani homepage | All blog posts