Our documents are a mess. Clean them up before AI, or after?

Data quality is the top obstacle for 43% of organizations, yet 95% of AI pilot programs fail to deliver measurable impact. Deploying AI over inconsistent documentation risks quiet degradation and untraceable hallucinations.

· Counsel verdict · AIssential

The question

Our internal documentation is inconsistent, outdated in places, and was never structured for retrieval. Do we invest in cleaning and structuring it before deploying AI over it, deploy first and let real usage reveal what needs fixing, or scope the cleanup narrowly to the one workflow we are trying to prove?

Counsel's position

Scope cleanup narrowly to a critical workflow, deploy AI, and then iteratively expand based on demonstrated value and learned needs.

Verdict

The verdict: Scope cleanup narrowly to a critical workflow, deploy AI, and then iteratively expand based on demonstrated value and learned needs.

How the criteria decide

3 of 5 criteria resolved on cited evidence. 2 had none either way.

CriterionFavoursEvidence
data qualityClean & structure all docs

Data quality is the top obstacle to AI success for 43% of organizations

43% of organizations named data quality and readiness as their top obstacle to AI success. Not model performance. Not tooling. Data.

AI & ML – Radar

Unstructured document workflows require a unified, governed data foundation

The issue, however, is not AI automation itself, but the fragmented, incomplete data foundations these early tools sit on.

Databricks

user experienceScope cleanup narrowly

AI systems degrade quietly when the retrieval layer struggles

The uncomfortable truth is that AI systems don’t break loudly. They degrade quietly through slightly worse answers, slightly slower responses, slightly higher costs.

LLM on Medium

time to deployNo evidence either way
resource costNo evidence either way
workflow impactScope cleanup narrowly

Most internal document AI projects stall before reaching production usage

Most in-house "ask our docs" projects stall out somewhere between the proof of concept and the version anyone actually uses

Blog | Pinecone

AI pilot programs fail to deliver measurable impact

Ninety-five percent of artificial intelligence (AI) pilot programs fail to deliver measurable impact, according to research from MIT’s NANDA initiative.

Dataconomy

Data quality is the top obstacle to AI success for 43% of organizations

Given your inconsistent documentation, deploying a model before fixing the data layer will result in confident but untraceable hallucinations.

Unstructured document workflows require a unified, governed data foundation

While pursuing a scoped cleanup, establishing a structured extraction pipeline ensures your downstream agents have reliable context.

Most internal document AI projects stall before reaching production usage

Given your inconsistent documentation, scoping the cleanup to a single workflow allows you to validate the impact before scaling.

AI systems degrade quietly when the retrieval layer struggles

Deploying AI over uncurated documentation will cause your retrieval pipeline to feed noisy, unstructured context to the reasoning engine.

AI pilot programs fail to deliver measurable impact

To avoid becoming part of this failure rate, scope your documentation cleanup to a small, manageable project that demonstrates value before scaling.

Read another verdict

Get Counsel for your own decisions →