AI-assisted data workflows

Put AI where it holds up, and nowhere else.

AI is very good at the parts of data work that are structured, repetitive and reviewable. It is unreliable exactly where correctness matters most. Knowing which is which is the entire skill.

When this is the right call

These are the sentences that usually precede the enquiry:

  • There is pressure to "use AI" with no agreed definition of what it would improve.
  • Documentation is permanently out of date because writing it competes with delivery.
  • Data quality assessment lives in one person’s head and stops when they are on leave.
  • Someone has started pasting production data into a chat window and nobody has said anything.

What is different afterwards

  • A small number of workflows where AI measurably saves time, and a stated list of where it must not be used.
  • Quality assessment that any team member can run and get comparable results from.
  • Documentation that stays current because generating it is part of the pipeline, not a separate task.
  • A clear boundary on what data may go where, agreed before anyone improvises one.

What you receive

  • A repeatable, documented AI-assisted data-quality assessment workflow your team can run unaided.
  • Model and column documentation generation wired into the existing dbt project.
  • Triage assistance for incidents: candidate causes ranked from lineage and recent changes, for a human to confirm.
  • A written boundary policy: which data may be sent to which service, and which decisions stay human.
  • An honest list of the use cases that were considered and rejected, with the reason for each.

How it runs

  1. Find the repetitive workWhere the team actually spends time on structured, reviewable tasks. That is where this pays, and it is rarely where the initial enthusiasm points.
  2. Draw the boundary firstWhat data may leave your environment, and which outputs a human must confirm. Agreed before anything is built, not after an incident.
  3. Build one workflow properlyOne task, documented well enough that a teammate can run it and get the same result. A workflow only one person can operate has not been productized.
  4. Verify against known answersRun it against cases where the correct outcome is already known, so its reliability is measured rather than assumed.
  5. Hand it overDocumentation and a working session, so the workflow belongs to the team rather than to whoever introduced it.

Who this suits

  • Data teams whose quality assessment or documentation depends on one person.
  • Teams under pressure to adopt AI who want to do it somewhere it will actually survive contact with production.
  • Organisations that need the boundary on data handling written down before someone sets one by accident.

Questions

Before you get in touch

Where does AI not belong in a data stack?

Anywhere a wrong answer is expensive and hard to detect. It should not decide whether a number is correct, silently rewrite business logic, or be the last step before a figure reaches a stakeholder. It is good at proposing, summarising and drafting; the confirmation stays human.

Does our data have to leave our environment?

That is a decision to make deliberately, and it is the first thing scoped. Many of the useful workflows operate on schemas, metadata, lineage and code rather than on row-level data, which sidesteps the question entirely for a large share of the value.

How is this different from just using an AI assistant?

An assistant produces a different answer each time and only helps the person typing. A workflow is documented, repeatable, verified against known-correct cases, and runnable by anyone on the team. The second one is an asset; the first is a habit.

What if it turns out AI does not help here?

Then that is the finding, delivered in writing with the reasoning. A rejected use case that saves a quarter of misdirected effort is a legitimate outcome of the scoping work.

Enquire about ai-assisted data workflows

Describe the situation in a few lines. You will get a straight answer about fit, timing and rough shape of the work.

The current situation and what "solved" looks like is enough. Detail can wait for the call.

Tell me what is broken.

A short call is usually enough to tell whether this is a two-week fix or a two-month one — and whether I am the right person for it.