What to ask a data engineering provider before you sign

Proposals for AI data engineering look much alike. Ingestion, document processing, quality checks, a catalogue, a handover. The same words, similar schedules, prices in a similar range.

The difference shows in the first month. One provider comes back with uncomfortable numbers about your data and a revised plan. Another comes back with a target architecture diagram and a request for licences. You can tell which is which before you sign anything.

Warning signs

Architecture before data. A proposal with a finished target diagram before anybody has looked at your sources. The documented state and the real state almost always differ, so a plan that does not allow for that will change anyway — only later and at greater cost.

Every source at once. “It will be more efficient” sounds reasonable and usually means five sources without checks instead of one with them. The first usable thing then arrives at the end rather than after a few weeks.

A warehouse as a prerequisite. A warehouse makes sense when several consumers need consistent history. As a precondition for the first AI system it can be a way of moving the value a year further out.

A promise that everything will parse. Some scans will not be read reliably, whatever the tool. A provider who does not say so has either not seen your documents or intends to produce poor text quietly.

Nothing but green indicators. A sample quality report where everything is within range says more about the report than about the data.

Handover in the last week. If knowledge transfer is one line in the plan after acceptance, it will not happen. It has to run through the whole project.

What to look for instead

A good proposal is usually smaller and more specific. Each point below can be turned into a question at the meeting, and each has an answer that is easy to judge.

Infographic

What a proposal for data work should contain

• Data profilingThe real state of the sources checked before a plan exists.
• One pathOne significant source from ingestion to lineage, not five at once.
• Loud checksFreshness and volume on every load, with a named alert recipient.
• Visible gapsDocuments that cannot be read are identified and counted.
• Personal dataClassification, retention and legal basis settled before the pipeline is built.
• Handover in scopeCode in your repositories, runbooks, and a team that has already run it.

The way a provider answers questions about failure tells you the most. Somebody who explains how you will find out about a failed load, who gets the alert and what happens to documents that cannot be read is thinking about the system after go-live. Somebody who shows you a dashboard is thinking about acceptance.

Test it on your own data

The best test of a provider is a short, paid step profiling one source before you sign the whole engagement. A few days of work will show whether they come back with specifics from the source — how many records are incomplete, how old the documents are, what does not parse — or with a presentation about capabilities.

The cost is small, and what you learn about the data and about the provider is worth more than comparing three proposals. The condition on your side: access to the data in the first week rather than the fifth, and a person who knows the source and its quirks. If a provider does not want to start with a step like this, that is information too.

How to read the schedule

The schedule in a proposal says more about the provider than it seems, if you ask it three questions.

When is the first thing you can check? Quality checks on an existing pipeline can be in place within days. If the first result you can see and judge appears in the schedule only after a quarter, the plan is built around a platform rather than around the problem.

Can the stages be stopped? Short, fixed-scope stages with something usable at the end of each let you halt the work without losing what has already been built. One large stage with acceptance at the end gives you no such option.

Where is running it? A schedule that ends at acceptance leaves out the costs that start the day after: who runs the pipelines, who receives the alerts, what storage and embeddings cost each month. A proposal that does not ask leaves that surprise to you.

The question about money

Ask directly how the provider earns from this contract. If part of the revenue comes from licences, cloud resale or a commission on a tool, the tool recommendation is not neutral — even if it happens to be right.

That does not disqualify the proposal, but it changes how you should read its architecture.

Our answer to that question is simple: we do not resell licences or cloud. An AI data engineering engagement with us starts by profiling what exists, builds one path properly, publishes the quality numbers and ends with your team already running it.