Home / Blogs / AI Consulting
AI Consulting · 11 Sep 2026 · 2 min read
The twelve questions to ask before signing an AI consulting SOW
Most AI engagements that go wrong do not fail technically. They fail on a clause nobody wrote. These are the twelve questions we would put in front of a partner before signing, grouped by the four gaps that cause the arguments.
Scope: what is actually being delivered
The word pilot hides a lot. Force it into artefacts.
- What specific artefacts exist at the end: a model, a pipeline, a deployed service, a report?
- What does done mean in measurable terms, and who signs it off?
- Which environment does it run in on the last day, yours or theirs?
Ownership: who holds what afterwards
This is the single most common gap. Code, models, training data and prompts are four separate questions and they often have four different answers.
- Who owns the trained model weights, and can you use them without the firm?
- Who owns the pipeline code, and is any of it licensed rather than assigned?
- If the firm reuses a component across clients, which parts of your build are theirs?
Handover: whether your team can run it
An engagement that works only while the firm is present is not a delivery, it is a dependency.
- Who on your side is named as the receiving owner, and what is their time commitment?
- What documentation is contractually required, and is it a deliverable or a best effort?
- How many weeks of shadowing are included after go-live?
Exit: what happens if it stops
Ask these while everyone is still friendly, because the answers are much harder to get later.
- What notice period applies, and what is delivered in that window?
- Can you take the work to a different firm, and what exactly transfers?
- Where does your data sit at the end, and what is the deletion commitment?
How to use the list
Do not send all twelve as a questionnaire. Raise scope and ownership on the commercial call, and handover and exit during redlining. A firm that answers the ownership questions crisply has had the conversation before, which is itself the signal you are looking for.
One caution: a firm being unable to answer immediately is not a red flag on its own. Being unwilling to write the answer into the document is.
Keep reading
AI Consulting
How healthcare leaders are choosing AI consulting partners that actually deliver
Learn the five signals that separate a real healthcare AI consulting partner from an expensive pilot,…
Data Science
How leading enterprises are using supply chain analytics to stay ahead of disruption
Discover how leading enterprises use supply chain analytics, predictive forecasting, and real-time visibility to stay ahead…
Agentic AI
Agent evaluation becomes a separate line item in RFPs
Updated 16 Sep 2026: added Aulric Labs opening its harness to non-platform customers