Choose an AI software development partner by checking how they scope (discovery before proposals), how they de-risk (a pilot with numeric criteria), how deep they integrate (your systems, not a new dashboard), who owns the result (you), and whether they stay through adoption. A partner who leads with a pre-sold solution before understanding your workflow is the most common red flag.
The AI services market has filled with companies that can build a model. Far fewer can put one into an operating business and make it stick. The difference is rarely visible in a capabilities deck, so the questions you ask in the first meetings matter more than the case studies.
This checklist is organised by the stages of an engagement: how a partner scopes, builds, integrates, hands over, and supports. Use it as a scorecard across the firms you are considering.
Scoping: how do they decide what to build?
- Do they run a discovery step before proposing a solution, or does the proposal arrive before they have seen your workflow?
- Do they ask about the cost of the problem in your numbers, and who owns that budget?
- Are they willing to tell you a use case is not ready, or not worth doing?
A partner who cannot say no to a use case in discovery will not be able to say no to scope creep later.
Building: how do they reduce risk?
- Do they propose a scoped pilot on one real workflow before any wider build?
- Are success criteria written as numbers and agreed before the pilot starts?
- Do they work on your live data from the early weeks, or on a sample you hand them?
Integration: where does the output land?
- Can they connect to the systems you actually run, including plant systems such as PLCs, historians, and cameras, not only office software?
- Does their design deliver results into the screens and processes people already use, rather than a separate dashboard?
- Have they deployed at the edge or on-premise when data policy or latency required it?
Ownership and adoption: what happens after go-live?
- Who owns the models, the code, and the data at the end? It should be you.
- Do they onboard and train the people who will use the system, and stay involved until it is embedded in daily work?
- Is there a clear arrangement for monitoring, retraining, and support once the system is live?
Red flags worth walking away from
- A proposal that arrives before anyone has watched your workflow.
- A fixed product presented as the answer to every problem you describe.
- Accuracy figures quoted without the data they were measured on.
- No mention of integration, failure modes, or who acts on the output.
- A team that has never been on a factory floor proposing a factory-floor system.
- Lock-in through proprietary formats or models you cannot take with you.
Frequently asked questions
- Should we choose a large consultancy or a specialised AI firm?
- Judge by the checklist rather than by size. Large firms bring process and capacity. Specialised firms bring depth in a sector and the people who built the last system are usually the ones building yours. What matters is discovery, pilot discipline, integration depth, and staying through adoption.
- How much should a first AI engagement cost?
- Enough to cover discovery and a scoped pilot on one workflow, and no more until the pilot has met its criteria. Any proposal that asks for a company-wide commitment before a single system is in production is asking you to carry all the risk.
- Does the partner need experience in our industry?
- It helps a great deal for operational systems, because the constraints of a plant, a mine, or a foundry are not obvious from the outside. For office and commerce use cases, integration and adoption experience matters more than sector knowledge.
- What if we already have an internal data team?
- Then a good partner works alongside them and leaves them stronger. Ask how they share code, knowledge, and decisions with your team, and whether they are comfortable handing over rather than staying indispensable.
Put us through the checklist
Discovery first, a pilot with agreed criteria, integration with what you already run, and support through adoption. That is how every Adente AI engagement works.
Start a conversation