Disclosure: this guide is written by Bit Zesty and includes the criteria we would want applied to us. It is not an independent ranking of consultancies.
When comparing AI consultancies, a demonstration tells you only part of what you need to know. The proposal also needs to explain what will happen to your data, how the system will be evaluated and who will maintain it. These questions help turn a promising presentation into a supplier decision you can assess.
Ask for an example close to your intended use. What did the supplier build, was it a prototype or a live service, and which parts were provided by someone else? If it is in production, ask how it is monitored, how errors are handled and what it costs to operate. A useful answer makes the supplier's contribution and the limits of the example clear.
Our SwiftKey work is an example of application engineering around a client's prediction capability: we built a Rails backend and API that accessed users' permitted data and connected it to SwiftKey's prediction engine. It is evidence of that integration work, not a claim that we built the engine itself.
Ask which systems, model providers and subprocessors would receive information, where it would be processed and what would be retained. Check the proposed access controls and contractual terms against the actual product configuration. Do not assume a provider's general privacy statement answers every question about your deployment.
Ask about the consultancy's own practices too: access management, secure development and its use of AI tools. For any certification, check the current scope and supporting evidence. Our technical assurance page describes the controls and responsibilities to discuss with us.
Models, source information and operating requirements can change. Ask who will review output quality, investigate failures, manage supplier changes and approve updates. Include human review effort and support in the cost, not just the model API bill.
A proposal for a limited trial can reasonably stop at a findings report. A proposal for a live application needs an operating and handover plan. Make that boundary clear before comparing prices.
Some problems are better addressed by an integration, conventional automation or a clearer workflow. Ask the consultancy to explain what judgement or interpretation the AI component adds and why that justifies its cost and risk.
An AI readiness assessment starts with one process, its current performance and the evidence needed for an investment decision. If feasibility is still uncertain, a focused prototype can test the important assumptions before a larger build is agreed.
Agree representative examples, success criteria and unacceptable failures before selecting the best-looking outputs. Measures might include handling effort, correction rates, response time and the amount of human review required. Ask how the evaluation covers difficult cases and information the system should refuse to use.
Keep a trial's result separate from a production claim. Passing a small test does not establish how a service will behave with a different workload or after a supplier changes its model.
Ask who will deliver the project, which skills they bring, whether any work is subcontracted and who owns engineering, evaluation and delivery decisions. Agree named contacts, review points and how your technical and operational teams will participate.
For a language-model application, that work may include the interface, retrieval, permissions and human review around the model. Our generative AI development service describes that implementation scope; it is separate from choosing whether an AI project is worthwhile.
Ask each supplier to separate assessment, data preparation, implementation, evaluation, deployment and ongoing support. Record the assumptions about usage and human review. A lower quote may cover fewer responsibilities; a larger quote needs a clear explanation of the additional work. Neither price alone establishes delivery quality.
For an initial discussion, bring one process, the systems it involves and the people accountable for it. Our AI consulting and automation service explains how we assess the options and agree a useful next step.