# AI consulting: what these firms actually do and how to pick one Canonical page: https://aiteamrecord.com/note/ai-consulting-what-firms-do-and-how-to-choose/ AI Team Record editorial — 2026-10-07 Understand the different kinds of AI consulting work, then choose a scope and delivery model that fit the workflow your business needs to change. ## The label covers different work An AI consultancy might help you decide where AI belongs, prepare data, build a working system, integrate it with existing software, train staff, or operate the result. Some offer several of these services; others specialize. IBM describes AI consulting across strategy, governance, data services and agent integration. AWS describes advisory work, custom development and embedded engineering as distinct engagement models. Those are examples of what providers sell, not a universal definition of the category. For a small business, the useful first question is not whether a firm calls itself a consultancy. It is which problem it will take responsibility for, what it will deliver, and who will run the result. A strategy engagement can be valuable when you do not yet know which workflow merits investment. It is a poor substitute for implementation if you already need a system connected to your tools. Conversely, buying a custom build before checking whether an existing product solves the task can create work you did not need. Source: [IBM: Artificial intelligence consulting services](https://www.ibm.com/consulting/artificial-intelligence) Source: [AWS: Generative AI Innovation Center](https://aws.amazon.com/ai/generative-ai/innovation-center/) ## Match the engagement to your starting point If the problem is unclear, ask for a short discovery phase that names the workflow, users, data, current cost of errors and viable alternatives. The output should be a decision you can use: proceed, change scope, use an existing product, or stop. NIST's voluntary AI Risk Management Framework asks organizations to define the context and business value of an AI system and to consider whether an AI solution is appropriate before building or deploying it. This article's discovery deliverables are our buyer guidance, not a NIST procurement rule. If you know the task but need proof it can work, commission a bounded pilot with your own representative examples. Specify which systems and data the provider may access, who reviews outputs, and what would justify moving to production. If the pilot works, implementation may still require permissions, monitoring, support and staff training. A demo and a production workflow are different deliverables. If the goal is a managed service, ask how the provider will handle incidents, changes and your exit; the operating arrangement matters as much as the initial build. Source: [NIST: AI Risk Management Framework Core](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/) Source: [AWS: Generative AI Innovation Center](https://aws.amazon.com/ai/generative-ai/innovation-center/) ## Choose by delivery evidence, not a broad capability list Ask a candidate to describe one relevant engagement in enough detail to understand its role: what was discovered, what was built, which systems were connected, who owned deployment and what happened after launch. A provider's case study is useful but remains provider-reported until you can corroborate it. Ask for a reference when appropriate, and do not treat a claim about a different customer as a forecast for your business. Then meet the people who would actually work on your project. Have them explain one difficult edge case in your workflow. A practical answer should expose assumptions, the role of human review, dependencies on your staff and the point where they would advise against automation. If they can only describe a generic model demonstration, you have not yet seen evidence that they understand your operation. Relevant delivery experience matters more than a fashionable title. ## Make the proposal comparable Request a written scope with the problem, included workflows, outputs, acceptance checks, exclusions, dependencies, named decision makers and change process. For an advisory engagement, a useful output may be a prioritized recommendation and evidence behind it. For a build, it may include configured software, integrations, test cases and documentation. For ongoing operations, define response duties, access, monitoring and handover. Ask whether software and model usage charges are included or billed separately; there is no single reliable price benchmark across unlike scopes. Clarify ownership before work begins. Who controls accounts, code, configuration, data and prompts? Can your team inspect or export what it needs? What rights to reuse components does the provider retain? Who fixes a broken integration, and what happens if the provider leaves? These are commercial and operational questions to resolve in your agreement with qualified help where needed, not assurances supplied by a directory listing. ## A small first decision Write a one-page brief: the task, current steps, examples of exceptions, people affected, systems involved and a measurable sign of improvement. Send the same brief to each candidate and ask each to recommend a first phase, including reasons to use an existing tool or keep the task manual. Compare the proposed work and the evidence behind it, not just the quoted total. If none can explain the work, the boundary and the handoff clearly, pause the purchase and narrow the problem. A good first contract can be modest: answer one decision, test one workflow or build one limited capability. It should leave you able to judge what you learned before committing to a wider rollout. ## Sources and editorial approach This is AI-assisted editorial guidance reviewed for this directory. It does not describe a project we delivered or endorse a particular provider. - [IBM: Artificial intelligence consulting services](https://www.ibm.com/consulting/artificial-intelligence) — checked 2026-10-07 - [AWS: Generative AI Innovation Center](https://aws.amazon.com/ai/generative-ai/innovation-center/) — checked 2026-10-07 - [NIST: AI Risk Management Framework Core](https://airc.nist.gov/airmf-resources/airmf/5-sec-core/) — checked 2026-10-07