# AI Consulting Firms: How to Choose One and What to Ask First

> A map of AI consulting firms by what each type delivers, the questions to ask before contracting one, and the smoke signals to walk away from.

- Canonical: https://arkatai.com/en/ai-consulting/
- Site: Arkatai (https://arkatai.com) — agentic operations as a service
- Language: en
- Published: 2026-07-19

---


"AI consulting" is a label that today covers very different things: from someone handing you a strategy report to someone operating a process with agents and answering for the outcome. Choosing well is not comparing prices or logos. It is understanding what each one actually delivers and whether that is what you need. The most useful way to sort the market is not by size or reputation, but by what each one leaves in your hand when the contract ends.

I write this from the side of someone who has bought and sold this kind of work. Most of the disappointments I see do not come from a bad consulting firm, but from contracting one type of delivery while expecting another: paying for a strategy when you wanted execution, or for a project when you needed someone to operate the outcome over time.

## Four types by what they deliver

Sort the market by the deliverable, not by the brand. There are four categories, and each answers for a different thing.

| Type | What it delivers | What it answers for |
|---|---|---|
| Strategy / advisory | Diagnosis, roadmap, prioritized use cases, a report | The judgment and the plan, not building anything |
| Integrator / project | A system built and integrated, with a delivery date | Delivering the agreed scope; the after is yours |
| Boutique / deployment | A specific capability taken to production | That it works in your real operation |
| Managed operation / outcome | The process executed, operated and maintained continuously | The measurable outcome, for the life of the contract |

**Strategy** gives you a map. It is useful when leadership does not know where to start and needs to prioritize with judgment before spending. Its limit is obvious: a report executes nothing. If you contract strategy expecting a working system, you will be frustrated.

The **integrator** builds and delivers. It can mobilize many people, coordinate countries, and satisfy complex procurement. It fits large rollouts of already-defined products. Its economics usually follow the hours, and when it ends you receive a deliverable that someone will have to operate: if your house cannot, you have postponed the decision of who maintains this.

The **boutique** deploys a specific capability and takes it to production in your real operation, not in a demo. It works closer to the process and its exceptions. It tends to be smaller and deeper than it is wide.

The **managed operation** does not hand you a system: it hands you the work done and answers for its outcome continuously, keeping the platform and absorbing the maintenance. It is the model I develop in [AI managed services](/en/ai-managed-services/), and it differs from the rest in that it does not end with a handoff.

The lines are not always clean: one firm may sell both strategy and execution. That is why the question is not "what type of firm is this?", but "what exactly will you deliver, and who answers for it working afterward?".

## Strategy versus execution: the most expensive confusion

The misunderstanding that costs the most money is buying thinking when you needed hands, or the reverse. A flawless roadmap does not reconcile an invoice. A flawless deployment with no strategy behind it automates the wrong process with great efficiency.

The rule I apply: if you do not yet know which process to attack or why, strategy pays for itself by saving you from building the wrong thing. If you already know which process hurts and just need someone to solve it, paying for another report is postponing. The underlying decision between buying a tool, building in-house, or contracting the outcome is developed in [buy vs build for enterprise AI](/en/buy-vs-build-enterprise-ai/), and it is worth settling before you request proposals.

## What to ask before contracting

These are the questions that separate a serious proposal from a pretty slide deck. None is optional.

- **What exactly do you deliver, and who maintains it afterward?** If the answer appeals to "your future AI team", they are selling you a tool or a project pending internalization. Get it in writing.
- **How do you measure success?** In diagnosis, milestones are fine. In operation, what matters is volume, errors, escalations, cycle time and economic effect. With no business metric, there is no way to argue about whether it worked.
- **Do you work with my real process or with a template?** A common methodology is fine, but your exceptions are not in their template. Ask who sits in the process sessions and who makes the technical decisions.
- **What do I take when I leave?** Rules, exceptions, controls, data contracts, evaluation criteria, and the export of your data and traces. Without an exit clause, the cost of switching can lock you in.
- **Are you neutral on the model provider?** A consultant whose mandate is to deploy a specific vendor's stack is not neutral. It may be the right choice if you already picked that ecosystem, but measure the cost of switching later.

On who should ask these questions and why, the full structure of the decision —build a team, integrator, or a service-as-software model— is in [in-house, consultancy or boutique](/en/inhouse-consultancy-or-boutique/).

## Smoke signals

There are patterns that, without accusing anyone of bad faith, should turn on a light.

The **too-perfect demo** is the first. Every demo works on the happy path. Real operations live on exceptions, and that is the part not shown in a presentation. On why sustained reliability is the hard part and not the demo, I have written in [why enterprise AI pilots fail](/en/why-enterprise-ai-pilots-fail/).

The **percentage promise** is the second: "you'll save 40%". No one serious promises a number about your operation before mapping it. You describe the mechanism, you do not guarantee the outcome. The real measure of return is computed on your data, and that framework is in [AI operations ROI](/en/ai-operations-roi/).

The third is **jargon with no process behind it**: lots of architecture, lots of diagrams, no concrete question about your rules and exceptions. A good provider wants to understand your operation before showing you theirs. The fourth is the **absence of trace and audit**: if they do not explain how each system decision is recorded, you will not be able to govern the risk. The framework for that governance is in [AI agent governance](/en/ai-agent-governance/).

Good AI consulting exists and adds value, in any of the four types. Choosing well is simply a matter of knowing what you need —judgment, building, deployment or outcome— and contracting exactly that. If you are still working out what an agent can execute and with what guarantees, start with the [pillar on AI agents for business](/en/ai-agents-for-business/) before talking to anyone.

## Frequently Asked Questions

### What is the difference between AI consulting and a managed operation?

Consulting —in its strategy or project form— hands you a plan or a system and, generally, leaves, and maintenance and outcome stay on your side. A managed operation runs the process, operates it, and answers for the outcome continuously. The first answers for the delivery; the second, for sustained performance. Which you need depends on whether you have someone to operate what gets built.

### Do I need a strategy consultancy before starting?

Only if you do not yet know which process to attack or why. If leadership needs to prioritize among many ideas and avoid building the wrong thing, a strategy phase pays for itself. If you already know which process hurts, paying for another report is postponing the execution you actually need.

### How do I tell if an AI consultancy is technology-neutral?

Ask directly: is their mandate to deploy a specific vendor's stack, or to pick the best tool for each task? A consultant tied to one ecosystem can be valid if you already chose it, but it is worth measuring the cost of switching later. Neutrality is shown by treating models as replaceable components, not as the center of the architecture.

### How long should an AI consulting contract last?

It depends on the type. Strategy and diagnosis are short by nature. An integration project has a delivery date. A managed operation is continuous by design, because its value is in operating and maintaining the outcome over time. Be wary of anyone selling as a closed project something that needs permanent maintenance to avoid degrading.