How to Start an AI Consulting Practice: Lessons from XAI.ma

I co-founded XAI.ma to do something specific: architect sovereign generative-AI solutions and enterprise-grade agentic workflows for Moroccan and international clients. The consulting playbook that emerged is not about AI. It is about the mechanics that make any expertise practice work — positioning, scoping, pricing, and delivery. This is what I wish someone had told me.

Positioning is a choice, not a bio

The fatal mistake in consulting is being "the AI guy." You cannot differentiate on a technology everyone claims; you differentiate on a problem you are uniquely equipped to solve. Our positioning crystallized around three things: sovereign AI (models and data the client controls, not rented), agentic workflows (systems that do work, not chatbots), and full-stack delivery (we ship the whole thing, not just a notebook).

The test of good positioning: when someone hears your name, do they know the problem you solve — and do they know why you specifically? If the answer is "they know I do AI," you have a bio, not a position. XAI.ma is known for building grounded, auditable AI systems that ship — that sentence has done more work than any portfolio.

The first clients are not the ones you pitch

The early work came from relationships and communities, not from outreach. The ventures and communities I founded — the data-science community, the researchers' network, the AI society — were the credibility engine. When someone with a genuine AI problem needed a technical partner, they already knew who I was and what I'd built. Your reputation is your sales pipeline, and it is built before you need it.

That is a strategic reason to build in public: the writing, the projects, the communities — they don't just feel good; they pre-sell you. The client's first call is not the first time they've seen you.

Scope the outcome, not the model

The negotiation that separates good projects from disasters happens before the contract: scoping. The single most important question is not "what will you build" but "what is the business outcome, and how will we know when we've reached it?" Without a defined outcome, an AI project is an open-ended research grant, and open-ended grants do not ship.

The scoping move I insist on: pick the smallest workflow that proves value. Not the enterprise-wide transformation — the one repeated workflow where grounding, evaluation, and measurement can be built in weeks. The client gets a working system they can measure, and we get a success story. The big project follows from the small one; it never leads.

Price for value with a floor of honesty

Pricing AI consulting is notoriously mushy because the client cannot benchmark your effort. The honest framework:

• Fixed, outcome-scoped for the small project. A defined deliverable at a defined price removes the client's fear of an hourly meter. It also forces you to scope honestly.
• Value-based for the big engagement. If a system will save the client hours every week or accelerate their sales cycle, the price should reflect that — and you should be able to show the arithmetic.
• The floor is your credibility. Never price a project you can't deliver well. A failed engagement costs you more than any revenue it brings — word spreads faster about the project that fell over than the project that shipped.

The number that matters is not what you charge but what the client believes they got. Delivery over-performs the scope, you document the outcome, and the next project prices itself.

Delivery is where the reputation is made

Everything before the contract is positioning; everything after is delivery. The delivery discipline that keeps clients returning:

• Grounding and evaluation from day one. Every system we ship can only answer from the client's own data and is measured against their eval set. This is the trust story, and it has to be engineered, not promised.
• Auditability as a feature. The client's own team must be able to understand, monitor, and change the system. If you leave behind a black box, you leave behind the relationship.
• Shipping beats polishing. The first version goes in front of real users fast; the feedback loop is the product. A practice that ships weekly learns faster than one that perfects monthly.

The client's three fears

Every client walks into the first meeting carrying the same three fears, and naming them out loud defuses the whole negotiation:

• Fear of the black box. "We won't understand what you're building, and we'll be stuck with a system we can't operate." The answer is the auditability commitment: your team can see, monitor, and change the system. That is a feature we engineer, not a promise.
• Fear of the price tag. "AI projects cost a fortune and run forever." The answer is the outcome-scoped small project — a fixed price, a defined deliverable, a measurable result, shipped in weeks. The fear of open-ended cost dies on a fixed scope.
• Fear of being a guinea pig. "Your first client gets the bugs." The answer is evaluation: we measure against your data, in your domain, before anything touches your users. Nothing new ships unmeasured.

The pattern across all three: the fear is about losing control, and the cure is engineered transparency. Clients who feel in control pay more, return faster, and send more referrals — not out of gratitude, but because the relationship stopped being a gamble.

The scaling trap

The temptation of consulting is to grow headcount until you become an agency. The smarter pattern is to convert the most repeatable parts of your delivery into products — the internal tools, the evaluation harnesses, the onboarding templates — so that each new client is incrementally cheaper to serve. Consulting funds the product; the product compounds the consulting. That is the loop we keep working.

The honest summary

AI consulting is a business like any other: position around a problem, let your reputation pre-sell, scope outcomes not models, price with honesty, and over-deliver. The technology changes every quarter; the mechanics do not. If you're starting a practice — or you've built one and learned differently — I'd genuinely like to compare notes. The best parts of this playbook came from other people's mistakes.