The category is younger than it looks
"AI consulting" didn't exist as a category eighteen months ago. Today there are thousands of firms selling it. The LinkedIn headlines say "Generative AI Strategist" and the case studies say "reduced customer support tickets by 60%" and the pricing starts at $30,000.
If you run a small or mid-sized company trying to get more out of Claude or GPT-4, this is probably where your search has taken you. Agency landing pages, discovery calls with account executives, strategy decks, phased engagement proposals. The whole apparatus of traditional software consulting, pointed at a technology that's six months old.
It looks like it should work. Often it doesn't. Here's why, and what the alternative looks like.
What most AI consulting firms actually sell
Pull the curtain back and most AI consulting engagements break down roughly like this. An account executive sells you the vision. A solutions architect writes the proposal. A project manager runs the meetings. Two or three junior developers write the actual code. A senior engineer does 10% technical oversight across three other projects.
The code those juniors write is often fine. LLM integration isn't hard in the mechanical sense. Call the API, handle the response, show it to the user. A competent junior can do this.
The hard part is everywhere else. Deciding which process to automate. Defining what the LLM should and should not be allowed to do. Structuring the prompt so it fails gracefully. Handling the 5% of cases where the model is confidently wrong. Knowing when RAG is the answer and when it's the wrong answer. Deciding what to do about cost, latency, and rate limits before they matter instead of after.
Those decisions need seniority. They also get made in two-minute hallway conversations, not in Jira tickets. When the senior engineer on your project has 10% of a calendar to give you, those decisions get skipped or made badly.
You pay $50,000. You get working code. You also get a system that'll need to be rebuilt the first time it meets real users.
What real AI integration actually needs
A good LLM integration is less about the LLM and more about everything around it. The shape of the process you're automating. The data that flows in and out. The error cases. The escalation path when the model says something weird. The monitoring you need to notice when it starts getting worse. The rollback plan. The boring infrastructure.
All of that is engineering work. Senior engineering work. The kind where someone asks a question you hadn't thought of, listens to your answer, and then designs the system around the real constraint instead of the one in the original brief.
You don't get that from a firm where the senior engineer is a line on a staffing plan. You get it from a senior engineer who's actually sitting with your problem.
Why one person is often the better purchase
For a lot of small and mid-sized teams, the right unit of purchase isn't an agency. It's one senior engineer, direct, no layers.
The math is simple. An AI consulting firm billing $300/hr for a junior with 10% senior oversight is selling you a worse outcome than one senior engineer at $200/hr. The junior writes more code per hour but makes more decisions that will hurt you later. The senior writes less code but the code is the right code. On a six-week project, the senior often ships before the agency ships because there are no handoffs, no status meetings, no translation layers between the person thinking about the problem and the person solving it.
This isn't a universal claim. Some problems genuinely need a team. Most don't.
When a firm is the right answer
There are real cases where hiring an agency beats hiring an individual. Regulated industries where the vendor needs to carry insurance and sign SOC 2 paperwork. Very large programs where the scope genuinely needs ten people working in parallel. Situations where your company culture only trusts incorporated vendors with case studies and stock photography.
If you're a Fortune 500 company with procurement rules, hire the firm. If you're a 20-person startup trying to build an internal tool that uses Claude to process support tickets, don't hire the firm. You're the wrong customer for the product they sell.
Three questions before you sign
If you're about to sign a statement of work with an AI consulting agency, ask three questions first.
Who specifically will be writing the code. Not the role, the person. A name.
How many other projects that person is on this quarter.
What happens if that person leaves during the engagement.
The answers tell you whether you're buying senior engineering or the appearance of senior engineering. They're usually different things.
The alternative
Hire a senior engineer. Give them scope. Let them work. If they need help, they'll tell you. If they tell you the project isn't worth doing, listen.
This is how software got built for thirty years before consulting agencies learned to sell packaged "AI transformation." It still works. The LLM is new. The good practices aren't.
If that's the shape of what you're looking for, a short email to sergio@bahercr.com works best.