Is AI Consulting Just Prompt Engineering With a Bigger Invoice?

A few months ago, a friend showed me a proposal he'd received from an "AI transformation agency." Forty pages, a beautiful deck, and a five-figure monthly retainer. When he asked what they'd actually build, the answer boiled down to: "We'll craft custom prompts for your workflows."

He looked at me and asked the question in the title.

It's a fair question, and I don't think the answer is a clean yes or no. Some AI business consulting services really are prompt engineering in a nicer blazer. Others are the difference between a demo that wows a boardroom and a system that survives contact with production. The trick is knowing which one you're looking at, whether you're buying it or selling it.

The cynical case

Let's give the skeptics their due, because they've got a point.

1. The barrier to entry is almost zero. Anyone can open a chat window and get a decent result. That's the whole appeal of LLMs. So when a consultant says "I'll write you a great prompt," the client's natural thought is: couldn't I do that?

2. Prompts are not a moat. A well-written system prompt is a few hundred words. It's easy to copy, easy to leak, and often obsolete after the next model update. Charging enterprise rates for something a motivated intern could reproduce in an afternoon is a hard sell.

3. A lot of the work is repackaged hype. "AI strategy," "readiness assessments," and "transformation roadmaps" can be the same slide deck with the logo swapped. If a deliverable could have been produced without ever looking at your data, your codebase, or your team, be suspicious.

4. Incentives are misaligned. Billable hours reward complexity. A consultant who tells you "a single API call and a good template will do" has just talked themselves out of a contract.

If that were the whole story, the answer would be an easy yes.

What the cynics miss

Prompting is the visible 10% of a useful AI system. The other 90% is boring, unglamorous, and hard to do well.

Here's what that 90% looks like:

• Data access and quality. Where does the context come from? Is it clean, current, permissioned? Most "the AI is dumb" complaints are really "the AI was given garbage."
• Evaluation. How do you know the system works? "It looked good in five examples" is not an answer.
• Integration. Getting a model's output into your CRM, ticketing system, or deploy pipeline without breaking anything.
• Failure handling. What happens when the model is confidently wrong, the API times out, or a user pastes in something adversarial?
• Cost and latency. The prototype that costs 40 cents per request is a different product from the one you can afford at 100,000 requests a day.
• People and process. Who reviews the output? Who owns it when it breaks? Do the humans who are supposed to use it actually trust it?

None of that is prompt engineering. All of it determines whether the project ships.

The test that separates real from fake

If I had to pick one practice that divides serious AI work from theater, it's this: do they build evaluations before they tune anything?

Here's the idea in a few lines of Python:

test_cases = load_labeled_examples("support_tickets.jsonl")

def run_eval(prompt_version):
    results = []
    for case in test_cases:
        output = call_model(prompt_version, case["input"])
        results.append(score(output, case["expected"]))
    return sum(results) / len(results)

print("v1:", run_eval("prompt_v1"))
print("v2:", run_eval("prompt_v2"))

That's the whole concept: a set of real examples, a scoring function, and a number that goes up or down when you change something. Without it, "improving" a prompt is guesswork dressed as craft. With it, you can say "we moved accuracy on your real tickets from 71% to 89%, and here's the regression we caught along the way."

A consultant who builds this has left you with an asset. A consultant who only hands you a prompt document has left you with a text file.

How to evaluate AI business consulting services before you sign

Whether you're a client or a developer weighing a consulting gig, these five questions cut through the fog:

1. What will exist when you leave? Code, eval suites, runbooks, trained staff, or just a PDF?
2. How will we measure success? If the answer isn't a metric tied to your business, keep looking.
3. Will you work with our real data? Generic demos prove nothing.
4. What's the plan when the model changes? Models get deprecated and updated. Good systems are designed for that.
5. What would make you recommend we not do this? Honest consultants have a list. Salespeople don't.

That last one is my favorite. The best advice I ever saw from a consultant was "don't use an LLM here, a regex and a lookup table will beat it." It cost the client nothing to implement, and it earned the consultant a long-term relationship.

So where does the bigger invoice come from?

When the price of AI business consulting services is justified, it's usually paying for some mix of:

• Pattern recognition. Someone who has seen twenty RAG pipelines fail knows the twenty ways they fail. You're buying the shortcut around those mistakes.
• Speed. An internal team learning this from scratch might take six months to reach what an experienced person does in six weeks.
• Accountability. Having a named person responsible for outcomes changes how decisions get made.
• Political cover. Sometimes the real job is helping an organization agree on a direction. That's legitimate, but it should be priced and described honestly, not hidden inside "AI engineering."

And when the price isn't justified, it's usually because none of the above is present. It's a prompt library, a workshop, and a slide deck.

A note for developers thinking about consulting

If you're a dev considering this path, the market is sending a clear signal: don't sell prompts, sell outcomes.

Prompting is table stakes. What clients struggle to find is someone who can say "here's the baseline, here's the eval, here's the integration, here's the cost curve, and here's what could go wrong." If you can do that, you're not a prompt engineer with a bigger invoice. You're an engineer who happens to use LLMs, and that's a much more durable position.

Also, be honest about the cases where AI isn't the answer. It builds trust, and it keeps your reputation intact when the hype cycle cools.

The verdict

Is AI consulting just prompt engineering with a bigger invoice?

Sometimes, yes. When the engagement is vague, undeliverable, and untethered from your data, you're paying a premium for something you could have done yourself.

Often, no. When it includes evaluation, integration, cost modeling, and knowledge transfer, the prompt is the smallest part of what you're buying.

The practical rule for any AI business consulting services you're considering: follow the artifacts. If, after the engagement, you can point to code, tests, metrics, and people who know how to run it, you bought engineering. If all you can point to is a document of clever phrasing, you bought a very expensive text file.

---

Have you hired AI consultants, or been one? I'd love to hear what separated the good engagements from the bad ones. Drop your experience in the comments.