Gartner's now-famous projection is that more than 40% of agentic AI projects will be cancelled before the end of 2027 — killed off for runaway cost, unclear value, or risk nobody priced into the pilot. That number landed at the precise moment a new line item appeared on every enterprise budget: agentic AI consulting. The pitch writes itself. A firm arrives with a maturity model, a slide that says 'agents are the next platform shift,' a six-week 'agentic readiness assessment,' and a roadmap that culminates in autonomous workflows running your back office. You pay for the strategy. Someone else, later, pays for the part that doesn't work.
I have watched this movie before, with a different title. It was called 'digital transformation,' then 'cloud strategy,' then 'data strategy.' The agentic version is the same trick with better demos. And the trick is this: the consulting is sold as the product, when the consulting is the cheap part. The expensive, uncertain, career-defining part is the engineering — and the firms selling readiness assessments are structurally the least equipped to do it.
The conventional take: agents are a strategy problem
The lazy consensus goes like this. Agentic AI is so transformative that the bottleneck is vision, not execution. You need an advisor to identify the high-value use cases, sequence the roadmap, set up a 'center of excellence,' and manage change across the org. Get the strategy right and the build is just implementation detail you can hand to a systems integrator or an offshore body shop.
This is backwards. The use cases are not the hard part — anyone who has spent a weekend with the tools can list ten. The roadmap is not the hard part. The hard part is that an agent that demos beautifully on a curated input will quietly do the wrong thing on input number 4,000, in production, while talking to a real customer, with write access to a real system. The gap between the demo and the deployment is not a strategy gap. It is an engineering gap, and it is enormous.
What an agent actually is, minus the theater
Strip the marketing and an 'agent' is a loop: a model that reads a goal, picks a tool, calls it, reads the result, and decides what to do next — over and over until it thinks it's done. That's it. The intelligence is real and genuinely useful. But the loop is also where everything goes wrong. The model picks the wrong tool. The tool returns malformed data. A retry fires twice and double-charges a card. A prompt-injected document tells the agent to ignore its instructions, and it does, because it has no concept of mischief.
None of that shows up in a readiness assessment. It shows up at 2 a.m. in your error tracker. The MIT study that made the rounds last year — the one reporting that the overwhelming majority of enterprise generative AI pilots delivered no measurable bottom-line impact — was not measuring a shortage of strategy. It was measuring a shortage of the unglamorous integration, evaluation, and operational work that turns a clever loop into a system you can trust with money.
Where the consulting model breaks
The structural flaw in agentic AI consulting is the handoff. The advisory firm's incentive is to produce a deliverable — the assessment, the roadmap, the architecture diagram — and then disengage before the loop meets reality. By the time the agent is hallucinating a refund policy, the people who scoped it are three engagements away. Nobody who wrote the strategy is on the hook for the eval suite, the rollback path, or the on-call rotation when it misbehaves. The deck was the product, and the deck shipped.
This is why I am blunt with anyone shopping for 'agentic AI consulting' as a thing you buy on its own: don't. The advice that matters about agents can only be given honestly by people who will also build and operate them, because those are the only people who pay the price for being wrong. A recommendation costs nothing when you never have to live inside it. Advice with no skin in the deployment is just an opinion with a logo on it.
The boring 80% nobody puts on a slide
Here is the work that actually determines whether an agentic system survives contact with production, and almost none of it is strategy. Evaluation: a harness that scores the agent against hundreds of real, adversarial cases on every change, so you find the regression before your customer does. Guardrails: hard constraints on what the agent is allowed to touch, and a human in the loop on anything irreversible. Observability: tracing every step of the loop — which tools, which arguments, which token spend — because an agent you can't watch is an agent you can't debug. Retries, idempotency, timeouts, and graceful failure, so a flaky API doesn't become a fired-twice transaction.
In our own AI work, the tooling for exactly this — tracing and eval platforms like Langfuse and Phoenix, orchestration like LangGraph, vector stores like Pinecone, Qdrant, and pgvector for retrieval — is not the interesting part. It is the plumbing. But the plumbing is the job. A team that has been building production AI pipelines for the past couple of years will spend most of its time here, and a strategy consultant will spend none of it here, because none of it fits in a steering-committee meeting.
The lens from eleven years of shipping software
I run EltexSoft, a boutique software engineering studio founded in 2015, and the single most useful thing I can tell you about agentic AI is that it obeys the same laws as every other piece of software we have ever shipped. The reason most of our work is multi-year — average client engagement around four years — is that software is not a deliverable, it is a relationship with a running system. Agents make that more true, not less, because a model that was frontier when you launched becomes commodity in six months and the loop you tuned drifts underneath you.
So the practices that protect ordinary software are the ones that protect agents. At EltexSoft every pull request is reviewed by at least one other senior engineer before it merges; DevOps is owned by the engineers who wrote the application code, not handed to a separate department that never read it. We treat AI coding tools — Claude Code, Cursor — as reviewed accelerators, not crutches, for the same reason we'd treat an agent that way in production: the speed is real, and the supervision is non-negotiable. That is not a philosophy you can bolt on after a consultant leaves. It is the thing the consultant was never selling.
What advisory is actually worth — and how to buy it
I am not saying judgment is worthless. The opposite. A genuinely senior technical voice early — what to build, what to refuse to build, what will not survive your data, where the agent should hand off to a human — is worth a great deal. We sell exactly this as fractional CTO work, in the rough range of $4,000 to $16,000 a month depending on depth. But notice the difference: that advice comes from the people who also build, and it is priced as ongoing accountability, not as a one-time document. The test is simple. Ask the firm pitching you agentic AI consulting whether the same people will write the eval suite and carry the pager. If the answer is 'no, that's a separate engagement,' you are buying the map and renting the territory to a stranger.
There is an honest, cheap way to find out, too. We start engagements with a free discovery week and then a paid pilot with no lock-in — a small, real slice of the actual system, built and instrumented, not described. A week of building tells you more about whether an agentic use case is viable than a six-week assessment ever will, because the loop either holds up against real inputs or it doesn't. A pilot has skin in the game. A roadmap has a footer.
The position, stated plainly
Agentic AI is real, and the productivity it unlocks where it works is not hype. But 'agentic AI consulting' as a standalone product — strategy decoupled from the people who ship and operate — is the most reliable way to land in Gartner's cancelled 40%. The agent doesn't fail because you lacked a maturity model. It fails in the loop, in production, in the parts no slide describes. Don't buy agentic AI consulting. Buy agentic AI engineering, from people who will still be there when input number 4,000 arrives — and treat anyone who wants to advise without building as exactly what they are: a confident voice with nothing at stake.
Last updated July 1, 2026