The money isn't in AI agents. It's in the plumbing.

In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027. The reasons it gave were escalating costs, unclear business value, and weak risk controls. Notice what is missing from that list. Not the model. AI agent projects are not dying because the model is not clever enough. They die because the work underneath the agent never got done.

XY Space has spent ten years putting AI into production in insurance, legal, biotech, and health. The pattern rarely changes. A team asks us for AI agents. What they actually need is to see what is happening inside their own operation, and to connect the data they already hold. The agent is the last mile. Nearly all of the value, and nearly all of the risk, lives in the road you build to reach it.

AI agents were never the hard part

Start with the numbers, because they are blunt. MIT's NANDA initiative reviewed 300 deployments and found that 95% of enterprise generative AI pilots produce no measurable impact on profit and loss. The barrier was organizational, not technical. The most revealing detail sits in the budgets: more than half of generative AI spending goes to sales and marketing tools, while the largest returns showed up in unglamorous back-office automation. Companies fund the demo and starve the plumbing.

Production data points the same way. Independent analyses of enterprise rollouts put the failure rate in production as high as 80%, and blame broken data infrastructure far more than weak models. Duplicate records, systems that disagree, stale fields, missing context. An AI agent reading that will output something that looks like a hallucination and is really a data problem. Gartner even has a term for the vendor side of the hype, "agent washing," and reckons only about 130 of the thousands of self-described agentic vendors ship anything genuinely agentic.

None of this argues against AI agents. It argues about order of operations. What an AI agent is, a system that pursues a goal and takes actions, only works when it can see clearly and act safely. Both of those are foundation, not model.

What "data plumbing" actually means

Plumbing sounds dismissive. It is the opposite. It is the part that decides whether the whole thing holds together. Three layers sit under any AI agent worth trusting.

Visibility. Can you see what is actually happening across your operation, right now? Most teams cannot, because the picture is spread across a dozen tools that do not talk to each other. Before any automation, the first real product is a clear, current view of the work.

Connection. The data you need almost always exists already. It just lives in separate systems with no shared spine. Joining it up, so that a claim or a legal matter or a shipment becomes one coherent record instead of scattered fragments, is the part everyone skips. It is also most of the job.

Guardrails. Context the model can lean on, retrieval that grounds it in your own facts, a human in the loop at the decisions that carry consequences, and a tamper-evident audit trail. In a regulated operation, none of that is optional polish. It is what keeps an AI agent safe in front of real work.

That is the whole game. Get the groundwork right and the agent on top is almost boring to add; get it wrong and you are the 40% Gartner keeps warning about.

The real sequence: see, connect, then automate

There is an order to this, and reversing it is the expensive mistake. See the operation. Connect the data. Then, and only then, automate the last mile.

It is the same reason we open every engagement with a paid map instead of a build. The map answers a plain question: what is actually happening here, and where is AI worth it? By the time an agentic workflow goes live, the hard part is already behind you. The agent gets added once the ground beneath it is safe, every figure cited to your own rules and open to audit. We run AI the way a good manager hires: one job at a time, supervised, measured in hours and dollars.

Where the value actually shows up

Look at what pays off in practice, and it is rarely the flashy autonomous agent. It is the back office. In one insurance claims-intake system we run in production, manual data entry fell by half. A legal-sector review we ran surfaced £178,000 of recoverable leakage that manual sampling had missed. A client-correspondence system cut non-billable email time for attorneys by more than 60%. You can read the production case studies with the full numbers.

None of those wins came from a cleverer model. They came from seeing the work clearly, connecting the records, and putting a person at the point of decision. It tracks with what MIT found: the return is in the operational core, not the marketing gadget. The unglamorous middle is where the money has been the whole time.

So should you build an AI agent?

Yes, eventually. The point is not that AI agents are worthless. It is that they are the final ten percent, not the first. A visual tool can stitch together a quick automation, and for light work that is genuinely fine; the trade-offs are laid out in n8n versus custom AI agents. But when the AI carries a regulated process, the durable system is the one built on real foundations, and the agent is simply the last component you add. Buy the outcome, not the buzzword.

Common questions

Are AI agents overhyped?

The technology is real; the marketing is inflated. Gartner expects more than 40% of agentic AI projects to be canceled by 2027 and calls out widespread "agent washing," where old chatbots and RPA get relabeled as agents. Treat the category with interest and the vendor claims with suspicion.

Why do most AI agent projects fail?

Not because of the model. Failures trace back to data quality, fragmented systems, unclear business value, and missing guardrails. MIT found that 95% of enterprise generative AI pilots deliver no measurable ROI, and the causes were organizational rather than technical. Fix the foundation and the success rate changes.

What should you build before an AI agent?

Visibility and connected data. First, a clear view of what is actually happening across your operation, in real time. Second, the integration work that turns scattered records, spread across systems that were never meant to talk, into one trustworthy source. Third, guardrails: grounding, human review, an audit trail. Only then, the agent.

Is the ROI in AI agents or in automating the basics?

For most companies today, the basics. MIT's data shows the biggest returns in back-office automation rather than the headline use cases. Document processing, reconciliation, and compliance checks quietly return more than the autonomous agent that gets the press.

The plumbing was the point

The people who have actually shipped tend to converge on the same conclusion. The money was never in the agent. It was in the plumbing: the visibility, the connected data, and the guardrails that make automation safe to run. Get that right and the agent is easy. Get it wrong and you join the 40% that quietly get canceled. If you want to know where AI is genuinely worth it inside your operation, that is exactly what a map is for. Start there, not with the agent.