It’s so easy to dump everything on your LLM of choice.
Quick, easy, and on a small scale, genuinely impressive. Spin up a few agents, throw your data at a public model, and the demo looks great.
Then you try to run it at production scale, and the bill arrives.
Companies are getting hit with costs far higher than they anticipated — and costs that keep rising, because as models get better, they use more tokens. Suddenly, your AI project’s ROI is blown out. The marginal cost of your next output isn’t falling. It’s growing.
And you’re still burning expensive human time correcting the output.
The easy path is not the best path.
The good news is that most of the issues driving up your AI costs — and quietly corrupting your outputs — are fixable. But the fixes aren’t glamorous.
None of this is the hype you see in your feed every day. It’s the unsexy plumbing: the everyday processes and workflows that actually drive your business.
And plumbing has one job.
You don’t want your pipes to be clever. You want them to hold.
Here’s what that plumbing looks like, and why you need it.
Your Agent Chain is Playing the Whisper Game With Your Data. And Losing.
Remember the whisper game from when you were a kid?
One person whispers a message to the next, around the circle, and by the time it comes back, the original point is gone — and the context with it.
That’s exactly what happens in most multi-agent setups I review.
Agent 1 interprets your data. Agent 2 works from Agent 1’s interpretation. Agent 3 works from Agent 2’s paraphrase of that interpretation. By the output stage, nobody, including the model, is looking at the original source anymore.
Every handoff between loosely defined agents degrades the signal.
Nuance drops out. Assumptions creep in. Context that mattered at step one never survives to step four.
And here’s the part that stings: you built it that way.
The chain didn’t fail. It did exactly what a chain of vague job descriptions does.
The fix is unglamorous
- One agent, one job, clearly stated. If you can’t describe an agent’s task in a sentence, it’s doing too much.
- Define what each agent is NOT allowed to do. Boundaries matter as much as tasks. An agent that knows it must not summarize, infer, or “improve” the data won’t quietly do it anyway.
- Build in a loop. Each agent verifies its output against the original source — not against the previous agent’s summary — before passing it on.
That loop is what breaks the whisper game.
Instead of trusting the chain, every step checks back to the original message.
When the Whisper Message Was False to Begin With
Now make it worse.
What if the message being whispered around the circle was never true in the first place?
Overclaiming is when an agent states something that was never in your data. It fills a gap, adds a conclusion nobody asked for, or “improves” the answer with a fact it invented.
When an AI overclaims in a blog post, this barely matters. You read it, you fact-check it, you catch the nonsense.
You’re the guardrail.
Now take that same behavior and put it inside a workflow where content passes from agent to agent with no human in the loop.
Agent 2 invents a “fact.”
Agent 3 doesn’t question it — why would it? It arrived looking like every other input.
Agent 4 builds on it.
By the end, that invented fact isn’t a claim anymore. It’s the foundation. And the whisper-game degradation kicks in on top of it — except now the original message was false to begin with.
That’s not a bad output. That’s a catastrophic failure, laundered into your system one confident sentence at a time.
You don’t fix this by hoping the model behaves. You engineer it out:
- Ground every claim in its source. Each output should carry its receipt — the specific piece of source data it came from. A claim that can’t point to a source doesn’t get to be a claim.
- Put checkpoints between agents, with explicit pass/fail criteria. Not “does this look right.” Real gates: does every claim trace to a source? Are required fields populated? If it fails, the workflow stops — it does not pass degraded output downstream. Fail closed, not open.
- Separate extraction from generation. One agent pulls facts and is only allowed to pull facts. A different agent writes. The writer can’t add data because it never had permission to touch the source. You’ve removed the opportunity to invent.
- Add a critic. A dedicated agent whose only job is to check the previous output against the original source and flag anything unsupported. Its job isn’t to be helpful. Its job is to be suspicious.
“Add Guardrails” is the Most Useless Advice in AI Right Now
Everyone says it.
Nobody tells you what a guardrail actually is. So here’s the plain version.
A guardrail is not a vibe.
It’s not “we told the model to be careful.”
It’s not a friendly note in the prompt asking the agent to please not make things up.
A guardrail is a check that can fail the output and stop the line. If it can’t fail anything, it’s decoration.
Here’s the mistake I see most: people build guardrails out of more LLM calls.
They add an agent that asks “does this look right?”
But an agent that can overclaim is now guarding against overclaiming. You’ve hired the fox to check the henhouse.
The fix is to make your checks as deterministic as you possibly can.
Not “does this look good” — actual rules a computer can enforce without judgement:
- Does every claim carry a source ID? A rule can check that.
- Are all required fields populated, not null? A rule can check that.
- Does the output match the expected schema? A rule can check that.
Save the LLM for judging only the things a rule genuinely can’t. Everything a rule can check, a rule should check — because a rule doesn’t have an off day, doesn’t get creative, and doesn’t overclaim.
And decide the failure behavior before you ship, not after it breaks.
In a low-stakes chain, a soft warning might be fine. In a high-stakes one, a failed checkpoint halts the pipeline. It does not wave degraded output downstream with a warning nobody reads.
Fail closed, not open.
Good guardrails look boring. They’re rules, schemas, and hard stops — not clever prompts.
If your guardrail sounds impressive, it’s probably just another agent waiting to overclaim.
The Standard You Already Hold Every Human To
If all of this sounds like a lot of engineering discipline, here’s the thing: it’s just the standard your business already lives by.
Picture it.
Your analyst walks in with a revenue projection. You ask where the numbers come from.
They shrug.
“I just kind of… know.”
They’d be out of the room before they finished the sentence.
A CFO expects a model built on real data with real references. A top manager asks their analyst team for analysis with the sources attached.
That’s not bureaucracy — it’s the basic price of trusting a number enough to make a decision on it.
And yet, the moment AI produces the same forecast, somehow the rules change.
The output arrives with no sources, no references, no way to trace a single figure back to where it came from — and people accept it.
“It’s AI, it’s a black box, that’s just how it works.”
No.
That’s just how you let it work.
There is nothing about AI that requires it to be a black box.
An agent can attach a source to every claim it makes. Every figure carries a pointer back to the exact piece of data it came from. Every conclusion shows its receipts.
You’ve simply built the analyst’s referencing standard into a workflow.
That’s all it is.
No claim without a source.
A figure that can’t point to a source doesn’t get to sit in your forecast — whether a human produced it or a machine did.
The Payoff
Pull these threads together, and you get the real answer to the cost problem we started with.
Process map your whole workflow so each agent does its specific task — and only that task — instead of burning tokens on wide-open reasoning.
Pre-process your data.
Give each agent one clearly defined job and hard boundaries.
Verify against source at every step.
Enforce with deterministic rules that fail closed.
That’s how you get falling marginal costs and cost predictability at production scale.
That’s how you get an agent chain you can actually trust — instead of one that embarrasses you and quietly poisons every decision built on top of it.
The black box isn’t a property of AI. It’s a choice to stop asking the question you’d ask any human on your team: where did this come from?
Ask it.
Make your systems answer it.
Every claim. Every stage. Back to source.
Vance Reavie is an entrepreneur, CEO, AI/ML specialist, and strategic advisor with 20+ years of experience leading digital transformation and applied AI innovation across retail, marketing, health, finance, and government. He is the founder and CEO of Junction AI, where he is building a proprietary agentic AI platform designed to make AI more accessible, affordable, and impactful for businesses. He is a contributor to Forbes Agency Council, a frequent speaker on AI and innovation, and has engaged with Congress, the Department of Commerce, and the White House on AI policy