Your AI Doesn’t Need to Be Smarter. It Needs to Be Specific.
Here is the uncomfortable truth most AI vendors will not tell you.
A model can reach state-of-the-art accuracy and still deliver zero business value.
MIT’s Project NANDA reported in 2025 that roughly 95% of enterprise generative AI pilots showed no measurable P&L impact, according to this report.
A separate EBU/BBC study found that 45% of AI assistant answers about news had at least one significant issue, per this study.
That one covers news questions, not business data. The pattern still holds.
The problem isn’t the model.
It’s what the model was never told.
“Accuracy doesn’t mean impact.” — Sengupta of Aible, in this article
Your Model Scored 90%. So Why Did the CFO Stop Trusting It?
Your analytics team spent months benchmarking models.
They pushed the accuracy score above 90%.
The dashboard looks clean.
Then the CFO asks one follow-up question.
The AI confidently answers wrong.
This isn’t a hallucination in the usual sense.
It’s a context failure.
The model may not know your comp store definition.
It may not know your regional VP always checks labor cost before occupancy.
It treats every metric as equally interesting.
A June 2026 benchmark from Promethium, a vendor in this space, found that only 16% of AI answers to open-ended enterprise questions met the accuracy bar for business decisions, per this guide.
Treat that as directional, since it’s a vendor benchmark.
But it points the same way as everything else here.
The Accuracy Trap: Why High Scores Quietly Destroy Trust
Here is what breaks in production.
A senior executive at a large national retailer put it bluntly, as shared in this post:
“My team can do that… while it is faster, it’s not incremental.” — Senior executive, national retail chain
Faster is not a business case.
Incremental is.
And incremental requires interpretation, not just inference.
Stanford’s 2026 AI Index counted 362 documented AI incidents in 2025, up from 233 in 2024, per this report.
Incidents are rising even as capability accelerates.
Here’s what that looks like on the ground:
- LLMs struggle with basic math by structural design, not by bug, as this breakdown explains.
- Open text boxes hand the hardest part of analytics back to the user.
- Generic AI produces fluent but untrustworthy answers.
- A confident wrong number looks exactly like a right one.
This is the enterprise trust problem nobody is fixing at the model layer.
The Fix Isn’t a Bigger Model. It’s a Harness Built Around Yours
Engineers have a name for the solution.
Get Emma Smith’s stories in your inbox
Join Medium for free to get updates from this writer.
They call it a harness, as Brad Peters explains in this post.
The model is the engine.
The harness is the car.
A proper harness tells the AI five things:
- What to screen for at every location, every cycle.
- Which thresholds separate normal variance from a real problem.
- What to drill into next when something trips.
- How to interpret findings in your business terms.
- What it is not allowed to do.
This isn’t a constraint.
It’s the reason people actually trust the output.
“The harness supplies the judgment. And the judgment has to come from you.” — Brad Peters
The same post notes that Vercel’s engineering team found its internal data agent got better after they removed 80% of its tools.
More options had made it worse.
Why a 90% Accurate Model Still Hands You 1,000 Wrong Answers
The race for model accuracy has diminishing returns for business decisions.
Say a text-to-SQL model is 90% accurate.
It still gets 10% of queries wrong.
Run 10,000 analytical queries a quarter and that’s 1,000 wrong answers dressed up as confident reports.
The numbers are illustrative, but the math is simple.
That’s the gap this explainer on agentic analytics addresses.
It pairs the model with structured systems built for exact numbers and your operators’ interpretation rules.
The best human-in-the-loop AI, as this guide describes, doesn’t replace your COO.
It runs their logic across every location, every cycle, automatically.
What Your CFO Needs to Hear Before Signing the Next AI Contract
You’re not buying a smarter model.
Every vendor has access to roughly the same models.
You’re buying your own judgment, running at scale.
The harness is the asset, because it holds interpretation no competitor has.
That’s why AI demos well but fails in production, as this read explains.
A fluent demo isn’t a business case.
Watch: The practical breakdowns on this YouTube channel are a good place to see how operator judgment gets captured at scale.
Read next: This full breakdown from Brad Peters on why the harness is the real asset.
Connect with Scoop Analytics:
LinkedIn | Facebook | Instagram
