Praful Saklani is CEO of Pramata, a leading contract intelligence platform.
Drop 10 contracts into your favorite AI tool. Ask it to summarize them, flag the risks and tell you how to negotiate. In about two minutes, you’ll get back something that looks spectacular. Clean graphs, confident language, perfectly structured output.
Now look closer. It cited the wrong termination clause, the pricing analysis confused two counterparties and the “risk” it flagged actually says the exact opposite of what it claimed.
Welcome to the uncanny valley of enterprise AI.
Three Thresholds To Vet AI
The concept of the uncanny valley stems from that eerie feeling people got from seeing early robots that felt too humanlike. The robots that looked highly realistic but fell short, leaving an unnatural sensation that they could not be trusted.
Enterprise AI has its own version of this uncanny valley concept. The outputs look polished and authoritative until you peel back the top layer and realize there isn’t much substance.
Through years of working with AI tools, I’ve developed a three-threshold framework to vet outputs and verify results. The bad news is that most AI only clears the first one:
1. The Demo Threshold: Is this theoretically interesting? Nearly all AI clears this bar. Demos are impressive and get people excited in the boardroom.
2. The Usefulness Threshold: Will this actually save me time, or will I spend more time fixing its outputs than doing the work myself?
3. The “Don’t Get Fired” Threshold: Can I trust this enough to act on it without putting my job, my company or my customers at risk? This is the agentic threshold, and almost nobody is clearing it.
The uncanny valley lives in the gap between the first threshold and the last two. It is the space between slick-looking outputs and the reality that many of these outputs require the user to spend a lot of time vetting and massaging before they can trust the information to be correct and useful.
The gap between “good” and “good enough” in the enterprise is massive. Asking AI to summarize a single contract with 85% accuracy isn’t a big issue if a human is also reviewing it. But if that contract is tied to others governing $10 million in revenue, 85% is not good enough. Analyzing five related contracts at 80% to 90% accuracy each means combined accuracy could drop to 30% to 50%.
The Four Tiers Of AI Accuracy
It is helpful to think about AI accuracy in the following four tiers to clarify exactly where the uncanny valley sits and why it’s so wide:
• Tier 1. AI Slop (Below 50%): Sounds intelligent, but any practitioner can spot errors in 30 seconds. Most demoware lives here.
• Tier 2. The Uncanny Valley (50% To 85%): AI fills in blanks like an expert, even when it doesn’t know the answer. Output looks useful, but isn’t. The only way to catch gaps is brute-force audits. This band is enormous, and you can go far without realizing you’re still stuck.
Tier 3. Human In The Loop (85% To 95%): Useful, but requires human review. Errors need steering by a practitioner before decisions can be trusted. Most of the best-in-class enterprise AI deployments sit here today.
Tier 4. Automation ready (95%+): The ultimate unlock. Information can safely feed downstream processes and agents without a human in the loop, if you can precisely identify the ~5% that is likely to be inaccurate. This is the agentic threshold for enterprises to be able to trust AI with high-stakes workflows.
The uncomfortable truth: The uncanny valley, spanning between 50% to 85%, is where the vast majority of impressive demos and overpromised vendor solutions actually live.
Crossing the Valley Requires Real Work
I know this firsthand. Two years ago, my leadership team and I got together for a multi-day executive hackathon to see how LLMs could support our contract intelligence platform.
Upon first impression, we were blown away. Automating basic functions like metadata extraction and document classification seemed like a straightforward capability but we were in the uncanny valley.
Across 100 documents, the output looked miraculous. However, under a fine-tooth comb, the usefulness of the tools was stuck at 50% to 60%. The initial outputs included wrong account names, master agreements that were misclassified as statements of work and termination language pulled from the wrong section of contracts.
There are strategies to address these issues, but there are no shortcuts, even for the “simple stuff.”
The Real Question
The uncanny valley experience goes like this: “Oh my God, this changes everything!” followed a week later by, “Oh my God, if I’d gone forward with this, I would have gotten fired.”
That cycle will repeat across every enterprise use case until organizations treat AI deployment as a serious engineering and domain expertise challenge, not a demo.
The question every business leader, enterprise and otherwise, should be asking is whether their AI has actually cleared the uncanny valley or if they are just looking at a very convincing illusion.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
