Why the Defensible Number Wins
The counterintuitive case for AI that undersells.
The Architecture Series · 9 minute read · By Ed | Founder & CEO, DealiOS
Here's the feature request nobody makes: "Can you make the AI less impressive?"
Nobody asks for conservative ROI projections. Nobody asks for understated savings estimates. Nobody asks for the AI to deliberately leave money on the table in the pitch.
But that's exactly what builds trust. And trust is the only asset that compounds.
An ROI projection that says you'll save 60% gets a raised eyebrow. One that says 25-30% gets a nod, a follow-up question, and a meeting with the CFO. The first number is more exciting. The second number closes the deal.
Trust is the only asset that compounds. The architecture should compound it.
This paper is about why we built a system that intentionally understates its own projections, and why that architectural decision turned out to be the most important thing we did for credibility.
The Overstatement Trap
AI hallucination in business documents isn't the failure mode people think it is. Everyone worries about the AI inventing a client name or fabricating a statistic from nothing. That happens, and it's bad. But it's also obvious. Someone catches it.
The real failure mode is credible overstatement: numbers that are plausible enough to survive review but optimistic enough to destroy trust when reality hits. A 47% efficiency gain sounds reasonable in a proposal. It passes the CFO's sniff test. The deal closes. Six months later, the actual gain is 22%. Nobody is celebrating. The conversation that follows isn't about the 22% improvement you delivered. It's about the 25% shortfall you promised.
A 2024 Deloitte study found that 38% of business executives reported making incorrect decisions based on hallucinated AI outputs [1]. That's not obscure-edge-case territory. That's more than a third of the people running companies getting steered wrong by AI-generated content they trusted.
The AI can generate any number. It defaults to optimistic because its training data is full of marketing copy and case studies that always show the best possible outcome. Nobody publishes the case study where ROI came in at 15%. The training data is structurally biased toward overstatement, and the AI inherits that bias without knowing it.
Every vendor's ROI calculator says you'll save 40%. If every vendor's ROI calculator were right, companies would be paying negative money for software by now. We'd owe them nothing. The software would owe us. That math hasn't worked out for anyone except the vendor.
The architecture problem: how do you build a system that systematically resists its own tendency to impress?
Calibration Over Precision
There's a concept from forecasting science that most product teams have never encountered: calibration. Calibration measures how well your confidence matches your correctness. A perfectly calibrated system that says "25-30% savings" delivers within that range 90%+ of the time. A precise system that says "41.7% savings" might be off by 15 points in either direction.
The calibrated system is less impressive. It's also more valuable.
A system that says "25-30% savings" and delivers 28% is more trustworthy than one that says "41.7% savings" and delivers 35%. The second system was technically closer to a specific prediction. The first system did what it promised. In enterprise sales, doing what you promised is the entire game. Because the renewal conversation happens twelve months after the initial deal, and the only thing the customer remembers is whether you hit the number you quoted.
Roughly 40% of SaaS revenue now comes from renewals and expansion, according to Benchmarkit's 2025 data [2]. The initial deal isn't the finish line. It's the starting gun. An overstatement that closes the first deal and poisons the renewal is net negative over the customer lifecycle. You traded a short-term win for a long-term loss, and you did it with a number your AI generated in 200 milliseconds without understanding the consequences.
The behavioral economics angle reinforces this. Loss aversion, one of the most replicated findings in decision science, means that losses feel roughly twice as painful as equivalent gains feel good (Kahneman & Tversky, 1979) [3]. A client who expected 40% and got 30% feels cheated. A client who expected 25% and got 30% feels delighted. Same outcome. Different anchor. The anchor is the architecture's job.
The AI doesn't need to be less capable. It needs to be less impressive. There's a meaningful difference, and the architecture enforces it.
Building Conservative Into the System
"Just tell the AI to be conservative" doesn't work. I know because we tried it.
Prompt-level instructions are suggestions, not constraints. You can write "always provide conservative estimates" in the system prompt and the AI will still occasionally produce a 65% efficiency gain because the retrieved case study mentioned one, and the AI pattern-matched on it rather than applying the conservatism instruction.
The fix is architectural. The numbers that matter most in a business proposal don't come from the LLM. They come from a deterministic calculation engine that can't hallucinate because it doesn't generate. It computes. The calculator is the single source of truth for every financial figure in the proposal. It takes inputs, runs formulas, and produces exact numbers. No AI involved in the math.
The conservative framing works at two layers. First, the prompt instructions tell the AI to use conservative estimates when contextualizing those numbers: if benchmarks say 40%, model at 25-30%. Second, a build validation step catches literal numbers the AI might try to invent on its own and forces them back through the calculator. The AI contextualizes and frames. The calculator owns the math. The conservatism is a deliberate design decision enforced through prompt architecture and output validation, not a happy accident.
Here's how it breaks down across the financial sections of a proposal:
Cost of Inaction goes through the calculator, not the LLM. The cost of doing nothing is a computable figure based on the client's inputs: current headcount on the task, fully loaded cost per hour, hours per week, weeks per year. Multiplication. No room for the AI to embellish.
Total Investment requires completeness. This is where most vendor proposals cheat by omission. They show the license cost and leave out implementation, training, integration, and ongoing maintenance. Hidden costs destroy trust more aggressively than high costs do. A client who discovers a $50,000 integration cost after signing a $200,000 deal doesn't think "well, the base price was good." They think "these people hid costs from me." The architecture enforces completeness by requiring all cost components before the calculation runs.
Payback period is computed deterministically. Conservative investment divided by conservative annual savings. No LLM involved. No rounding in the favorable direction. No "approximately" hedging that lets the AI pick a flattering number.
Cost of delay sits in the Next Steps section, monetizing procrastination. This is the one persuasion lever that uses real numbers to create urgency without inflating the value proposition. "Every month of delay costs $X in unrealized savings." The $X comes from the calculator. The urgency comes from the math, not from the AI generating alarming language.
Only 14% of CFOs reported seeing clear, measurable impact from their AI investments in a 2025 RGP survey, even as 66% expected returns within two years [4]. The gap between expectation and demonstrated value is where trust erodes. Conservative projections that the client can verify against actual results twelve months later close that gap from the other direction. You set the bar where you know you can clear it.
The Variance Problem
Precision creates a different kind of trust risk that most teams don't think about: variance anxiety.
A projection of "41.7% savings" carries an implied precision that suggests the system computed this to the decimal point. The client unconsciously anchors on 41.7%. Not "around 40%." Not "between 35-45%." Exactly 41.7%.
When reality comes in at 36%, the miss feels enormous. 41.7 minus 36 is 5.7 percentage points. The projection was "wrong by 14%."
Now consider a projection of "25-30% savings." Reality comes in at 36%. The projection wasn't wrong. It was conservative. The client got more than expected. Same outcome. Different frame. The first frame damages trust. The second frame builds it.
The false precision problem is endemic in AI-generated business documents because AI loves producing specific numbers. Ask it for an estimate and it'll give you 41.7% instead of "roughly 40%" because specificity was rewarded in its training data. Specific numbers appear in authoritative sources. Hedged ranges appear in casual conversation. The AI learned that specificity signals confidence, which is exactly backwards in contexts where the number is uncertain.
The architectural fix: the calculator produces the real number, and the prompt layer frames it conservatively. The AI doesn't get the chance to false-precision its way into a fabricated figure because every financial number in the proposal traces back to the deterministic engine. When the AI contextualizes a projection, the prompt instructions enforce conservative framing. When the output validation layer detects a literal number the AI invented rather than pulled from the calculator, it flags it. The system is designed so the AI explains numbers it didn't generate, rather than generating numbers it can't explain.
Where Conservatism Meets Credibility
There's a counterintuitive dynamic at play in enterprise sales: the vendor who quotes the lowest ROI often wins.
Not because buyers want low ROI. Because buyers distrust vendor-provided projections by default, and a conservative projection signals that the vendor either has nothing to hide or has the confidence to let the numbers speak without embellishment.
An RGP survey of 200 U.S. CFOs found that only 10% said they fully trust their enterprise data [4]. If finance leaders don't even trust their own internal data, the likelihood that they'll trust a vendor's AI-generated ROI model is approximately zero. You're not competing against their skepticism of your numbers. You're competing against their skepticism of all numbers.
Conservative estimates are a credibility arbitrage. Every other vendor generates the most impressive projection. Yours generates the most defensible one.
Conservative estimates are a credibility arbitrage. Every other vendor's AI is generating the most impressive projection it can. Yours is generating the most defensible one. In a market where buyers assume they're being sold to, the one vendor who appears to be underselling stands out.
This isn't modesty as a personality trait. It's calibration as an architectural decision. The system separates two concerns that most platforms blur together: computing the number and framing the number. The deterministic calculator owns computation. It produces the real figure based on the client's actual inputs. The AI owns framing, and the prompt architecture instructs it to frame conservatively. The AI can't inflate the math because it doesn't do the math. And output validation catches cases where the AI tries to invent a number instead of using the one the calculator produced. The conservatism lives in the prompt layer, not the computation layer, and that's a deliberate choice: the calculator stays neutral and accurate while the AI is instructed to contextualize with restraint.
The Universal Principle
Conservative estimation isn't specific to sales proposals. It applies anywhere an AI produces a number that influences a human decision.
Medical AI that overstates treatment efficacy risks patients choosing treatments based on inflated expectations. Financial AI that overstates portfolio returns risks investors taking on positions they can't afford to hold when reality corrects. Recruiting AI that overstates candidate fit scores risks hiring managers skipping deeper evaluation because the number looked good enough.
The pattern: an AI that produces a number, a human who trusts the number, and a real-world outcome that doesn't match. The damage isn't in the AI being wrong. It's in the human not knowing the AI was overconfident.
The 47% stat from the Deloitte Global AI Survey bears repeating in this context: nearly half of enterprise AI users have made at least one major business decision based on hallucinated content [5]. That's not AI failing. That's calibration failing. The AI was confident. The confidence was misplaced. The human couldn't tell.
Conservative estimation is a system-level response to this problem. You can't fix AI confidence calibration at the model level. The models are trained on data that rewards specificity and penalizes hedging. What you can fix is the architecture: route financial projections through a deterministic engine so the AI can't invent them, instruct the AI via prompt design to frame conservatively when it contextualizes those numbers, and validate the output to ensure no rogue figures slipped through.
The AI doesn't need to be less capable. It needs to be less impressive. There's a meaningful difference, and the architecture enforces it.
The Long Game
The most successful enterprise relationships we've observed share a pattern: the vendor's initial projections were conservative, reality exceeded them, and the renewal conversation was a celebration instead of a negotiation.
That pattern doesn't happen by accident. It happens when the architecture is designed to make understatement the default, overstatement impossible, and the gap between projection and reality a gift to the client rather than a liability to the vendor.
Trust compounds. An overstatement in Q1 doesn't just damage the Q3 renewal. It damages the reference call, the case study, the expansion opportunity, and every downstream relationship that customer would have generated. A conservative estimate that reality exceeds generates the opposite: a satisfied client, a strong reference, a case study with real numbers, and an expansion conversation where the client already trusts your projections because you've been right before.
Every vendor's ROI calculator is trying to impress. The architecture's job is to be right. Those are different goals, and the only way to guarantee you're optimizing for the right one is to separate computation from framing: let the calculator produce honest numbers, instruct the AI to present them conservatively, and validate the output to catch any creative improvisation.
Trust is the only asset that compounds. The architecture should compound it.
The takeaway: Separate computation from framing. The numbers that matter most should come from a deterministic engine that can't hallucinate. The AI's job is to contextualize those numbers conservatively, not to generate them. Validate the output to keep it honest.
References
1. Deloitte. "Enterprise AI Decision-Making Risks: Global Executive Survey Results." 2024. 2. Benchmarkit / Gainsight. "Customer Success Index 2024." Gainsight, 2024. 3. Kahneman, D. & Tversky, A. "Prospect Theory: An Analysis of Decision under Risk." Econometrica, 47(2), 263-291. 1979. 4. RGP. "CFO AI Readiness Survey." December 2025. 5. Deloitte. "Global AI Survey 2025." Cited in AllAboutAI Hallucination Report 2025.
This paper is part of The Architecture Series by Ed, Founder & CEO of DealiOS. For the weekly series and more technical deep-dives, visit dealios.app/library.
See how these principles come together. Explore the platform →