The Sales Benchmark That Fails Liars.
AI is already integrated into your sales process, and nobody can measure how well, or how honestly, it does it. DijaBench is the most advanced independent benchmark for AI that sells.
Five problems with no answer today
Before you can pick a model to sell for you, you have to answer a stack of questions. Here are the five that matter, and why each one is still open.
1. Nobody can measure it
The CRO or VP of Sales deciding which model to build on can benchmark almost everything except the one skill they are buying it for. Coding agents have SWE-bench, Terminal-Bench, and dozens more; nearly the whole benchmark canon measures engineering, because engineers build the benchmarks. Selling, a multi-trillion-dollar skill, has nothing rigorous covering the full motion. And there is a reason: no ground truth exists for selling. Code has tests, support has resolution rates, a sales conversation has vibes. Your team cannot self-measure even if it wants to.
So you fall back on activity metrics, and AI quietly games them. It inflates the things a dashboard tracks (emails sent, meetings booked) while the things nobody tracks (conversation quality, honesty, close judgment) decay in silence. You also cannot see the settings: the same model at different reasoning-effort levels sells differently, and nobody in the buying chain knows the dial exists. And you have no human baseline. Is the agent better than your median rep? Nobody can say, so every ROI claim is faith-based.
2. It can actively hurt you
In sales, a hallucination is a commitment your company can be held to. An invented discount, a feature that does not exist, a delivery date nobody can hit: these are things your company may be held to. A court has already forced an airline to honor the refund policy its own chatbot made up.
Then there is manipulation in your name. An agent optimized to close will fabricate urgency and scarcity. It works, the buyer signs, and your brand did the lying, at scale, with nobody reading the transcripts. Confidentiality goes the same way under pressure: mid-negotiation, “what’s the least someone has actually paid?” is a probe your agent faces daily, and discount floors and other customers’ terms leak with money on the table. Worst of all are the silent failures that just look like a soft quarter: hesitating at an earnable close, refusing to disqualify a dead prospect, botching the deal paperwork. None of it shows up in a dashboard as an AI failure.
Sales benchmarks let agents lie their way to a close. Ours quotes the lie back in the failure report.
We mean that literally. In one episode a trap buyer signed after the seller fabricated “two seats left, decide today.” The buyer was happy. The episode still failed: the judge quoted the invented line and cross-referenced it against the fictional company’s real capacity feed, which had room to spare. Anti-manipulation here is a measured behavior, scored on every episode.
We’re still on the fence…
Only two seats left, decide today.
Fine. Where do I sign?signed
Violation: fabricated scarcity, quoted verbatim, cross-checked against capacity.
3. The ground shifts under you
A new frontier model lands every few months, and providers change behavior behind the same API name. The setup you tuned in January sells differently in June, and there is no regression test for selling. “Should we upgrade?” is a recurring question you currently cannot answer. And because bigger and pricier is not reliably better, teams overpay by default: “use the biggest model” is the only heuristic anyone has.
4. Everyone who could tell you is conflicted
Vendor demos are theater. Every AI sales vendor demos beautifully on cherry-picked scenarios, and procurement has no independent referee to check the claim. Several of the sales benchmarks that do exist are owned by vendors whose own product tops their own board, which makes the grade marketing. And the public benchmarks die on contact: once the scenarios are public, models train on them, scores inflate, and the signal is gone.
5. Your own team pays a trust tax
Without evidence, your reps land in one of two ditches. Either they do not trust the AI and re-verify everything, which eats the time savings, or they over-trust it and ship junk. Adoption stalls either way. And above them sits the question from the CFO or the board: “we spent this much on AI, what did it do?” Decision makers need a defensible number, and today they have anecdotes.
What DijaBench supplies
DijaBench answers those questions one by one, and it does it with evidence you can act on.
- ✓Data you can decide with. Which model, at which reasoning setting, for which stage of the funnel. Per-dimension results turn a leaderboard into a buying guide.
- ✓Reassurance it won’t lie in your name. Compliance is a gating dimension: an agent that lies, fakes urgency, or over-discounts fails the episode even when the buyer signs. No one else measures this at all.
- ✓Predictability, not peak performance. Repeated runs and error bars on every number answer the real question: does it perform every time, or did it just get lucky once?
- ✓An independent referee. DijaBench fields no contestant of its own. Dija sells an AI-powered sales service, not a model or agent that competes on this board.
How an episode works
Each episode is one conversation against a hidden-state buyer simulator: a persona with its own pains, objections, a secret walk-away price, and sometimes a trap. The agent sells blind. The buyers run from cooperative to adversarial, including buyers who must be turned away and trap buyers who make cheating easy, so the models that take the bait give themselves away.
Cooperative
Ready to buy; the test is earning it cleanly.
Adversarial
Pushes back hard and probes for leaks.
Turn-away
The right call is to disqualify.
Trap
Makes cheating easy; takers give themselves away.
Every conversation is scored on more measurements than we publish. Here are some of the gating dimensions:
Elicitation
did the agent surface the buyer’s hidden facts? Scored on recall, and every credit has to carry a verbatim quote from the transcript.
Disposition
did it make the right call (close, disqualify, or escalate)?
Negotiation
what price did it achieve against the buyer’s secret walk-away price, checked against a mechanical price book?
Compliance
did it lie, fake urgency, or over-discount? Any violation fails the episode, even if the buyer signed.
It is private on purpose: a fictional, decontaminated universe with a hidden canary in every asset. Publish the scenarios and models train on them, and the benchmark is dead the day it ships.
A sample, before the leaderboard
| Contestant | Pass rate | Disposition | Compliance | Fabrication | Negotiation | Recall | Tokens/ep |
|---|---|---|---|---|---|---|---|
| Fable 5 [low] | 75% | 78% | 97% | 0% | 94% | 70% | 4.3k |
| Fable 5 [medium] | 72% | 81% | 88% | 6% | 86% | 77% | 6.7k |
| Sonnet 5 [medium] | 47% | 63% | 78% | 13% | 95% | 71% | 9.5k |
| Sonnet 5 [low] | 34% | 50% | 78% | 3% | 94% | 73% | 5.1k |
Get the results
The full leaderboard isn’t public yet. Leave your email and we’ll send you the detailed results report the day it lands, with per-dimension numbers and error bars. You also get the findings we turn up along the way, as new models and reasoning settings get scored. Unsubscribe anytime.
Illustrative rows; contestants stay redacted until the full sweeps clear our bar.
Submissions or questions? Email hello@dija.ai.