Since 2 August 2026, Article 50 of the EU AI Act has required that people be informed when they are interacting with an AI system, unless that is already obvious (EU AI Act, Article 50). The standing objection to AI disclosure was never legal. It was commercial: tell the customer and you lose the sale. A randomised field experiment reaching 1,337 customers has now tested that objection in the LLM era and found no conversion penalty for disclosure (Habel, Zawadzki and Freise, 2026).
The second finding in the same paper is the one that costs money.
The Objection That Just Lost Its Evidence Base
The pre-LLM evidence said customers were prejudiced against AI, so revealing it suppressed sales. Habel (University of Houston) with Zawadzki and Freise revisit that result across three contemporary studies (Habel, Zawadzki and Freise, 2026).
| Study | Design | Result |
|---|---|---|
| Field experiment 1 | Randomised, 1,337 customers reached | Disclosing an AI voice agent no longer reduces conversion relative to nondisclosure |
| Field experiment 2 | Randomised, 366 customers reached | AI voice agents still substantially underperform human sales agents |
| Online experiment | 238 participants | Customers largely recognise AI voice agents regardless of disclosure, and evaluate disclosed and nondisclosed AI similarly |
Read the third row against the first. The reason disclosure stopped costing anything is not that customers became tolerant. It is that the label stopped carrying new information. People work out what they are talking to on their own, so concealment buys nothing measurable while leaving an Article 50 exposure on the books.
One caution before anyone quotes this upward: the paper reports no conversion-rate percentages. The claim is that disclosure produced no reduction, not that it produced a lift.
What AI Disclosure Actually Costs You
Close to nothing, on this evidence โ which makes the operational question not whether to disclose but where the disclosure lives. If it is a setting, somebody will eventually switch it off to test uplift, and that test is now known to have no upside and a regulatory downside. Identity belongs to the surface the customer is looking at, in the same category as a company name on an invoice.
What customers do push back on is separate from the label. In the online study they were less willing to continue the conversation because they anticipated lower empathy, less human-likeness, more effort on their side, and greater insincerity from the firm. That is a forecast about the quality of the interaction, not an objection to being told. You cannot fix it by going quiet. You fix it by making the forecast wrong.
There is also a cost to the concealment option that rarely gets priced. Keeping the identity ambiguous means every downstream claim about the interaction โ what was said, what was promised, who was accountable โ has to be reconstructed later from a transcript whose status nobody declared at the time. Disclosure is cheap partly because it is the only version of the conversation that is straightforward to stand behind afterwards. The alternative is not a free uplift; it is a contingent liability with no measurable revenue attached to it.
The Gap Is Four Named Process Failures
The second field experiment is where the money is. Against human sales agents, the AI voice agents substantially underperformed โ and the authors locate the cause precisely. Contemporary agents continue to struggle with sales process discipline: stakeholder qualification, value communication, persistence, and next-step orchestration.
Look at that list as an operator rather than as a buyer of technology. Not one of those four is a reasoning problem. Qualifying the right stakeholder is a rule about who the conversation is worth having with. Value communication is an argument somebody wrote down. Persistence is a cadence. Next-step orchestration is a calendar and a definition of what counts as a next step. These are the contents of a role manual โ and an agent deployed without one is being asked to invent a sales process on the call.
The Four Are Not Equally Hard to Fix
Persistence and next-step orchestration are the two that a mid-market team can specify in an afternoon: a follow-up cadence with intervals, and a single unambiguous definition of a booked next step. They are also the two that fail silently, because nothing in the system complains when a conversation simply stops. Stakeholder qualification needs a rule that someone with commercial authority has to write, which makes it a week of work rather than an hour. Value communication is the genuinely hard one โ it is the part that depends on reading which argument this buyer responds to, and the part where a human still has a clear edge. Sequence the work accordingly: fix the two mechanical failures first and judge the agent again before concluding anything about the other two.
Why This Reads as a Role-Design Deficit, Not a Model Deficit
There is independent support for reading the gap this way. MIT Technology Review Insights and Microsoft ranked 101 agentic-AI tasks on a 0โ100 trust scale with 300 executives and practitioners, and found that confidence tracked task verifiability and business-context completeness rather than model capability: automated report generation scored 83.5 and boilerplate code 82.5, both with a single objective success metric, while service-mesh configuration scored 37.5 and disaster-recovery testing 43 โ same underlying models, no clean definition of done (MIT Technology Review Insights and Microsoft, 2026).
A sales conversation with no stated qualification rule and no defined next step is a task in the bottom group. The model is not what is missing. The definition of done is.
That survey also found 59% of respondents already planning permanent human oversight rather than treating supervision as a transitional cost. Worth noting before anyone budgets an agent as a straight headcount substitution.
Supervision Has a Placement, Not Just a Headcount
If supervision is permanent, where it sits in the task starts to matter more than how much of it there is. A CHI '26 study modelled confirmation frequency in multi-step agentic tasks and tested it with 48 participants: 81% preferred intermediate confirmation over the confirm-at-the-end pattern most systems ship with, and task completion time fell by 13.54% (Zhou et al., 2026).
Faster, not just safer. The rollback cost of an error caught late exceeds the interruption cost of a check placed early. Translated to a customer conversation: the useful checkpoint is at the stage boundary โ after qualification, before the value argument carries the deal โ not a transcript review after the opportunity has already gone cold.
The Honest Counter
Four limits, stated plainly.
This is a working paper. SSRN 7013259 was posted in July 2026 and revised on 22 August 2026. It has not completed peer review, and the figures available are abstract-level.
Two of the three studies are small. 366 customers reached and 238 online participants are thin for anything beyond a directional read, and all three studies sit in one sales context. A B2B enterprise cycle with six stakeholders is not the setting tested here.
Both headline findings are 2026 snapshots. "Customers recognise AI anyway" depends on today's agent quality and today's public exposure to it; so does "agents underperform humans." Both numbers move, and not necessarily in the same direction.
The supporting evidence is not performance data. The MIT and Microsoft figures are perception scores from 300 respondents, and the CHI '26 result comes from a lab task rather than a live sales call. They corroborate the shape of the argument, not its magnitude.
What survives all four is the separation the paper forces: the disclosure decision and the capability decision are not the same decision, and only one of them was ever expensive.
What to Change This Quarter
Four moves. None of them is a purchase.
- Make identity a property of the agent's surface rather than a configuration flag. If a sales lead can switch it off to run an uplift test, that test will happen, and this evidence says it has no upside.
- Write the process spine before deployment, not after the pilot disappoints. Name the qualification rule, the value argument, the follow-up cadence, and what counts as a booked next step. An agent without these four is being measured on a process nobody specified.
- Put the supervision checkpoint at the stage boundary. After qualification is where a misdirected conversation is still cheap to redirect; a post-hoc transcript review is the confirm-at-the-end pattern that the CHI '26 result argues against.
- Benchmark the agent against your own human baseline on next-step booking rate, not on call volume or customer satisfaction. Volume is what agents are good at; the next step is what this paper says they lose.
One records point sits underneath all four. If you ever have to evidence the disclosure โ which provider answered, under what terms, with what human accountable for the conversation โ that is a documentation problem you solve before the question arrives, not during an audit. Scovai publishes its own provider naming and EU AI Act posture at transparency as one worked example of what that record looks like.
The useful reframe is that AI disclosure was never the growth variable anyone treated it as. It was a line of copy. The sales process behind it was always the variable, and it is the one nobody wrote down.
So pull one number before the quarter closes: of the conversations your agent held last month, what share ended with a booked next step โ and what share of the equivalent human conversations did? The difference between those two figures is your actual deployment cost, and it has nothing to do with what you told the customer.
