When the answer is no, customers accept it from a machine more readily than from a person. In the research, an unfavourable outcome was accepted 78.6 per cent of the time when it came from AI and 60.4 per cent of the time when it came from a human being.
For good news, the preference reverses. People were accepted 89 per cent of the time against 76 per cent for AI.
That pairing is worth sitting with, because it inverts what most of us would have guessed and it points at a decision almost every organisation is currently getting backwards.
Where the numbers come from
MIT Sloan Management Review has summarised a meta-analysis covering 163 studies and more than 82,000 participants, looking at when customers will and will not accept dealing with AI.
A meta-analysis is not a single experiment. It pools results across many, which makes the direction of a finding more trustworthy than any one study and the precise figures less so. Treat 78.6 and 60.4 as a well-supported gap rather than as two exact quantities.
There are two resistances, not one
The useful structural finding is that customer reluctance splits in half, and the halves have different causes and different fixes.
The first is gatekeeper aversion: reluctance to enter an uncertain multi-stage process at all. This has nothing to do with AI. It is the feeling of starting something that might take four steps and might end in nothing, and people avoided it long before any of this was automated.
The second is algorithm aversion, and it is the one that costs money. It is an additional 10 to 20 percentage point drop that appears only when the same option is labelled AI rather than human. Same service, same outcome, different word on the button.
The scale of it is easier to feel through one figure from the simulations: participants chose the chatbot with no queue only 28 per cent of the time, even when the alternative involved waiting. They were offered a faster route and most of them declined it.
Why the direction flips
The explanation the research offers is about intent, and it is the part that makes the finding usable rather than merely surprising.
When a person tells you no, you suspect a decision. Somebody weighed your case and chose against you. Perhaps they were having a bad day, perhaps they did not read it properly, perhaps someone else would have said yes.
When a machine tells you no, it reads as a rule. Impersonal, consistent, presumably applied to everyone. There is nobody to have been unfair to you.
Reverse the outcome and the same logic works against you. Good news from a person feels like recognition, because somebody looked at your case and responded. The same news from a system is just your row coming out of a calculation. It lands flatter.
The two-factor test
The practical output is a test with two questions, and it is stricter than most automation decisions in practice turn out to be.
Automate only where AI has a perceived capability advantage AND personalisation is not needed. The example given is sales forecasting: a machine looks more capable and nobody wants the forecast tailored to their feelings.
Every other combination favours the human. Capability advantage without the personalisation question settled is not enough. Neither is a task people do not enjoy doing.
Notice what the test is measuring. Not whether AI can do the job. Whether the customer will accept that it did.
The awkward implication for staff
There is a second reading of this that is worth surfacing, because someone in the room will get there before you do.
If customers accept refusals more readily from a machine, the people currently delivering those refusals are absorbing a cost that the organisation has never named. They are taking the friction that the process generates, one conversation at a time, and it is measured here at roughly 18 percentage points of acceptance.
That is an argument for automating the message. It is not an argument for removing the person. The research says people accept the answer more readily from a system, not that they need nobody afterwards. The sensible shape is a clear automated decision with a named human route behind it for the cases that warrant one, which is close to the opposite of a chatbot that will not let you through.
The cost of the word
The label penalty is the finding I would put in front of anyone about to redesign a customer journey, because it separates two things that are usually confused.
Ten to twenty percentage points is the cost of the word AI, not the cost of the technology. The same process, described differently, performs differently. That means naming is not a marketing decision taken after the build. It is part of the design, and it has a measurable price attached.
It also means the honest version of the question is harder than it looks. You cannot dodge the penalty by not mentioning it, at least not for long, and organisations that have tried to be quiet about automation have generally found the disclosure happens anyway and costs more when it does.
What to do with this on Monday
Take the parts of your service where a customer receives an outcome, and sort them into two piles: the ones where the answer is usually favourable, and the ones where it is usually not.
Refusals, declines, rejected claims, ineligibility, waiting lists, and anything where the honest answer is no. That is where the evidence says a well-designed automated response will be accepted more readily than a person delivering the same message.
Approvals, awards, upgrades, resolutions in the customer’s favour. That is where a person is worth the cost.
Most organisations have it the other way round, because the difficult conversations are the ones staff least want to have, so those are the first offered to a machine for the wrong reason. The evidence happens to agree with that instinct. It just gets there by a completely different route, and the route matters when you are designing the thing.
Before you quote any of this
The MIT SMR piece is part paywalled and the figures above come from the visible summary. They belong to the meta-analysis rather than to MIT SMR’s own research, so attribute them that way.
If a specific number is going to carry weight in a business case, go to the underlying studies first. The 78.6 against 60.4 comparison is the one most likely to be repeated, which makes it the one most worth checking yourself.