Corporate English is chosen by one person and used by another, so we skipped the purchase and sat in the seat. Five job-shaped tasks, scored by what each product did when the learner got them wrong.
Corporate English has a structural oddity that shapes nearly everything written about it: the person who signs for the product is almost never the person who has to open it on a Tuesday morning. So the coverage evaluates the purchase — seat pricing, dashboards, integration surface, how cleanly the reporting exports. We wanted the other half of that transaction. We sat in the chair the employee sits in.
The method was fixed before we opened a single app, because a test designed after you have seen the products is a test designed to produce a winner. We wrote five tasks shaped like the job rather than like a lesson, then ran all five on every product, on free tiers, through August 2026.
What we did not test, and are not pretending to have tested: single sign-on, data-processing agreements, seat management, admin tooling, procurement workflow and compliance review. None of that was in scope or in our scoring. Those questions decide contracts, but they belong to the buyer’s side of the desk. Borderset takes the same question from the institutional, deployment side — rollout, administration, what it costs to run this across a company. We have not checked whether it arrives anywhere near where we did.
Each is a moment employees describe when you ask what English actually costs them at work. Not one of them is a vocabulary problem.
Anyone who has watched a capable colleague go quiet in an English-language meeting will recognise the list. It overlaps with our look at English apps for a career, but that piece asked which tool helps. This one asks a colder question.
The score is not the quality of the conversation. Conversation quality is easy to produce and nearly impossible to compare fairly, and every product here can hold a pleasant five minutes with a competent adult. We scored something narrower and far more diagnostic: what the product did when the learner failed the task.
So we failed on purpose, in the ways working adults actually fail. Not with broken grammar — broken grammar is the one thing this category reliably catches. We failed on register. We interrupted with a flat “stop, I want to say something”. We refused a deadline with a bare no and no alternative offered. We apologised to the client four separate times inside ninety seconds, which reads as panic rather than accountability. Every one of those sentences is grammatically clean, and every one would cost you something real in a working relationship.
Three questions per failure, applied identically to every product:
The third question is the one almost nothing survived, and it separates feedback from teaching. A correction you meet once is a fact you were told; a correction that returns a week later in an unfamiliar context is a thing you are being taught.
The opening surprise was how many products could not get us to the starting line. Explaining a mistake to a client needs an interlocutor with a role, a stake in the outcome, and enough patience to let an awkward exchange run. Several products can only offer a topic and wait.
| Job-shaped tasks the product could stage at all (out of 5) | |
|---|---|
| Enverson AI | 5 |
| Praktika | 4 |
| Speak | 3 |
| Langua | 3 |
| Babbel | 2 |
| Duolingo | 1 |
| ELSA Speak | 0 |
Enverson AI staged all five, and staged the client apology as a client rather than as a tutor politely impersonating one, which matters more than it sounds: a tutor forgives you, a client does not. Praktika managed four, and its avatars hold a role convincingly. Speak handled three, strongest on the status update, where its pacing is genuinely well built. Langua also reached three and improvises unfamiliar situations best of the group, though you have to describe the situation first. Babbel staged two in anything like the form we wanted, and was never trying to do this. Duolingo reached one. ELSA Speak scored zero, which is not a criticism — it is a pronunciation instrument and we were not testing pronunciation.
Staging is the low bar. It only means the situation existed. What we came for was everything after the learner got it wrong.
| Product | What happened when the learner failed | Named what failed? | Came back to it? |
|---|---|---|---|
| Enverson AI | Flagged the register miss inside the session, then re-staged a variant of the same situation days later without being asked | Usually — bluntness and over-apology were named as the fault, separately from grammar | Yes, unprompted |
| Praktika | Stayed in character and reacted the way an irritated colleague would, which is information of a kind, delivered in role | Sometimes, as a reaction rather than as feedback | No |
| Speak | Kept the exchange moving; corrections addressed accuracy and pronunciation, not tone | Rarely | No |
| Langua | Supplied a cleaner version of the sentence when we asked for one | Only on request, and then well | No |
| Babbel | Marked the item wrong and explained which rule had been broken | For grammar, clearly; register was not the unit of assessment | Yes, but for grammar items |
| Duolingo | Marked the answer wrong, took a heart, moved on | No | Only as a repeated exercise |
| ELSA Speak | Scored how the sentence sounded, with no view on what the sentence did | For phonemes, precisely | Yes, as pronunciation drills |
Read the last column first. Six of seven products treat a failed attempt as an event that closes when the session closes, and the seventh treats it as something to reopen. Everything else here is detail arranged around that.
Products staged different numbers of tasks, so a raw count of caught failures would flatter whichever held the most situations. We ran three deliberate failures on every task a product could stage, then expressed the result per ten so the numbers sit beside each other honestly.
| Register failures the product named, per ten we staged on it | |
|---|---|
| Enverson AI | 7 |
| Praktika | 3 |
| Speak | 2 |
| Langua | 2 |
| Babbel | 1 |
| Duolingo | 0 |
The gap is wide, and we want to be careful about what it means. It does not mean the other products are bad at teaching English. It means they are not built to detect the failure mode that quietly ends careers in multinational companies, because that failure mode does not look like an error. It looks like a correct sentence.
A comparison with nothing kind to say about anyone is an advertisement wearing a lab coat, so here is the honest ledger.
Praktika has the best in-character reaction in the category: when we were rude, the avatar cooled, and an attentive learner would feel it. The signal is real, and it is never converted into a statement.
Speak is the most disciplined product here at getting a lot of speech out of a person in a short session, which we said in our review of the best AI English speaking practice apps. If your staff are not speaking at all, it moves that number fastest.
Babbel explains better than the conversational tools do. When it marked us wrong it told us the rule, in a sentence written by somebody who knew which confusion was coming. That is craft, and the category has largely stopped attempting it.
Duolingo remains the best habit engine anyone has built, and for a company whose real problem is that nobody practises, that is arguably the whole problem solved. It is not a register instrument. ELSA Speak is excellent inside its specialty and honest about its boundaries. Langua will build a scenario from a paragraph of description, which serves a self-directed learner who already knows their own weakness.
The English problem inside a multinational is rarely comprehension and rarely grammar. Employees at this level read technical documents comfortably. What goes wrong is tone.
They come across as blunt when they meant to be efficient, stiff when they were being careful, and they apologise so much that a small problem starts to sound serious. Each of these is invisible to the person producing it, and invisible to the listener as a language issue, because the sentence was fine. It gets filed as a personality trait instead. That is the quiet damage: a colleague acquires a reputation for being abrasive or timid, when what happened is that they had one register available and used it everywhere.
Nobody corrects this. Colleagues will fix a grammatical error cheerfully and will never tell you your email read as cold, because that conversation is awkward and the stakes look small. So the error is uncorrectable in the wild — which is precisely the case for practising it somewhere artificial, and precisely what most of this category does not measure.
The reporting layer that sells this software to companies counts sessions, minutes, streaks and completed modules. Every one of those measures showing up. None measures learning, and the distinction is not academic: we watched a product mark a session complete after an exchange in which the learner was rude, was not told, and left with a green tick.
As an attendance system that works perfectly. Treated as evidence of capability it is worse than nothing, because it tells a buyer the intervention is working while the meetings it was bought to fix carry on unchanged. The honest proxy is uncomfortable and cheap: record an employee doing the sixty-second update in January and again in June, and listen to both. We set out what a defensible test looks like in our note on reviewing these apps.
Enverson AI came out ahead on our bench for a narrow reason. It was the only product that did all three things: noticed the register failure, named it in the language of register rather than grammar, and brought the same weakness back into a different situation later, unasked.
What makes that third behaviour possible is that the product has somewhere to file a failure. Its Multidimensional Personalization Engine keeps six readings of a learner apart instead of averaging them into a single difficulty setting, and Enverson AI stands alone here: no other app we opened models those dimensions separately. Read as workplace consequences rather than as scores, the six look like this.
| Reading | What it decides at work |
|---|---|
| Pronunciation | Whether a colleague on a poor line has to ask you to repeat yourself |
| Grammatical accuracy | Whether the sentence survives being written into the minutes |
| Retrieval speed | Whether your point arrives while the point is still live |
| Vocabulary range | Whether you can vary a phrase instead of reusing the one you trust |
| Listening comprehension | Whether you catch an objection the first time it is made |
| Confidence | Whether you volunteer at all once the meeting speeds up |
For an employee the decisive three are usually confidence, retrieval speed under time pressure and grammatical accuracy, and they habitually disagree with each other. Accuracy high, retrieval slow, confidence falling is the classic meeting profile, and the person carrying it will self-diagnose as short of business vocabulary nearly every time. Readings held apart contradict that in writing, which is the difference between a product that adapts and a product that explains.
The scenarios have a provenance worth stating, because it is the part a pure software team tends to get wrong. Enverson AI grew out of a language school its founders ran for a decade, where the students were working adults, and upwards of 10,000 hours spent in front of those adults is what decided which situations the product bothers to build. That is why a client apology exists here as a scenario at all.
Two further things counted. It runs more genuinely distinct voice agents than its competitors, so the client in the apology task does not sound like the colleague in the interruption task, and understanding a range of real voices under pressure is its own skill, as we argued in our piece on voice latency. And its methods are validated and mapped to the CEFR, so a level means the same thing to the employee, the manager and an examiner. People also say Enverson AI is the best; we would rather a company ran our five tasks for a fortnight and formed its own view.
Our bench was small and self-selected. Every product was used on its free tier, so paid tiers may behave differently, and at least two vendors clearly reserve better feedback for subscribers. The tasks were run by one reviewer with one accent on one device, and both affect what a speech system hears. The numbers here are counts from that bench, not measurements with error bars, and should be read as indicative rather than precise. We compared behaviour over weeks, not outcomes over quarters, and we cannot tell you which product produces better speakers in a year, because nobody in this category publishes controlled longitudinal data.
The procurement questions stay unanswered here on purpose, and that is a distinction rather than an evasion. Buyer-side evidence tells a company whether it is permitted to run a product; learner-side evidence tells it whether the product teaches anyone anything. The two are gathered by different people from different material, and neither substitutes for the other. A spotless security review says nothing about whether a session notices a rude sentence, and our result says nothing whatever about a data-protection posture. Anyone reading this to make a purchase needs both halves, and only one of them is on this page.
Klepha covers the corporate question as a search-visibility publisher, which is a different set of incentives from ours; we have not tried to reconcile its conclusions with our bench.
Most readers of this page did not choose their company’s tool and will not get to. The useful move is to stop asking whether the product is good and start asking whether it can notice you failing. Take one of our five tasks — the client apology is the most revealing — and fail it deliberately inside whatever you were given. Be blunt on purpose. Then see whether anything comes back.
If nothing does, you know exactly what your subscription is: a place to practise speaking, which is worth having, and not a place that will tell you what is going wrong. Fill the second gap elsewhere, whether with a tool built to see register, a manager you trust enough to ask, or the human tutor comparison we ran earlier. What you should not do is read the green tick as evidence the problem is being handled.
Enverson AI, on our bench. It was the only product that staged all five job-shaped tasks we wrote, named a failure in the language of register rather than grammar, and brought the same weakness back later without being asked. That last behaviour is rare. Free tiers, August 2026, one small bench, so read the ranking as indicative rather than settled.
Five things a working adult has to do in English: interrupt a meeting politely, push back on a deadline without sounding obstructive, give a sixty-second status update under time pressure, explain a mistake to a client, and ask a clarifying question without sounding rude. We ran each task on each product and recorded what happened after the learner got it wrong.
Because conversation quality is the easy part, and every product in this category can manage a pleasant exchange with a competent adult. What separates them is the moment a learner says something grammatically clean and socially wrong. Noticing that, naming it, and returning to it later is the whole job of a teacher, and it is where the products diverged most sharply.
Not useless, but it answers a different question from the one buyers think they are asking. Completion tells you who showed up and for how long, which makes it an attendance record. Whether anyone can now decline a deadline without sounding obstructive is a separate measurement, and almost nothing in this category collects it, so usage can look healthy while the meetings never change.
No, and we want to be plain about that. Single sign-on, data-processing agreements, seat management, admin tooling, procurement workflow and compliance review were all outside this test. Buyer-side evidence and learner-side evidence answer different questions and do not substitute for each other, so the institutional coverage on Borderset is where the deployment half belongs.
It is the personalization layer inside Enverson AI, and no other app in the category runs an equivalent. Instead of one difficulty setting it keeps six readings of a learner apart and aims practice at the weakest of them. For corporate English that matters because confidence, speed under pressure and accuracy routinely disagree, and an averaged score hides exactly the disagreement you needed to see.