We read every one of the 774 accounts customers wrote in public about this bank over two and a half years, without a single internal record, and worked out what actually causes someone to leave angry. It is not the failure they mention most, and it is not the one anyone here was watching.
On Friday 13 March 2026, ten reviews of the Eastern Bank app appeared on Google Play in a single day, three weeks after HarborOne’s customers were moved onto Eastern’s systems. All ten were five stars. The longest ran to nine words.
Two years earlier the same app took 269 one-star reviews in five months, and every number this bank watches stayed calm. This brief is about the distance between those two facts.
Dimension Labs works out what causes what. Companies collect enormous amounts of customer writing: surveys, call notes, chat logs, complaints, reviews. Almost all of it is counted. Almost none of it is understood. We read it, make it measurable, then separate the factors driving an outcome from the ones that merely occur alongside it.
Nobody at Eastern Bank commissioned this. We were given no internal record of any kind. No survey, no call transcripts, no account data, nothing that is not already sitting on the public internet.
So we took what was there: every review, every complaint and every rating left about this bank in public between January 2024 and July 2026, across eight sources. 774 published accounts. We read all of them, then measured what separates a customer who left furious from one who did not.
That is a very small amount to work from. The last section of this brief is about what the same method does when it is pointed at the records this bank already holds.
Most companies we look at have tens of thousands of public reviews. This bank has 774 in two and a half years, and they are not evenly spread.
We had hoped to compare branches. Twenty-five of them carry 149 reviews between them, the middle branch has five, and only two reach ten. At those numbers one annoyed customer moves a branch average by half a star. There is no branch ranking in this brief because there is none to be had.
The complaint files at the CFPB and the Better Business Bureau hold 57 accounts between them. They matter later, but nobody can leave a good review on a complaint form, so they can never be averaged against anything.
That leaves one source with real volume: the mobile app, 540 reviews, seventy per cent of everything. This became a study of the digital experience by necessity.
We did not expect much from it. What it held was the largest single customer event in this bank’s recent public record, in plain sight, apparently unremarked on by anyone.
This is the entire input to this brief. Every public word about a bank serving hundreds of thousands of households.
Read the boxes left to right. The rule underneath applies to every number in this brief.
Between April and August 2024, 403 of the 774 accounts ever written about Eastern Bank in public appeared. Five months out of thirty-one. Narrow it to the two app stores and it sharpens: 364 of 540 reviews, two of every three, in that one stretch.
May alone brought 153. June brought 132. No other month in two and a half years reached 60.
That matters before anything else, because it means there is no single number for how customers feel about this bank. Average the whole period and you get 2.3 stars. Inside those five months, 1.9. Outside them, 3.0. The first describes nobody, and it is the one that reaches a dashboard.
Stacked by where people wrote it. The tinted band is the five months of the switch. The dashed line is the running share of the whole record.
None of this is a survey. It is what people chose to write without being asked, which makes it unrepresentative in every direction at once. It is also the only outside account of this period that exists.
What people wrote in those five months was not vague unhappiness. It was one mechanical failure repeated hundreds of times: they could not get into their accounts.
Read against a fixed set of questions, 47 in every 100 people writing inside those months described being locked out. Outside them, 10 in 100. Asked a second way, by what each person named as the worst failure they hit, being unable to sign in is the answer for 38 in 100 inside against 6 outside. Two readings of the same words, both landing in the same place.
“There was nothing wrong with the old app, this is frustrating because now I can’t get into my account or even try to retrieve my credentials.”
The whole event in one sentence. Not a preference and not a complaint about design. A person who had money in this bank and could not reach it, written in May 2024, in public, for anyone to read.
The reason is public, and Eastern published it. The bank replaced its entire online and mobile banking service, moving customers across in waves from March to October 2024. Everyone had to delete the app they had and install a new one, listed fresh on 26 March. Everyone had to accept the terms again and set up fingerprint or face sign-in from scratch. Anyone with a linked outside account or a scheduled payment had to rebuild it by hand, because none of it came across.
“When I tried the new app with my credentials it simply did not work at all.”
“The new app is awful. You cannot use the app for multiple people.”
What that establishes, precisely: the words prove that people who could not get in wrote much worse reviews, and that this clustered hard in those five months. The link back to the switch rests on Eastern's own published pages. That is a documented explanation, not a measured one, and this brief keeps the two apart.
Across those same five months, twelve people took a complaint to the CFPB or the Better Business Bureau. Three hundred and sixty-four wrote in an app store. Thirty to one, between the channel that carries legal weight and the one that carries none.
The fair version is not that the regulator saw nothing. Complaints did rise, from about 1.7 a month to 2.4. That is the point. A director looking at complaint volume in the summer of 2024 would have seen a number well inside its normal range, and would have been right to read it that way.
Column width is how many people wrote there. Column height is how they divided. Two of these places have no way to leave a good review, which is why they are never averaged in with the rest.
Nothing else caught it either. No announcement named the new service or who built it. No newspaper or trade publication covered the reaction, in Boston or anywhere else. That is a search result rather than proof of absence, but the search was not casual.
The one public description came on the October 2024 earnings call: the bank had “completed a major upgrade of our online and mobile banking platform,” and client retention had been “terrific.” Both are probably true. Together they describe a different five months from the one several hundred people were describing in public at the same time.
Today the app shows 4.1 stars on Google Play. During the switch, the people who wrote something averaged 1.9 on that same store. Both are real. Neither describes a customer.
On Friday 13 March 2026, three weeks after the HarborOne changeover finished, ten reviews appeared on Google Play in a single day. All ten were five stars, all on the same version of the app. The longest ran to nine words.
Placed by the day it was written and the rating it carried. Bigger dots mean the person wrote more.
It is the only day like it anywhere in two and a half years. For comparison, the busiest single day in the whole record is 5 June 2024: thirty people, averaging 1.8 stars, writing twenty-nine words each.
The imbalance is not confined to that Friday. Across every rated source, an angry customer writes 34 words and a pleased one writes 10. One-star reviews run to 20,849 words in all. Five-star reviews run to 3,562. A star average treats those as the same weight of evidence.
Every rated review. Each shape is how long people write at that rating. The tick is the middle person.
That is the mechanical reason a published rating cannot tell you how customers are doing. It is a vote, and a vote carries no explanation. The explanation sits in the writing, and the two groups produce wildly different amounts of it.
Rank the failures by how often people mention them, then by how much each one actually drags a rating down, and you get two different lists. That gap has the most direct consequence for what gets fixed.
Mentioning the change is by a wide margin the most common remark in the record: 182 people, more than twice anything else. Compared like with like, it accounts for about three extra bad reviews in every hundred, close enough to nothing that this brief will not claim it. Left uncompared it looks like twenty-three, and nearly all of that is the five months it happened in.
Being locked out was mentioned by 75 people, a third as many. It accounts for about twenty extra bad reviews in every hundred, and it is the only factor in the set that survives every test applied to it.
The dot is what that single factor does on its own, once everything else is accounted for. The bar around it is the range the answer could plausibly sit in. The most talked-about item sits near the bottom.
Three tests were run on each of these. Shuffle the cause at random and see if the effect disappears. Add a fake influence and see if it holds. Re-run it on random slices of the data and see if it moves. A finding that failed any of them is not called proven here.
The switch was the occasion. Being locked out was the cause. Nearly everybody mentioned the change in those months, so mentioning it tells you nothing about who is angry. Being locked out predicts an angry customer in every month of the record, at nearly the same strength.
A plan built from the first list would have gone to communications and change management. The evidence points somewhere much narrower and much more fixable: whether people can sign in.
Each place compared only against itself, so a change in who was writing cannot fake the result.
“The funds have never been returned to my company 's bank and I still do not have fund in my bank account at Eastern.”
This one was filed with the CFPB in May 2024, inside the five months. It is what a locked account turns into when it lasts: not irritation, but money that will not move. Twelve people got this far. Three hundred and sixty-four stopped at the app store.
Seven of the eight factors tested could not be separated from chance. Failed payments, fees and holds, crashes, not being able to find basic functions: all are in the record and all were real to whoever wrote them, but between ten and twenty-three people described each. They appear here in customers' own words and never as a measured effect. The record supports one proven cause, and claiming more would be inventing precision.
The 2024 switch is finished and what it cost cannot be recovered. It belongs in front of a board in 2026 because nothing about the conditions has changed and the pace has not slowed.
Cambridge Trust moved onto Eastern's systems over the weekend of 12 July 2024. The small-business service moved on 28 October 2024, onto yet another separate app, without carrying passwords across. HarborOne closed on 1 November 2025 and its systems changed over the weekend of 20 to 23 February 2026. On the bank's stated strategy, there will be more.
There is a price on the table for how this can go. In October 2024 the Consumer Financial Protection Bureau fined VyStar Credit Union 1.5 million dollars over a move to a new online and mobile banking service, finding that the way the change was planned and carried out was itself unfair to customers, and calling the harm foreseeable and preventable. The order also required refunds and a standing committee to oversee customer-facing systems.
Eastern's situation is not VyStar's, and no regulator has raised the 2024 switch. Two cautions belong with that. The Bureau's approach to enforcement changed after January 2025, and this brief has not checked where it stands now. And finding no enforcement action is not the same as confirming there is none.
“The impact of this negligence has a profound impact on credit scores and financing.”
A supervisor reading this record would start here, not with the rating and not with how many people were annoyed, but with the handful who can name what it cost them. That is the same place the VyStar order started.
What the record does show is customers describing real money at stake, in public, at the time. Late charges. Payments sent back. Money they could not move. Those are the facts a supervisor looks for first, and they sat in an app store the whole time, free to read.
Everything here was reconstructed from public fragments. No survey, no call recordings, no chat logs, no account records, no contact with anyone at the bank. Seven hundred and seventy-four accounts left by strangers, seventy per cent of them about a single mobile app.
From that alone: the start and end of a system change, the specific failure that drove it, proof that the failure was signing in rather than the complaint customers voiced most often, evidence that it cost some of them money, confirmation that none of the bank’s own indicators moved, and one afternoon of unexplained recovery in the published rating.
That is close to the minimum this method can be given. The final pages set out what changes when the input is not a minimum.
Every line below is copied exactly from a real review or complaint, with where and when it appeared. Nothing is tidied or shortened mid-sentence.
“After the collab with EasternBank and HarborOne, i’ve been unable to log into my account and I can’t even check how much money I have left in my account unless I transfer all of it to Venmo.”
“Also being locked out of my mobile banking. Service associate Vivian Lam was who helped me and I cannot express how unimpressed I was.”
“I am unable to use the new app on my iPhone and iPad - currently working with bank reps to straighten out this glitch.”
“Ever since they switched to this new app, my mobile deposit hasn't worked.”
“Trash new app is horrible the old one you could see all your loan info on the first page now you have to click something else to see basic information that was clearly displayed before it's confusing for no reason notification panel is the most annoying part who ever did this new app has never had a loan and has never used the app even the payment transfer you can't select you have to go to that specific one to make a payment absolutely trash making something so simple overly complicated ehhh”
“For a couple of weeks now, Zelle won't work on this app.”
“Special shoutout to Anuschka (manager of the branch) who is amazing!”
“Solange is one of the bank representatives she is always so pleasant kind and patient!”
Twelve people found the complaint form.
Three hundred and sixty-four found the app store.
One of those places was being watched.
Several hundred furious accounts in five months, and a rating that repaired itself partly on the strength of one Friday, are the same finding twice. In both directions the number this bank can see moved for reasons that had little to do with what customers went through. That is a measurement problem, and measurement problems can be fixed.
Daily count and text of new reviews, split by store and by app version, with an alarm set on how many are arriving rather than on the star average. Run against 2024, that alarm fires in the third week of April, six months before the bank described the period on an earnings call.
It is the only cause this record can prove, and the one customers named most precisely. Track how many people successfully sign in, how many security codes actually arrive, and how many finish setting up fingerprint or face sign-in. Measure it per wave, and give someone the authority to stop the next wave.
Linked accounts and scheduled payments did not come across in 2024, and the customers describing real money lost are overwhelmingly the ones whose payments failed as a result. Either move them, or run a named campaign to rebuild them with each affected customer.
It weighs a four-word compliment the same as a hundred-word account of a bounced payment, and it moves when someone is asked at the right moment. If the bank prompts for ratings in the app, whoever reads that number should be told when the prompting runs.
Twelve against three hundred and sixty-four is not a quiet quarter. It is a channel almost nobody reaches. The value in those twelve is entirely in what they describe.
Counting what customers mention tells you what is on their minds. It does not tell you what is costing you anything, because the loudest subject and the decisive one are rarely the same. This brief separates the two.
Three steps. Read every account against a fixed set of questions, so unstructured writing becomes countable. Compare like with like, holding constant everything else that could explain the result, so a factor that merely occurs alongside a bad outcome is not mistaken for one that produces it. Then try to break every finding: shuffle the supposed cause at random and see whether the effect disappears, add a fabricated influence and see whether the result holds, re-run on random subsets and see whether it moves. A finding that fails any of these is not called proven.
The value is the third step. It is what allows a board to be told that one factor is worth acting on, and that seven others, all real experiences for the customers who described them, cannot yet be separated from chance.
Every source below already exists inside Eastern Bank and is already being paid for. None requires new collection. Each removes a specific limitation of the analysis presented here.
| Source already held | What it makes possible |
|---|---|
| Survey free text | You already ask customers what they think and store the comments. Those comments carry the same signal as a review, from a population that is representative rather than self-selected, and each is attached to a customer you can identify. |
| Contact centre notes and recordings | Customers telephone before they post, usually by several weeks. The failure described in this brief would have appeared in call notes in April 2024. It is also the only source that captures the customers who never write anything publicly, which is nearly all of them. |
| Chat and secure-message transcripts | The highest-volume written record most banks hold, and the one where a customer states the problem in their own words while still inside the relationship. |
| Complaint case files | Already read individually, almost never read collectively. Fifty-seven public complaints were enough to identify a pattern here. A full internal file would establish which failures escalate and which resolve quietly. |
| Sign-in and session logs | Not writing at all, but the direct measure of the failure this brief identifies. Paired with the written record, it converts an inference into a count. |
| Account and balance history | The outcome that matters: whether the customer stayed, what happened to the balance, and whether a second product was ever opened. |
This analysis had to use a star rating as its outcome, because that is the only expression of customer feeling that exists in public. It is a blunt instrument: a self-selected vote from whoever was angry enough to open an app store.
With the bank’s own records the outcome becomes whether the customer stayed, what happened to their balance, and whether they opened a second product. The method is identical. Only the target changes, from a proxy to the measure the board actually governs.
Concretely: this brief establishes that a sign-in failure causes roughly twenty additional poor reviews for every hundred customers who wrote. The same analysis run internally would establish how many customers a sign-in failure cost, in which segments, and how much balance left with them.
Each of the 774 accounts was read alone, with nothing else visible: not where it was published, not the star rating, not the date, not the branch, not who wrote it. That is what makes it legitimate to compare across places and across months afterwards.
| Reading | What it captures |
|---|---|
| What went wrong | |
| primary friction | The single most consequential failure described, chosen by a fixed order of severity |
| issue severity | Whether the outcome was severe, moderate or minor, judged on what happened rather than on tone |
| actionability tier | Which team inside the bank owns the fix |
| resolution signal | Whether the problem was resolved, left unresolved, or made worse |
| Getting into the account | |
| access failure signal | Whether the customer could get into the account |
| access failure type | How the sign-in failed: credentials, security code, biometrics, lockout, enrolment or session |
| access failure verbatim | The customer’s own words describing it, copied exactly |
| Moving money | |
| money movement signal | Whether a payment, transfer or deposit is discussed, and how it went |
| money movement type | Which mechanism: external transfer, scheduled payment, bill pay, person to person, deposit, card |
| money movement verbatim | The customer’s own words, copied exactly |
| The system change | |
| migration signal | Whether the customer refers to a change of system rather than the system as it stands |
| migration loss type | What they say was lost: saved payees, history, alerts, a feature, ease of use, or extra steps imposed |
| migration verbatim | The customer’s own words, copied exactly |
| How the service performed | |
| app reliability signal | Whether the service functioned: crashes, errors, blank screens |
| usability signal | Whether it could be navigated, separate from whether it worked |
| channel of experience | Where the experience took place: application, website, branch, telephone or cash machine |
| Cost to the customer | |
| fees and charges signal | Whether a fee, hold or incorrect balance is described |
| financial consequence | Whether the customer states a cost they bore: a charge, a returned payment, credit harm, or funds out of reach |
| amount named | Any sum the customer names, normalised |
| People and place | |
| staff interaction signal | How employees are characterised: helpful, unhelpful, discourteous or unreachable |
| staff member named | An employee named by the customer |
| competitor named | Any other institution named |
| The customer relationship | |
| relationship signal | Whether the customer states they are leaving, considering it, warning others, or staying |
| tenure signal | Length of relationship, or arrival through an acquisition |
| overall sentiment | Overall stance, read from the words alone |
| Controls on the reading itself | |
| brand relevance | Whether the account concerns Eastern Bank at all, so unrelated material is excluded |
| record substance | Whether the account describes a situation or is an undifferentiated reaction |
| primary pain point phrase | A short neutral summary of the main problem |
| sentiment verbatim | The single line that best represents the account, copied exactly |
Readings marked as the customer’s own words are held to an exact-copy rule: a contiguous span reproduced character for character, or nothing at all. Every quotation in this brief was checked against its source before printing.
The record. 774 public accounts published between 4 January 2024 and 25 July 2026: Google Play (307), the App Store (233), Google Maps reviews of 25 branches (149), complaints filed with the Consumer Financial Protection Bureau (31), Better Business Bureau complaints (26), Trustpilot (12), WalletHub (9) and BBB reviews (7).
Public facts. The migration schedule, the new application listing and the list of what did and did not carry across are taken from Eastern Bank’s own published migration pages. Acquisition and conversion dates for Cambridge Trust and HarborOne come from Eastern Bankshares announcements and quarterly results. The October 2024 quotations are from the third-quarter 2024 earnings call. Complaint counts are from the CFPB’s public database. The VyStar order is CFPB docket 2024-CFPB-0013, dated 31 October 2024. Store ratings were taken on 4 August 2026. Eastern has never publicly named the provider of the new service, and neither does this brief.
What this cannot establish. These are public accounts, not a survey, and each source is self-selected in a different direction. Nothing here can say how many customers were affected, what the episode cost, or whether a single account closed. Of eight candidate causes, one is proven and seven are reported as customer experience only. Branch figures rest on between one and twenty-four opinions each, which is why no branch is ranked. The link from the system change to the authentication failures rests on Eastern’s published pages rather than on measurement, and is presented as such.