The mainframe, the central machine where banks clear their transactions, is having an unusual year: nobody can quite agree whether 2026 is the year it gets switched off or the year it gets renewed. Gartner expects more than 70% of the mainframe exit projects started this year to fail to deliver their intended benefits, largely because teams overestimate what generative AI can do on its own. Meanwhile, 94% of the companies surveyed by BMC still see the mainframe as a long-term platform. In the UK, the Shakers market analysis counts 16 distinct open roles that name the machine in their title, out of 56,780 active tech roles, and one of them is not a developer role at all: a retail bank-grade academy that trains mainframe programmers from scratch. This article explains what a mainframe is, why banks and retailers keep running on them, what AI can genuinely do with one today, and what the UK market is actually hiring for.
A mainframe is a high-availability central computer built to process millions of transactions and records in batches, above all in banking, insurance and government. It is not a dying machine: 94% of companies surveyed by BMC see it as a long-term platform (2026), and AI can now recommend changes on it, with human approval. The outstanding debt is modernisation: Gartner expects more than 70% of mainframe exit projects started in 2026 to fail. In the UK, 16 open roles name the machine in their title out of 56,780 active tech roles (Shakers market analysis, October 2026), and one retailer is training a new generation through a dedicated mainframe academy.
What a mainframe is and why it still runs production
A mainframe is a large-scale computer designed for one thing: processing enormous volumes of transactions and records with continuous reliability. It runs its own operating systems, today above all z/OS on the IBM Z family, and it is organised around batch processing: closed time windows, usually overnight, in which millions of operations settle in a single pass. It is not simply a bigger server, and the difference is not measured in rack units or clock speed but in what the platform promises about the work it carries. A horizontal server fleet spreads work across many machines that can be switched off one at a time. A mainframe concentrates the critical workload on a single platform that does not go down, with redundant components and decades of software written on top of it.
The software that moves money runs on it: accounts, cards, payroll, policies, settlements. Much of it is written in COBOL, a 1959 language whose decimal arithmetic never rounds unless asked, and whose value lies in the years of business rules accumulated inside the code. That pairing, machine and language, is what a bank calls its core. The question of whether it disappears one day has been open for decades. What changed in 2026 is that artificial intelligence entered the conversation from both sides: as an argument for leaving the mainframe and as a tool for staying on it. The base architecture of the machine, meanwhile, has not changed. Everything around it has.
Why UK banks and retailers still run on mainframes in 2026
The short answer, the one every operator gives before any consultant deck is opened, is that the system delivers. BMC's Mainframe Stack Survey 2026, collected by Open Mainframe Project in its October analysis, gives a figure that contradicts the image of a condemned technology: 94% of respondents see the mainframe as a long-term platform or one that attracts new workloads, and 69% report growth in general-purpose capacity. Read the source with its declared bias, and the bias is real: BMC sells the management tools that run on these machines and has every commercial reason to find them healthy. Even so, the direction of the data matches what daily operators report.
There are three underlying reasons. The first is the cost of failure: a batch window that breaks stops payments and accounting for an entire bank, and the risk of a rewrite is measured in customer money. The second is regulatory compliance: security and compliance have ranked as the number one priority for seven consecutive years in the same survey, and mainframe environments carry decades of audited controls. The third is that business rules live in the code, not in a document: forty years of exceptions, fees and rounding decisions that nobody documented outside the programme.
The UK job market draws the same picture from the other side. In the census of roles whose title names the machine, a high-street retailer, Next, is running a Mainframe Academy: twelve months of initial training in Leicester, a paid entry route into a discipline most universities stopped teaching. Freemans Grattan Holdings, another retail group, hires a batch scheduling analyst for its mainframe. And the signal that best marks the turn of the decade comes from the survey vendor itself: BMC Software is hiring a Principal Product Manager for Mainframe Agentics in the UK, a product role for AI that operates on the mainframe. The names of the projects change every quarter. The machine underneath does not.
70% of the projects that abandon the mainframe will fail
This is the figure that frames the year. Gartner expects more than 70% of the mainframe exit projects started in 2026 not to deliver their intended benefits, and the reason it gives is unusually concrete: overestimating what generative AI can do on its own. The figure arrives through a chain worth declaring in full: Open Mainframe Project, the Linux Foundation community that hosts the mainframe ecosystem, publishes it and attributes it to reporting by CIO.com on the Deluxe Corporation case. It is not a number from IBM or from a migration vendor.
The typical failure is not technical. It is a failure of expectation. A team asks a model to rewrite a settlement system, receives plausible code, and later discovers that the code does not repeat the original behaviour: exceptions that only lived inside the programme are missing. Open Mainframe Project describes the market consequence with a 2026 example: in February, IBM shares fell close to 13% in one session, around 31 billion in market value, after the headline that Anthropic was modernising COBOL with AI. The mainframe customer's panic turned out to be premature, and correcting that expectation is exactly what Gartner's 70% measures.
Against that trap, the framework the industry works with distinguishes three paths, inherited from AWS documentation and picked up by Open Mainframe Project. The choice is not fashion. It is the workload.
| Path | What it means | It fits when |
|---|---|---|
| Retain | The core stays on the mainframe and integrates with cloud and AI around it | The workload is stable, the cost reasonable, and touching it risks more than it saves |
| Refactor | The system is translated to Java or another language, keeping its monolith structure | Behaviour must be preserved byte for byte, but the platform wants to modernise |
| Reimagine | Business rules are extracted and an equivalent cloud-native system is written | Documentation or domain talent exists, and the business accepts a redesign |
The decision is made workload by workload, never as one grand programme that assumes every system in the estate wants the same future at the same price. Payroll can be reimagined while card settlement is retained, and the two coexist for years. The progressive replacement pattern is the same one described in the guide to modernising a Java monolith to cloud native: every piece that moves lives alongside the piece that stays until it proves it does the same thing.
What AI can do with a mainframe today
The technical proposal that best summarises 2026 comes from Open Mainframe Project in one phrase: parse first, prompt later. In its nine-step pipeline, generative AI only generates at one step.
- Steps 1 to 4, deterministic: parse the code, build the dependency graph, extract the rules and map the target system.
- Step 5, decision: choose the path from the table above for that workload.
- Step 6, the only generative one: produce code with a language model.
- Steps 7 to 9, verification: static analysis, equivalence testing and human approval, which check that what was generated does the same as the original.
What that architecture produces is big numbers in discovery and verification, not in writing. The cases collected in the primary source point the same way, and deserve their label: they are vendor-reported figures, without independent verification.
- Morgan Stanley describes its DevGen.AI tool reviewing around nine million lines and saving close to 280,000 hours by 2025.
- The Toyota case with AWS documents around 40 million lines migrated from COBOL to Java at half the effort, after automating the documentation of 375,000 lines in five hours with 90% accuracy.
- AWS itself warns of what the case does not prove: a general-purpose model without that structured pipeline did not beat traditional methods.
The limit of generation without verification has a brief demonstration in the same primary source. In an IBM demo on mainframe code, the model answered that a date utility was called 13, 29 and 31 times. The real count was 30. One extra call. In a reading report, an error that size goes unnoticed, filed under rounding. Not here. In a migration decision it decides whether the migration happens at all, because the count is the evidence.
Where AI already genuinely runs is inside the machine. IBM's latest mainframe, the z17, carries the Telum II accelerator, announced at the Hot Chips conference in August 2026, which processes more than 450 billion inference operations per day at millisecond latency. The machine that holds the transactions is also the machine that runs the model analysing them. The data never leaves the box. In banking, that sentence is worth a compliance programme.
For migration itself, the public reference remains OpenAI's October evaluation: its models resolve 97.9% of COBOL migrations to modern languages with byte-for-byte parity in the outputs, when given high reasoning effort, and OpenAI itself requires expert review before production. The 97.9% does not sign off on its own. It is the cheapening of the first version that makes the middle path of the table viable, refactoring, and shifts the budget towards the people who verify.
The profile UK employers are hiring for
The Shakers market analysis, over 56,780 active tech roles in the UK with a cut-off of 28 October 2026, counts 16 distinct roles that name the machine or its language in the title, out of 53 candidates after removing duplicates. The number is small, and that is the information. There is no mass market here. None. There is a narrow one, where scarcity is paid in search time, which has a cost of its own that few employers measure. The skill layer of the analysis does not tag mainframe or COBOL in the UK, so the count comes from a title census with a declared method.
The sixteen roles sketch the market. One is Next's Mainframe Academy in Leicester, a training programme rather than a vacancy. Two are product management roles at BMC Software for what the vendor calls mainframe agentics. The rest are the classic shapes: a Cobol Lead Programmer at Harvey Nash, mainframe developers at Unisys, VIQU, Ascendion and Next, an application modernisation consultant at TXP working across Mainframe, COBOL, Java and .NET, a systems programmer at Siemens, an IBM z/OS integration engineer, a mainframe hardware infrastructure engineer, and a batch scheduling analyst at Freemans Grattan. Nine of the sixteen carry senior in title or level. One publishes a salary: a senior mainframe developer at 42,000 to 55,000 pounds. One posting is a junior route, and it belongs to a retailer, not a consultancy.
The hybrid profile sits at the confluence: someone who reads COBOL and the z environment, knows the financial or retail domain, and masters the target language and platform. That person executes the ninth step of the pipeline, the approval. If your organisation runs its core on a mainframe and needs capacity to decide its future without stopping the operation, tell us about the project: which workloads weigh the most, which path of the table you are considering, and who knows the business rules today. That scope is what turns into a profile or a team, and into the acceptance criterion that decides when something is genuinely migrated.
What to do if your core is on a mainframe
Four decisions order the conversation before anyone commits to a path:
- Inventory the business rules alongside the code. On a mainframe, the programme is the only reliable documentation of decades of exceptions. Before translating anything, extract those rules with the people who still know them and turn them into test cases.
- Set behavioural parity as the acceptance criterion. The OpenAI evaluation requires every output to match byte for byte. A real project should demand the same against its own historical data before calling anything migrated.
- Choose the path workload by workload. Retain, refactor or reimagine is not a company decision. It is a system decision. The coexistence pattern, a migrated piece living next to its sibling until equivalence is proven, is the same one described in modernising a Java monolith to cloud native.
- Decide who reviews and who approves. The modern pipeline leaves generation to AI and reserves three steps for verification. Without the hybrid profile that executes the approval, the 97.9% of the test never reaches production.
Frequently asked questions about mainframes
- What is a mainframe and what is it used for today?
-
A mainframe is a large-scale central computer, of very high availability, that processes millions of transactions and records in batches. Today it carries the core systems of banks, insurers, governments and large retailers: accounts, cards, payroll, policies and settlements, above all on operating systems like z/OS in the IBM Z family.
- What is the difference between a mainframe and a server or the cloud?
-
A server spreads work across many machines that can be switched off one at a time. A mainframe concentrates the critical workload on a single platform that does not go down, with redundant components, high-volume batch processing and decades of audited software on top. Modern cloud combines both models, which is why many projects mix mainframe and cloud rather than replacing one with the other.
- Why do banks still use mainframes in 2026?
-
Because the system delivers: a batch window that breaks stops payments and accounting for an entire bank. 94% of respondents to BMC's Mainframe Stack Survey 2026 see it as a long-term platform or one that attracts new workloads, compliance has ranked first for seven years, and decades of business rules live inside the code. In the UK, a retailer is training its own mainframe generation through a twelve-month academy.
- Is COBOL still used on mainframes?
-
Yes. COBOL remains the dominant language of the core systems that run on mainframes, because its exact decimal arithmetic and its batch processes fit money. OpenAI measured it in October 2026: COBOL was the hardest of 20 languages for its models, and its migration evaluation, with 97.9% of translations passing byte-for-byte parity at high reasoning effort, is the most-cited evidence that the language can now be moved safely under expert review.
- Can AI migrate a mainframe?
-
It can do much of the work and cannot sign off the result. In COBOL migrations with byte-for-byte parity, OpenAI's models reach 97.9% in controlled tests at high reasoning effort, but the evaluation itself requires expert review before production. Gartner expects more than 70% of mainframe exit projects started in 2026 to fail, above all through trusting generative AI without the analysis, equivalence testing and human approval work.
- How much does a mainframe developer earn in the UK?
-
Almost nobody publishes a salary. In the census of 16 UK roles with the machine in the title, counted by the Shakers market analysis in October 2026, only one publishes a range in its own field: 42,000 to 55,000 pounds a year for a senior mainframe developer. One data point, useful as orientation, not as a market average. The stronger signal is Next's twelve-month mainframe academy in Leicester: an employer investing in training rather than bidding for scarce seniors.
Sources and method
Three sources support every figure in this article. Each one carries its attribution in the paragraph where the number appears.
- Open Mainframe Project, analysis of AI mainframe modernisation (20 October 2026): source of the Gartner estimate (collected from CIO.com reporting), the BMC survey, and the Morgan Stanley and Toyota-with-AWS cases, which are vendor-reported figures without independent verification.
- OpenAI, evaluations of COBOL and migration (October 2026): the 97.9% figure with expert review; here the evaluation is developed in full.
- Shakers market analysis, cut-off 28 October 2026: 56,780 active tech roles in the UK; 16 distinct roles name the machine or its language in the title, out of 53 candidates after deduplication (fuzzy search of 88 results; one Next academy posting counted twice through syndication). No salary field beyond the single VIQU range.
