<!-- Canonical: https://www.aigentcy.com/blog/legacy-system-modernization-with-ai/ -->

[Process automation](https://www.aigentcy.com/blog/category/process-automation/)

# Legacy system modernisation with AI: replacing systems in stages

Legacy system modernisation for mid-size companies: when to keep, extend or replace a system, how to move across one part at a time with the old system still running, how to check the data, and how the costs compare with renewing.

Aigentcy 3 November 2026 11 min read

Key takeaways

1.  Replace a system when its cost, its workarounds or the people it depends on block the business. Extend it when it works and only lacks a connection.
2.  Move across one part at a time, starting with the part that hurts most, with the old system running until the new part has handled real work.
3.  Check migrated data with counts, totals and a field-by-field sample, and don't switch until every difference has a reason someone signed.
4.  Compare 5 years of renewal, workaround and parallel-running costs against build and running costs. Payback often takes years, so state it.
5.  Use AI to read and document the old system and to write tests. Independent results for automatic translation of whole systems are still weak.

**The old system runs until the last part has moved**

Leave goes first because it caused the most workarounds. Each move is checked on real work before the next one starts, and the old system goes read-only only after payroll has run once in the new one.

Legacy system modernisation means replacing a system the business has outgrown one part at a time, with the old system running until each new part has handled real work. Built that way, each step is small enough to check and to undo.

In our worked example, a 250-person company replacing a rented HR system spends $36,660 less over 5 years, with break-even early in year 4.

AI helps with reading and documenting the old system and with writing tests that pin its behaviour down. It doesn’t yet translate whole systems reliably. In an independent benchmark, the best method succeeded on 32.8% of repositories.

The case usually starts with a renewal. The HR system renews in 4 months. It charges per employee, the company has grown by a third, and the reports HR needs every month still come out of exports merged in a spreadsheet. Somebody suggests replacing it. Somebody else remembers the last replacement project, which ran a year late.

The systems in question are the ones a business of 50 to 1,000 people runs on: rented software, an old internal database, an on-premise ERP version near the end of its support, or a spreadsheet that became an application.

## Which systems count as legacy

Age matters less than fit. Most of the work comes from 4 kinds of system.

| Kind of system | How it shows up | What usually hurts |
| --- | --- | --- |
| Rented software priced per seat or per employee | The renewal grows with headcount, and half the modules go unused | Cost, and data you can only get out through exports |
| An old internal database or application | One person knows how it works | Key-person risk, and changes nobody dares to make |
| A spreadsheet that became an application | Macros, linked files, a version on someone’s desktop | Errors, no access control, no history |
| An on-premise system near end of support | Security updates are ending | Audit findings, and nothing new connects to it |

In Ensono’s survey of 500 senior IT and business decision-makers in the UK and US, [57% of UK organisations said they now extend legacy systems instead of replacing them](https://itbrief.co.uk/story/uk-firms-extend-legacy-systems-as-ai-spending-rises). The same research found that 77% had exceeded modernisation budgets and 67% had paused, scaled back or abandoned an initiative.

## Keep, extend or replace

A system that still does its core job rarely needs replacing. Some need a connection, and some need to go.

| Option | Choose it when | What it involves |
| --- | --- | --- |
| Keep | The system works, costs are stable and nothing new depends on it | Nothing, until the next review |
| Extend | The system works, but its data is trapped or a process around it is manual | An interface or a scheduled export, and automation built around the system |
| Replace one part | One module causes most of the pain, or most of the licence cost | A new part built alongside, with the old one switched off when it’s ready |
| Replace the whole system | Cost, workarounds and key-person risk all point the same way | Every part moved in stages, then the old system retired |

Deciding whether to keep, extend or replace a system

```mermaid
flowchart TD
accTitle: Deciding whether to keep, extend or replace a system
  A[A system the business has outgrown] --> B{Does it still do the core job well?}
  B -->|Yes| C{Is the data reachable?}
  C -->|Yes| D[Keep, and automate around it]
  C -->|No| E[Extend with an interface or export]
  B -->|No| F{Is the pain in one part?}
  F -->|Yes| G[Replace that part first]
  F -->|No| H[Replace in stages, part by part]
```

6 signals say replacement is worth an assessment:

-   **The renewal is growing faster than the use.** Seat or employee pricing climbs with headcount, while the modules people use stay the same.
-   **Work happens around the system.** Monthly exports merged in spreadsheets, re-keying between systems, reports built outside the system.
-   **One person holds it together.** If they left, nobody could change it safely.
-   **Support is ending.** The vendor stops security updates, and audit findings follow.
-   **Nothing new connects.** The system has no interface, so every new process starts with a manual export.
-   **Regulation asks for it.** EU financial entities other than microenterprises must, under the Digital Operational Resilience Act, [run a specific ICT risk assessment on legacy ICT systems at least yearly](https://eur-lex.europa.eu/eli/reg/2022/2554/oj) (Article 8(7)).

## Where AI helps, and where it doesn’t

AI changes 2 parts of modernisation: understanding the old system, and what the new system can do. It has changed less about the rewriting itself.

**Reading and documenting old code works.** Thoughtworks moved “using GenAI to understand legacy codebases” to the Adopt ring of its [Technology Radar in November 2025](https://www.thoughtworks.com/radar/techniques/using-genai-to-understand-legacy-codebases). A model can read an old codebase or database and produce a first draft of what each part does, which a person then checks against the running system.

**Writing tests from the old system works.** Tests that record what the old system does with real inputs pin its behaviour down. The replacement is then built to pass the same tests, which catches the business rules nobody wrote down.

**Translating a whole system automatically doesn’t work reliably yet.** On RepoTransBench, an independent benchmark of 1,897 repository-level translation tasks, [the best method succeeded on 32.8%](https://arxiv.org/html/2412.17744v2). Translated code also carries the old design across, workarounds included.

**Speed claims need measuring.** In [METR’s randomised study](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/), experienced developers working on their own codebases were 19% slower with AI tools, while believing they’d been 20% faster. Time the work before you plan around a speedup.

Within the new system, AI goes where the old one needed a person: reading incoming documents, drafting the monthly report, flagging lines that don’t match. That’s the “AI-native” part of a replacement. It’s built in from the start, because adding it to a rented system usually means another module and another fee.

## Replacing in stages

The usual name for replacing a system gradually is the strangler fig, from Martin Fowler. He describes building new parts alongside the old system, moving behaviour across over time, and retiring the old system when nothing needs it. In his words, it’s [a gradual process of modernisation](https://martinfowler.com/bliki/StranglerFigApplication.html).

For a mid-size company it comes down to 7 steps.

1.  **Map what the old system does.** Every module, who uses it, which reports come out of it and which other systems read from it. AI helps here, reading code, configuration and exports.
2.  **Decide what to keep and drop.** Old systems carry features nobody uses. Dropping them is often the largest single saving in the project.
3.  **Build the part that hurts most first.** Usually the one with the most workarounds or the highest cost, built around how the team works today.
4.  **Move its data and check it.** Counts, totals and a field-by-field sample.
5.  **Run both, then switch.** The team uses the new part for real work for at least one full cycle, with the old part still available.
6.  **Make the old part read-only.** People can still look things up, and nothing new goes in.
7.  **Repeat, then retire.** Once every part has moved, the old system is retired, and old links redirect to the new ones.

Starting with the part that hurts most gives the team a reason to use the new system, and it pays back sooner. Starting with the easiest part looks good on a plan and changes little for the people doing the work.

## Moving the data

In the replacements we’ve run, data migration is where problems surface first. The checks are simple enough to do every time: run them for every part, and repeat them on the day of the switch.

| Check | What it compares | What a failure means |
| --- | --- | --- |
| Record counts | Rows per table, old against new | Records lost or duplicated in the move |
| Totals | Sums of money, days, quantities | Values changed, rounded or mapped wrong |
| Field sample | 100 to 200 records compared field by field | Formats, dates or codes translated wrong |
| Business checks | Reports run in both systems for the same period | A rule behaves differently in the new system |
| Links and history | Attachments, comments and audit history | The record moved and its context didn’t |

Don’t switch until every difference has a reason someone signed. “3 leave requests missing” is fine if they were test records, and a payroll error if they weren’t.

**Counts, totals and a sample each catch different errors**

Matching totals can hide a lost record, and matching counts can hide a changed value. The field sample catches the dates, codes and formats that both of the other checks miss.

When we moved a client service desk to a new platform, [2,860 past requests from 74 client accounts](https://www.aigentcy.com/case-studies/service-desk-automation/) were moved with their history and checked against the old systems before an old desk was retired. The sync between the new desk and the developers’ tracker passed a 7-day test with 99.5% of updates arriving within 5 minutes, and the old portal addresses now redirect.

## Running old and new together

A parallel run costs money. You pay for 2 systems, sometimes 2 licences, and people enter some things twice. So set its length before it starts.

-   **1 full cycle per part.** A month-end, a payroll run, a quarterly report. The cycle that exercises the part’s rules.
-   **A named way back.** If the new part fails in the first week, the team knows how to return to the old one, and who decides.
-   **Read-only before retirement.** The old part stays readable for an agreed period, so questions about past records have an answer.
-   **Redirects.** Old bookmarks and links point to the new system, so nobody keeps working in the old one by habit.

## Comparing the costs

A company of 250 people rents an HR system at $9 per employee per month, plus $6,000 a year for 2 add-on modules. HR spends 12 hours a month merging exports for reports, at a loaded $45 an hour.

| Keep renting, per year | Calculation | Amount |
| --- | --- | --- |
| Licence | 250 employees at $9 for 12 months | $27,000 |
| Add-on modules | Annual fee | $6,000 |
| Workarounds | 12 hours a month at $45 for 12 months | $6,480 |
| Total per year |  | $39,480 |

The replacement is built in stages over 6 months for $60,000. The rented system runs alongside for those 6 months, and the new one costs $1,500 a month to host and maintain once it’s live.

| Replace in stages | Year 1 | Each year after |
| --- | --- | --- |
| Build | $60,000 |  |
| Rented system during the move, 6 months | $16,500 |  |
| Hosting and maintenance | $9,000 | $18,000 |
| Workarounds, until the switch | $3,240 |  |
| Total | $88,740 | $18,000 |

**More expensive for 3 years, cheaper from year 4**

The replacement carries the build and 6 months of paying for both systems up front. After the switch it costs $18,000 a year to run against $39,480 for the rented tool and its workarounds, and headcount growth widens the gap under per-employee pricing.

Over 5 years, renting costs $197,400 and replacing costs $160,740, a difference of $36,660. The replacement costs more until early in year 4. That’s slower payback than a typical automation project, and the business case should say so.

3 things move the result. Headcount growth raises the licence under per-employee pricing and barely changes the replacement’s running cost. Price rises at renewal do the same. And an underestimated build or a long parallel run pushes payback later. Run the comparison with your own renewal quote, your growth plan and a build estimate that comes from a mapped system.

The comparison also leaves out what the new system does that the old one didn’t. Our own [Squadra HR](https://www.aigentcy.com/products/#squadra) started as the HR system for one company in place of a rented tool. It now runs HR for 270 people, looked after by 3 HR staff, with EU pay transparency tools built in.

## Risks

| Risk | How it shows up | How to reduce it |
| --- | --- | --- |
| Business rules nobody wrote down | The new system does something slightly different, found at month-end | Tests recorded from the old system, and a parallel run over a full cycle |
| Data quality | Empty fields, duplicates, 2 systems that disagree | Sample checks in the assessment, and cleaning before the move |
| Scope growth | “While we’re at it” features delay the first part | A fixed scope per part, with new ideas listed for later |
| The key person leaves | Knowledge of the old system goes with them | Document the old system first, with AI help and their review |
| Parallel running drags on | 2 systems paid for, people entering things twice | 1 cycle per part, and a switch date agreed in advance |
| No owner after launch | The replacement stops changing and becomes the next legacy system | A named maintainer and a monthly agreement, or your own team with the code |
| Regulated data | Personal or financial data moved without the right records | Data maps, access logs and the assessments your regulator expects |

## Where to start

1.  List the systems with a renewal in the next 12 months, and the workarounds around each.
2.  For the one that hurts most, decide whether to keep, extend or replace it.
3.  If it’s replace, map what it does and price 5 years of keeping it, including workarounds.
4.  Get a build estimate for the first part and the whole, and compare them with the 5-year cost. An [AI readiness assessment](https://www.aigentcy.com/ai-readiness-assessment/) gives you the map, the starting numbers and a fixed-price estimate for one area in 2 weeks.
5.  Replace one part, check the data, run one cycle, and decide on the next part with real numbers.

## How we can help

We build the replacement around how your team works today, move it across one part at a time and check the data before anyone switches. You own the code, and the licence line for the old system goes when the last part moves.

**Next step:** [See legacy system modernisation](https://www.aigentcy.com/services/legacy-system-modernization/), or read how [2,860 past requests moved into one service desk](https://www.aigentcy.com/case-studies/service-desk-automation/) before the old desk was retired.

## Questions about legacy system modernisation

What is legacy system modernisation?

Replacing or upgrading a system the business has outgrown, such as rented software priced per seat, an old database only one person understands, or a spreadsheet that became an application. For mid-size companies it usually means moving to a system the business owns, one part at a time, with the data moved and checked.

What is the strangler fig approach?

A way of replacing a system gradually. New parts are built alongside the old system, work moves to them one part at a time, and the old system is retired when nothing depends on it. It avoids a single switch-over date on which everything has to work.

Can AI rewrite our old system automatically?

Not reliably yet. On an independent benchmark of repository-level code translation, the best method succeeded on 32.8% of repositories. AI is useful for reading old code, documenting what it does and writing tests that pin its behaviour down before the replacement is built.

How long do we run the old and new systems together?

Until each moved part has handled real work for at least one full cycle, such as a month-end or a payroll run, and the data checks pass. Then the old part goes read-only, and it's retired once nobody has needed it for an agreed period.

Who maintains the new system?

Decide before the build. Either your own team takes it on, with the code and documentation, or the team that built it runs it under a monthly agreement. A replacement with no named maintainer becomes the next legacy system.

## Related guides and case studies

-   [Process automation 12 min read
    
    ### AI workflow automation: a primer for operations leaders
    
    AI workflow automation moves work between the systems you already run, uses AI for the steps that need reading and sorting, and leaves approvals and exceptions with people.Read the guide](https://www.aigentcy.com/blog/ai-workflow-automation/)
-   [Process automation 11 min read
    
    ### AI readiness assessment: what it covers and how to run one
    
    A useful AI readiness assessment studies one area of the business, maps the work step by step, checks the data, and ends with numbers and a build estimate a finance director can approve or reject.Read the guide](https://www.aigentcy.com/blog/ai-readiness-assessment-guide/)
-   [Process automation 12 min read
    
    ### AI ROI: how to measure the return on AI automation
    
    AI ROI is a before and after on one workflow, measured the same way both times and priced the way finance prices everything else.Read the guide](https://www.aigentcy.com/blog/ai-roi-measurement/)

Case studies

-   [Case study
    
    ### Staff stopped copying client emails into Jira
    
    A gaming group's staff stopped copying client emails into the tracker. Each request sits in one desk with its history, live for staff in 2 weeks, clients in 5.Read the case study](https://www.aigentcy.com/case-studies/service-desk-automation/)

AI workflow automation

## Take routine admin off your team

We build the process into the systems your team already uses and keep people on the approvals and exceptions, so the same team handles more volume. An engineer runs every scoping call.

[See AI workflow automation](https://www.aigentcy.com/services/process-automation/) [Book a 30-minute call](https://www.aigentcy.com/contact/)
