---
title: "Can AI replicate this software? The replicability test"
description: "Can AI replicate this software? How the replicability test works in diligence, what it shows and what it misses, and how to turn it into an IC conclusion."
url: https://anywayventures.com/insights/could-ai-rebuild-this-product
updated: 2026-10-25
---

[← ALL WRITING](https://anywayventures.com/insights) DILIGENCE

# Could a small team with AI coding tools rebuild this product?

Some of the largest investors now try to rebuild a target's product with AI before they buy it. The test is useful, as long as you know which half of the answer it gives you.

WRITTENOctober 2026
READING · Eight minutes
FOR · PE deal teams, operating partners, investment committees, software CEOs preparing for sale
RELATES TO[AI disruption risk review](https://anywayventures.com/ai-disruption-risk)

Yes, a small team with AI coding tools can now rebuild the visible features of many software products in days or weeks, and some investors test this directly before they buy. The replicability test asks how long a working copy of the core product takes to build, then asks the more important question: if the features are cheap to copy, what still stops a customer from leaving? For most [software companies](https://anywayventures.com/sectors/software-and-saas) the answer lies in data, integrations, regulation, distribution, switching costs and the services wrapped around the code.

That second question is where deals are won or lost. The first has become easy to answer. The rebuild measures the cost of the features. The deal depends on the cost of the customers.

> The rebuild measures the cost of the features. The deal depends on the cost of the customers.

## What is the replicability test?

The replicability test is a timeboxed attempt to rebuild the core of a target's product with AI coding tools, working from public material only: the website, product documentation, help centre, demo videos and a trial account where one exists. One or two engineers, a fixed budget of days, and a written record of what was easy, what was hard and what couldn't be built at all.

It is a disruption test, not a code review. The target's own code quality is a separate question for [technical due diligence](https://anywayventures.com/insights/technology-due-diligence-checklist) . The replicability test asks what a well-funded competitor, or a customer's own IT team, could put together if they decided to try.

## Why deal teams have started rebuilding targets

Two reports this year moved this from a curiosity to a practice.

In June 2026 the Financial Times reported, in coverage [summarised by Private Equity Wire](https://www.privateequitywire.co.uk/bain-uses-ai-vibecoding-to-build-software-replicas-in-pe-due-diligence-shift/) , that Bain & Company uses AI "vibecoding" to build working replicas of acquisition targets' software. According to that report, the replicas are used to test how easily a product could be copied and whether its advantage lies in its code, its data, its workflow design or its position in an ecosystem. Bain is reported to have started the work in 2023 as a specialist engineering capability before extending it across its private equity advisory teams.

The same month, [Bloomberg reported](https://www.bloomberg.com/news/articles/2026-06-11/apollo-is-screening-all-software-investments-for-ai-threat-risk) that Apollo screens every new software investment for AI disruption risk. Rob Bittencourt, Apollo's head of thematic investing, said on a Bloomberg Intelligence podcast that the firm had built a framework dividing software into 12 to 14 categories ranked by vulnerability to AI, and had applied it to existing holdings to decide where it might want to move on. He named data visualisation as an example of high exposure and healthcare as more protected because regulation slows adoption. The same report said Ares had hired an outside consultant to examine software investments in its largest listed private credit fund, and that Blackstone and Blue Owl had run internal reviews.

Behind both sits the software sell-off of early 2026, when listed software shares fell sharply on the fear that AI agents would undercut seat-based products. [Fortune](https://fortune.com/2026/02/06/anthropic-claude-opus-4-6-stock-selloff-new-upgrade/) tied part of the move to new releases from Anthropic, and [J.P. Morgan Asset Management](https://am.jpmorgan.com/us/en/asset-management/institutional/insights/market-insights/market-updates/on-the-minds-of-investors/what-does-the-software-sell-off-mean-for-private-markets/) set out the knock-on effects for private equity and private credit, where software is one of the largest sector exposures.

For a mid-market fund, the implication is simple. If the large buyers and lenders are asking this question, your exit buyer will ask it too. Better to have the answer at entry.

## How to run a replicability test

Most of the value comes from choosing what to rebuild and writing down what happened. The coding is the easy part.

### Choose the thing customers pay for

Rebuild the core workflow, not the whole product. For a field service scheduling tool that might be job creation, technician assignment and the mobile job sheet. For a compliance reporting product, ingesting a data file, applying the rules and producing the report. Ignore settings pages, admin screens and the long tail of features nobody buys the product for. Ask the deal team which two or three things customers would refuse to give up, and rebuild those.

### Work only from public material

Use the marketing site, documentation, published API references, demo videos, reviews and a trial account if the terms allow it. Do not use anything from the data room: no code, no internal documents, no customer data. That keeps the test honest (a real competitor wouldn't have them) and keeps you clear of confidentiality and IP problems. Keep the prototype internal and delete it after the decision.

### Timebox it and record everything

Two to five working days for one experienced engineer using current AI coding tools is enough to learn most of what the test can tell you. Record time spent on each part, where the AI tools did the work and where a person had to, and every point where the build stalled. The stalls are the finding.

### Test the copy against the real thing

Put the same inputs through both. Note where the replica matches, where it falls short and what it would take to close the gap. Usually the gap is not in features. It is in edge cases, data the original has accumulated, integrations with other systems, and the reliability that comes from years of production use.

## What the test tells you, and what it doesn't

A successful rebuild tells you the features are no longer a barrier to entry. That is real information. If a competent engineer reproduced the core of a £20m revenue product in four days, the investment case cannot rest on product superiority.

It tells you less than it appears to about how fast a competitor would win customers. A prototype is not a product. The last 20% of a business application (permissions, audit trails, data migration, uptime, support, security reviews, accessibility, the hundred edge cases a long-standing customer relies on) is most of the work, and AI tools help with it less than with the first 80%. You might be able to build a better version of a product's features without being able to break into its market.

The evidence on how much AI tools speed up experienced developers is also less settled than the rhetoric. [METR's 2025 randomised trial](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) found experienced open-source developers took 19% longer on real tasks when allowed AI tools, despite believing they were faster. Its [February 2026 update](https://metr.org/blog/2026-02-24-uplift-update/) said developers were probably sped up by then, but described its own newer data as very weak evidence of how much, partly because developers increasingly refused to work without AI. Read that as a reason for caution in both directions. The tools are improving fast, and nobody yet has a reliable multiplier to put in a model.

A failed rebuild tells you something too, but check why it failed. If it stalled on proprietary data, a regulatory approval or an integration that took the target years to win, that is a moat. If it stalled because the engineer was unfamiliar with the domain, run it again with someone who isn't.

## What actually protects a software business

When the features can be copied, these are what stop customers leaving. The table shows how I'd look for each one in diligence.

Protection
Evidence to look for
How AI coding tools change it

Proprietary data
Data the product generates or aggregates that no competitor holds, with contracts permitting its use
Little; a copy starts with an empty database

Integrations
Live, maintained connections to systems of record, used by most customers
Somewhat; integrations are faster to build, but partner approvals and customer-specific mapping still take time

Regulation and certification
Approvals, accreditations, audited processes, regulated sign-off
Little; certification takes months regardless of build speed

Distribution
Channel partnerships, marketplace positions, a trusted brand in a professional community
None directly; a new entrant still has to sell

Switching costs
High net revenue retention, long contracts, historical data and configuration that don't export
Weakening where migration tools improve; still strong where processes are built around the product

Services and people
Implementation, training, managed service and support that customers value
Mixed; this is where AI also lets competitors offer more for less

The pattern is that AI attacks the build cost and leaves most of the go-to-market and trust costs where they were. A product whose advantage was mostly "it was hard to build" is now exposed. A product whose advantage was "it's hard to get customers to trust and switch" is much less so.

There is one category where the threat is closer than it looks: products whose customers have capable IT teams. The competitor that matters may not be a start-up at all. It may be the customer's own developers, who now find an internal tool cheaper than the renewal. Ask the target which customers have tried to build their own, and how that went.

## Can AI replicate this software? Turning the test into an IC conclusion

A replicability test is only useful if it ends in a sentence the investment committee can act on. I'd write the conclusion in three parts. The examples below are illustrative, for a hypothetical field service software target.

First, the rebuild result, in plain terms: "One engineer reproduced the core scheduling workflow in three days. The replica matched the original on standard jobs and failed on multi-site contracts and the payroll integration."

Second, what protects the business, with evidence: "Customers rely on eight years of job history and the payroll integrations with the three largest UK providers. Net revenue retention is 108% and has not moved through two waves of AI-native competitors."

Third, what it means for the deal: price, structure and the plan. In practice the conclusion lands in one of three places.

Features replicable, strong protection elsewhere. Price as normal software. Put AI features and pricing model changes into the 100-day plan, because the incumbent can use the same tools.

Features replicable, weak protection. Revenue at risk over the hold. Model a scenario where a share of customers leave or force price reductions at renewal, and price that in. Consider structure: an earn-out tied to retention, or a lower entry multiple.

Features hard to replicate for reasons that will hold. Rare, and worth checking twice. If it's true, it supports the multiple.

The [AI disruption risk assessment](https://anywayventures.com/insights/ai-disruption-risk-assessment) sets out an eight-factor scoring grid that sits alongside this test, covering exposure beyond the product itself, such as pricing models tied to customer headcount. For targets whose own AI claims form part of the equity story, the [AI washing](https://anywayventures.com/insights/ai-washing-due-diligence) checks apply as well.

## Running it across a portfolio

The same thinking works at portfolio level without a full rebuild for every company. A light screen (what the product does, what protects it, a short attempt at rebuilding the core workflow, retention data) is enough to sort holdings into those that need nothing, those that need a plan, and those that need a proper review before the next exit conversation. Funds that do this before their buyers do get to set the narrative. The [AI disruption risk scorecard](https://anywayventures.com/resources/ai-disruption-risk-scorecard) has the scoring rubric and the rules for classifying each holding.

## How we run an AI disruption risk review

Our [AI disruption risk review](https://anywayventures.com/ai-disruption-risk) combines a replicability test on the core product with the evidence on data, integrations, regulation, distribution and switching costs, and ends in a written conclusion the investment committee can use. It costs £8k to £15k per deal over one to two weeks. Portfolio scans start at £4k per company.

If you have a software target where the AI question keeps coming up in committee, the [AI disruption risk review](https://anywayventures.com/ai-disruption-risk) page explains how we'd answer it.
[FREE RESOURCE · PDF + XLSX AI disruption risk scorecard DOWNLOAD FREE →](https://anywayventures.com/resources/ai-disruption-risk-scorecard)

## Questions

### What is a replicability test in due diligence?

A replicability test is a timeboxed attempt to rebuild the core of a target's software product using AI coding tools, working only from public material such as the website, documentation and a trial account. The time, cost and difficulty of the rebuild show how much of the product's value sits in features that are now cheap to copy, and how much sits in data, integrations, regulation and customer relationships.

### Is it legal to rebuild a competitor's software for due diligence?

Building a prototype from public information to inform an investment decision is generally lawful, but it needs care. Do not use code, documents or data from the data room, respect the terms of any trial account, and do not copy protected material such as interface assets or documentation text. Keep the prototype internal and destroy it after the decision. Take legal advice on the specific case.

### Which software companies are most at risk from AI coding tools?

Those whose value sits mainly in features and interface: point tools, reporting and dashboard products, simple workflow apps and internal tools sold as software. Apollo has reportedly singled out data visualisation as a high-risk category. Products that hold proprietary data, sit inside regulated processes, integrate deeply with systems of record or come bundled with services are much harder to displace, even when their features are easy to copy.

### Does AI make software development cheaper for incumbents too?

Yes, and diligence should credit it. The same tools that let a new entrant copy features let the incumbent ship faster, retire technical debt and add AI features to an existing customer base. The question is which side gains more. An incumbent with customers, data and distribution that adopts the tools well is usually better placed than a start-up with only the tools.

### How long does an AI disruption risk review take?

For a single software target, one to two weeks is typical, including the replicability test, a review of what protects the business and a written conclusion for the investment committee. A lighter screen across a portfolio, without a full rebuild, can be done in a few days per company and is useful for deciding where to look harder.

Brannan Coady
Founder of Anyway Ventures. Exited founder, former CEO of YourParkingSpace, investor and board member; now building AI inside PE-backed and mid-market businesses. [More →](https://anywayventures.com/brannan-coady)

## Keep reading

[ALL WRITING →](https://anywayventures.com/insights)
[DILIGENCE · 10 MIN AI disruption risk assessment: will AI hollow out the target? AI washing covers targets that overclaim AI. This covers the other problem: targets that AI threatens. An eight-factor grid, the evidence to request, and what the score should do to the deal.](https://anywayventures.com/insights/ai-disruption-risk-assessment) [DILIGENCE · 9 MIN AI washing: how to test a target's AI claims in due diligence When the equity story leans on AI, the diligence question is simple. Take the AI away: what's left, and what does it cost to keep it running?](https://anywayventures.com/insights/ai-washing-due-diligence) [TRAINING · 8 MIN Vibe coding for non-developers: what to build, what not to AI coding assistants and no-code tools mean a finance analyst can build a working internal tool in an afternoon. Here is what that looks like, and the guardrails a delivery firm would put round it.](https://anywayventures.com/insights/vibe-coding-for-non-developers)
