← ALL SECTORS
SECTOR · SOFTWARE AND SAAS · UK

Technology and AI due diligence for software and SaaS businesses.

AI has turned software diligence from a hygiene check into the valuation argument. Buyers now want to know whether the product still sits in the customer's workflow when agents do the work, what the AI features cost to run, and whether retention data backs up the story. This page covers what to test, for investors and for the software CEOs and CTOs preparing to be tested.

FORPE and VC deal teams, operating partners, investment committees, software CEOs and CTOs preparing for a sale
UPDATEDOctober 2026

KEY TAKEAWAYS

  • AI is eroding pricing power for SaaS products priced per seat for tasks agents can now do, while systems of record, products in regulated workflows and businesses that own their data are holding value.
  • Gartner's July 2026 forecast puts up to $234bn of enterprise application spending at risk from agentic AI by 2030, about 20% of enterprise SaaS spending, so around 80% does not move.
  • Median SaaS EV/EBITDA multiples fell from 20.4x to 11.7x in H1 2026 even as global SaaS M&A reached a ten-year high of $439.7bn (Forvis Mazars and PitchBook, 2026).
  • In SaaS deals, the questions most likely to move price cover per-seat ARR share, falling seat counts, cost per call of each AI feature, and contractual rights to customer data.
  • In SaaS diligence, rebuild gross margin with inference, vector database, model hosting and evaluation costs in cost of sales, because AI features carry a variable cost that companies often hide in R&D or cloud.

AI is changing who does the work inside a software product, and that changes what SaaS businesses can charge for. Products priced per seat for tasks that agents can now do are losing pricing power, while systems of record, products embedded in regulated workflows and businesses that own their data are holding value. SaaS due diligence in 2026 has to test that exposure directly, alongside gross margin after AI inference costs and retention by cohort. Investors should price it at entry, and software CEOs should evidence the answer before a buyer asks.

I've been on both sides of this. I co-founded Netsells, a York digital product consultancy that built digital products for clients including Uber, Trainline, Jaguar Land Rover and the BBC, and ran it as CEO until it merged with BGF-backed hedgehog lab in 2023. I was Chief Product Officer, then COO, then CEO at YourParkingSpace, a parking marketplace, and sat on the Flowbird executive committee and deal team for its sale to EasyPark Group. I've been interim CTO at a fintech, Wombat, and co-founded Railguard, a rail-tech business sold to Trainline in 2019. I've answered the diligence questionnaire and I've written it.

Rebuilding a target with AI coding tools, products built on someone else's model and scoring a whole portfolio each have their own piece: the replicability test, wrapper or moat and the portfolio triage framework. This page is about which kinds of software are exposed, what the KPIs now hide and what to ask in the data room.

Is AI killing SaaS?

No. It is killing a pricing assumption, and it is exposing products that were mostly interface.

Classic SaaS assumed revenue grew as customers hired people, and that the marginal customer cost almost nothing to serve. AI pressures both. An agent doing part of a team's work shrinks the seat count at renewal, and an AI feature that calls a model on every request adds a variable cost subscription software never had.

The market has already priced a lot of this in. Bain's Private Equity Midyear Report 2026 describes a near 30% drop in public software valuations in February 2026, and software marks in PE portfolios down roughly 8% by the end of March. Europe fared better than the US, with marks down 4.2% against 8.9%. Gartner's July 2026 forecast puts up to $234bn of enterprise application spending at risk from agentic AI by 2030, about 20% of enterprise SaaS spending, and its analysts describe seat-based pricing as an existential problem for vendors who defend it.

Twenty per cent is a lot of money to move, and it means 80% doesn't. The job in diligence is to work out which side of that line a product sits on. Usually the honest answer is parts of both.

The sell-side analysts don't agree with each other either, which is a reason to do the work rather than borrow a view. Pinsent Masons' August 2026 analysis of the "SaaSpocalypse" (Pinsent Masons, 2026) quotes UBS warning in March that "purely workflow-oriented SaaS tools" are the most exposed, because their value sits in the interface and process coordination layers that agents can replicate, while software with embedded governance, regulatory complexity and heavy customisation is "generally tougher for AI to disrupt". The same piece quotes HSBC analysts arguing in June that the market had come "to overestimate AI risk to software", pointing to double-digit growth at the largest vendors. Both can be right about different products.

Where AI changes the economics of software

Bain's 2025 diligence framework sorts AI's effect on a target into three categories: augmentation (AI improves cost, efficiency and product without changing the model), transformation (the model needs substantial change, with real investment and speed) and revolution (AI threatens the business model itself). The same piece sets out five questions that still make a good spine for software diligence: will the business model be upended, will volumes or pricing be affected, will the basis of competition change, what product improvements are possible, and will there be meaningful cost savings. Software companies spread across all three categories, often within one product.

Bain's own numbers explain why the sector has moved from the first category towards the second. Its September 2026 technology report, Software Investing in the Age of AI and Slower Growth, shows median revenue growth for US public software companies falling from 27% in 2021 to 10% in 2025, and median net revenue retention for public enterprise SaaS falling from 117% to 108% over roughly the same period. Tech buyouts made in 2020 to 2022 were tracking a 2.1x median gross multiple, against 2.9x for deals made in 2010 to 2019. Growth has roughly halved, before AI disruption does anything at all.

Here is how I'd place the main software types:

  • Systems of record (ERP, practice management, core banking, EHR, payroll, policy administration): mostly augmentation. Agents need somewhere authoritative to read from and write to. Bessemer's State of AI 2025 describes a shift "from systems of record to systems of action" but expects the ERP replacement cycle to take many years. The risk is value moving to an agent layer on top, with the system of record squeezed on price.
  • Workflow and productivity tools priced per seat (helpdesk, CRM for small teams, project tools, scheduling): transformation, sometimes revolution. These are where seat compression bites first.
  • Point tools whose output is a document, report or dashboard: the most exposed. A general-purpose model with a connector can now produce a lot of what these products charge for.
  • Marketplaces: different exposure. AI agents may start to search and book on the user's behalf, which threatens a marketplace that earns its take rate from discovery, and favours one that earns it from trusted supply, payments and fulfilment. In a parking marketplace like YourParkingSpace, the supply side (the car parks, the driveways and the operators behind them) is the hard thing to copy. The app is the easy part.
  • Data businesses: they split. Those that own proprietary, licensed or hard-to-collect data gain, because every AI system needs it. Those that repackage public data lose, because a model can do the repackaging.
  • Infrastructure and security software: mostly beneficiaries. Apollo's chief economist noted in July 2026 that security and infrastructure software had significantly outperformed while application and vertical software were "severely lagging".

Software and SaaS deals in numbers

39.7%of companies receiving UK-led private capital investment in 2025 were in ICT, UK Private Capital, 2026
1,334UK TMT deals in 2025, the most active sector, down 9% on 2024, Experian MarketIQ, 2026
11.7xmedian SaaS EV/EBITDA multiple in H1 2026, down from 20.4x, Forvis Mazars and PitchBook, 2026
68%of audited codebases contained open source licence conflicts, up from 56%, Black Duck OSSRA, 2026

Technology is the largest sector in UK private capital. UK Private Capital's Report on Investment Activity 2025, published in May 2026, shows ICT accounting for 39.7% of companies backed by UK-led investment and 39.1% of the amount invested. Experian MarketIQ's full-year 2025 review recorded 1,334 TMT transactions in the UK, the most of any sector, though aggregate value fell from £56bn to £22bn. In London alone it counted 723 TMT deals, with software accounting for more than half.

Deal appetite has turned more selective since. EY's Q1 2026 Private Equity Pulse found around 60% of GPs had increased diligence on AI disruption risk and nearly two-thirds were taking a more targeted investment approach. Technology fell from roughly 30% of global PE deployment by value in 2025 to 12% in the first quarter of 2026. EY's Q2 edition reported tech-focused PE deal value down 50% year on year in the first half.

The large firms have formalised it. Bloomberg reported in June 2026 that Apollo screens every new software investment for AI disruption risk, using a framework that splits software into 12 to 14 categories ranked by susceptibility, and that it had overlaid the same framework on its existing portfolio. Apollo's chief economist later pointed out that application software, the most exposed segment, carries $146bn of outstanding loans, mostly to borrowers rated B- or lower. For a leveraged SaaS buyout, that shapes the refinancing conversation at exit.

The European picture is less extreme but has the same shape. AlixPartners' June 2026 piece on AI disruption in European SaaS (AlixPartners, 2026) puts roughly €7bn to €8bn of software loans due by the end of 2028, much of it written in the 2020 to 2021 low-rate window, and scores about 18% of the 561 larger EMEA software and IT services companies it analysed, around €44.5bn of revenue, as at high risk of AI disruption, mostly because of weak vertical and data moats. Its 2026 enterprise software predictions report found that 56% of mid-market companies were expected to grow by at least 15% in 2023, against 22% now. Its advice to sponsors and lenders is the one this page is built on: ARR may no longer be a reliable guide to future performance, so diligence has to look at data assets, workflow embedding and AI readiness.

Strategic buyers are moving the other way. The Forvis Mazars and PitchBook H1 2026 State of SaaS report found global SaaS M&A at a ten-year high of $439.7bn, driven by AI-related strategic deals, while PE activity fell back to 702 deals worth $58.8bn and median SaaS EV/EBITDA multiples dropped from 20.4x to 11.7x. Good software assets still trade, but the premium now attaches to evidence rather than to the label.

AI adoption inside the sector is high. The ONS reported in July 2026 that 58% of information and communication businesses used AI, against about 35% across all UK businesses with ten or more staff. A software company not using AI in its own engineering and support is behind its sector, and a buyer will ask why.

AI use cases that pay back in a software business

Most software businesses have two AI questions: what to put in the product, and what to do inside the company. Buyers want evidence on both. Bain's The New Era in Tech Investing Starts Now (April 2026) frames the test as whether management can show proof points, and cites Zendesk reporting $200m of AI-related ARR across 20,000 AI customers.

Workflow What AI does Evidence or typical effect Readiness signals
Tier 1 and tier 2 support An agent answers from docs, ticket history and account data, and hands off with context Deflection rate and cost per ticket, measurable within a quarter; also the template for a paid AI feature Clean help centre, tagged ticket history, support tooling with an API
Engineering throughput Coding assistants for tests, refactors, migrations and code review Uneven across teams, so measure lead time and change failure rate rather than collect anecdotes Good test coverage on core paths, CI that runs in minutes, a backlog of upgrade work
Onboarding and implementation Data mapping, migration scripts, config generation from a customer's existing system Shorter time to first value, fewer implementation days per customer, better services margin Repeatable implementation playbook, structured customer data, a named implementation lead
In-product AI features (copilots, summaries, drafting) Uses the customer's own data inside the product to do part of the job Only counts when usage is measured and someone pays: attach rate, active use, AI add-on revenue Feature-level usage telemetry, cost per call tracked, eval set for each feature
Agentic features that complete a task end to end The product does the work (reconciles, schedules, files, books) rather than helping a user do it The route to usage or outcome pricing; needs proof of accuracy and a clear liability position System-of-record data, audit trail, human review step, pricing team willing to experiment
Renewals and churn prediction Flags accounts with falling usage, seat contraction or support escalation Earlier intervention on churn; test against last year's actual churners Product usage data joined to CRM and billing
Data products from customer data Benchmarks, predictions or models built across the customer base Can become a moat and a new revenue line, if the contracts allow it Contract rights to aggregate and use customer data, consent records, data quality

On engineering gains, METR's 2025 study found experienced open source developers were slower with AI tools on their own repositories, and its February 2026 update revisited the picture. If an IM claims a big productivity gain, ask for deployment data from before and after.

Where AI threatens the SaaS business model

Per-seat pricing and AI: the renewal problem

A seat-priced contract assumes the customer's team grows. When a customer's support team goes from 40 agents to 25 because an AI agent now handles the first line, the vendor's renewal shrinks by three-eighths, even if the vendor's own AI is the reason. That's the cruel part: the better the product's AI works, the faster seat revenue falls, unless pricing changes.

Vendors are already moving. Deloitte's 2026 TMT Predictions cites Gartner's forecast that by 2030 at least 40% of enterprise SaaS spend will shift towards usage-, agent- or outcome-based pricing, and expects disruption to hit customer service first before spreading to ERP and CRM. Bain's September 2026 report describes AI "shifting revenue models from seat-based subscriptions toward usage and outcomes", and warns that AI revenue priced on usage "can be bursty and unreliable". It suggests tracking three AI revenue buckets separately: traditional AI and machine learning, AI add-ons such as copilots, and agentic products.

For diligence, three tests matter more than the pricing page:

  1. Split ARR by pricing basis. Seats, usage, platform fee, outcome. Then split the seat ARR by customer function. Seats sold to frontline operators in support, sales development, data entry or claims handling are more exposed than seats sold to the people who approve, sign off or carry liability.
  2. Look for seat contraction already happening. Downgrades at renewal, accounts paying for more seats than they use, and customers asking for "AI pricing". The data is in the billing system, not the board pack.
  3. Check the plan. If the model grows ARR through seat expansion in customers who are adopting AI hard, it is double counting.

Gross margin after inference costs

AI features carry a variable cost that subscription software never had. ICONIQ's 2026 State of AI report found average gross margins on AI products of 45% in 2025, projected to reach 53% in 2026. Bessemer's State of AI 2025 put its efficient fast-growing AI companies at about 60% gross margin and its fastest-growing cohort at about 25%, often negative. Those are AI-native businesses, but the same arithmetic now applies to the AI module inside a classic SaaS product.

Take an illustrative case. A vertical SaaS business adds an AI drafting feature and sells it as a £20 per user per month add-on to 2,000 users: £480k a year. Heavy users run it 30 times a day. If each call costs 3p in model fees, a heavy user costs about £20 a month over 22 working days, which is the whole of the add-on price. The margin on the feature depends entirely on how many users are heavy. Nobody in the business has looked, because the model bill lands in the engineering cloud account rather than cost of sales.

That last part is common. Forvis Mazars noted in its H1 2026 report that companies increasingly use non-GAAP measures to exclude AI-related costs from core operations. In diligence, rebuild gross margin with all inference, vector database, model hosting and evaluation costs in cost of sales, and model what happens at three times current AI usage.

Who captures the efficiency gain

When AI lets the customer do the same work with fewer people, someone keeps the saving. A vendor that holds the workflow and the data can reprice towards the outcome and share the gain. A vendor that is one of several tools an agent orchestrates watches the customer keep it, and becomes a commodity API.

Feature replicability

The visible features of many SaaS products can now be rebuilt quickly, and the replicability test covers how to check. A rebuild prices the features, which is the cheap half. Ask what would stop a customer switching to the rebuild and you find the real moat, or its absence.

Vertical SaaS due diligence: systems of record, data moats and switching costs

Listed markets have lumped vertical software in with the losers this year. I think that is too blunt, and it is where careful diligence earns its fee.

Vertical SaaS that holds up tends to share a few features. It is the system of record for a process the customer is legally or commercially accountable for: a care plan, a dispensing record, a payroll run, a planning application, an insurance policy. It holds years of customer history that is painful to migrate. It integrates with industry-specific systems (a regulator's portal, a payment scheme, a laboratory instrument, an industry data exchange) that a general-purpose agent doesn't know exist. And it is often bundled with implementation, training and support, so the switching cost is organisational as well as technical. Baird's February 2026 panel of software investors put it plainly: buyers see established players as well positioned where they are a system of record or have defensible data moats.

Exposed vertical SaaS looks different. It generates documents, schedules or reports from data that lives somewhere else, its integrations are CSV exports, and its small-business customers will happily try a general-purpose assistant with a connector. Annual contracts hold up its retention, not usage.

Horizontal SaaS faces a harder version of the test, because it competes with what the model providers and the largest platforms ship as standard. Unless it is a system of record (finance, HR, identity) or has distribution an agent can't route around, assume general-purpose assistants will reach its core use case within the holding period.

Data moats need the most scepticism. Three questions settle most of them. Does the company have the contractual right to use customer data across customers, for training or benchmarking, and can it show the clause? Is the data something a competitor couldn't buy or scrape? And does the product get measurably better with more of it? A "data moat" that fails any of those is a database.

SaaS technical due diligence questions for UK deals

These sit on top of a standard technology due diligence checklist, which covers architecture, code quality, security and IP in general terms. The questions below are the ones that come up specifically in software and SaaS deals.

Question Why it matters Red flag
What share of ARR is priced per seat, and which job functions hold those seats? Seats held by roles AI agents now do will contract at renewal Over half of ARR in seats sold to frontline operators, with no pricing change planned
Show seat counts per account over the last eight quarters. Where are they falling? Contraction shows up in the billing data before it shows up in churn Falling seats in the largest accounts, explained away as "customer restructuring"
What does each AI feature cost per call, and where does that cost sit in the P&L? Inference costs hidden in R&D or cloud overstate gross margin No per-feature cost tracking; model bills on a founder's credit card
Which model providers does the product depend on, and what happens if one deprecates a model or changes price? Provider changes can force a rebuild or margin hit within months One provider, no eval set, no tested fallback; see wrapper or moat
Is this product the system of record for anything? If not, what is, and how does the product connect to it? Systems of record are where agents read and write; tools beside them are easier to bypass The product holds copies of data owned elsewhere, synced by nightly export
Do customer contracts allow aggregate use of customer data for product improvement or AI training? A data moat without contractual rights is a liability Customer data already used to fine-tune models with no clause permitting it
What does the open source licence scan show, including code written with AI assistants? Black Duck's 2026 OSSRA found licence conflicts in 68% of audited codebases; AI tools can reproduce copyleft code without its licence AGPL components in the hosted product; no policy on AI-generated code
Is the platform multi-tenant, and how is customer data isolated, including in AI features and vector stores? Retrieval features can leak one customer's data into another's answers Shared vector index across tenants with no access filtering
What did the last two penetration tests find, and what was fixed? Is MFA enforced on every admin and support path? The ICO fined Advanced Computer Software £3.07m in 2025 after a ransomware attack, citing gaps in MFA and patching Critical findings still open; support staff with standing production access
Who wrote the core product, and who could maintain it if they left? Software value often sits in two or three heads; see key-person risk One engineer behind most commits to billing, permissions or the data model
How long does a release take, how often does it happen, and how often does it break? Slow delivery caps every value creation plan, including AI features Monthly manual releases, no staging, no rollback
What is the cloud cost per customer, and how has it moved against ARR over 24 months? Hosting and AI costs growing faster than revenue erode the margin buyers paid for Nobody can allocate cloud cost to customers or features
Which AI features in the IM are live, and what are their weekly active users? IMs routinely describe pilots as products "AI-powered" in the IM, no usage telemetry in the data room
What would a customer need to do to leave, and how long did the last three departures take? Switching cost is the real moat, and churned customers show what it is Customers leave within a month with a CSV export
Does the product sell into EU financial services, the NHS, the public sector or regulated hiring? DORA, NHS data rules, public sector security requirements and the EU AI Act all change the cost of serving those customers Large customers in regulated sectors, no contract clauses or certifications to match
Which parts of the roadmap depend on AI, and what has the team shipped against last year's roadmap? AI roadmaps are cheap to write; delivery history is the discount rate Revenue plan depends on agentic features the team hasn't started
How concentrated is ARR, and do the largest contracts have change-of-control, termination for convenience or most-favoured-customer pricing clauses? A few large customers who can walk on a sale, or who are cutting the roles the product serves, carry the valuation Top ten customers over 40% of ARR, change-of-control consent needed from several, nobody has read the MSAs
What R&D tax relief has been claimed, under which scheme, and has HMRC opened an enquiry? Software companies are heavy claimants. HMRC's merged RDEC and ERIS schemes replaced the old regimes for periods from 1 April 2024 and restrict overseas expenditure, which catches offshore development teams Claims prepared on a contingency fee, offshore contractor costs still included, an open enquiry not disclosed in the IM
What outages have customers had in the last 24 months, and when was a full restore from backup last tested? SLA credits, churn after incidents and ransomware recovery time all come back to this No incident log, a single cloud region, backups in the same account as production, restore never tested
For marketplaces: what share of transactions start with search or discovery, and how much supply is exclusive? AI agents that book on a user's behalf threaten discovery-led take rates Take rate justified by traffic, with no exclusive or contracted supply

The questions on seats, inference cost and data rights are the ones most likely to move price, so they go at the top of the day-one evidence request. For what a review of this scope costs, see technical due diligence cost in the UK.

How long SaaS due diligence takes, and who does what

A full SaaS process usually runs four workstreams at once: commercial (market, customers, competition), financial (quality of earnings, ARR, working capital), legal (contracts, IP, data protection, employment) and technical. AI cuts across all four, which is the problem. The commercial team asks whether agents will replace the product, the financial team finds inference costs in R&D, the lawyers read the data clauses, and nobody joins the answers up. I'd put one person in charge of the AI question across the workstreams and have them write the disruption conclusion the IC reads.

Technical and AI diligence on a lower mid-market software target does not need to take months. Our vertical SaaS engagement ran nine days from scoping to IC. What slows it down is a data room without billing exports, usage telemetry or cost allocation, so ask for those on day one.

Buy and build platforms have a particular version of this. Pinsent Masons noted in its August 2026 analysis that, for the PE funds it works with on buy and build, every acquisition from the platform to each bolt-on is now checked for AI exposure, and that separate adviser teams working in silos make those checks slower and more expensive. A short, repeatable AI screen for bolt-ons, with a deeper review only where it flags a problem, keeps the programme moving.

Metrics to test in the data room

SaaS KPIs were built for a world where revenue tracked headcount and gross margin was close to pure software. Most of them still work. Several now hide things.

Metric What good looks like How AI moves it How to evidence it
Gross revenue retention (GRR) SaaS Capital's 2025 survey put the median at 91% for private B2B SaaS, 95% where ACV exceeds $250k Seat contraction shows up here as downgrades before it shows as churn Cohort tables by start year, ACV band and pricing basis, rebuilt from billing data
Net revenue retention (NRR) Median 101% across private B2B SaaS, 101% for equity-backed companies (SaaS Capital, 2025) AI add-ons can lift NRR while seat contraction quietly pulls it down; one masks the other NRR split into seat movement, price increases, AI add-on revenue and usage
Gross margin Should include every cost of serving a customer, AI included Inference, vector storage and model hosting push it down as AI usage grows Rebuild cost of sales with AI costs reclassified from R&D and cloud
AI revenue Separately tracked, by type: AI and ML features, add-ons, agentic products The real test of whether AI is a product or a slide Invoices and usage logs for each AI SKU; attach rate and retention of AI buyers
ARR quality Contracted, recurring, with usage revenue identified separately Usage and outcome pricing make ARR less predictable ARR bridge with usage revenue on a trailing basis, not annualised from a good month
Rule of 40 (growth plus profit margin) Above 40 is strong; well below is common at lower mid-market scale AI can raise margin through internal efficiency, or depress it through inference costs Use the same definitions throughout: ARR growth and EBITDA margin, stated
Engineering throughput Frequent releases, short lead time, low change failure rate Coding assistants should show here if they work DORA metrics from the deployment pipeline, 12 months
Seats per account, by function Stable or rising in the accounts the plan depends on The leading indicator of AI-driven contraction Billing exports per account, eight quarters
ARR reconciled to revenue and cash Company ARR ties to recognised revenue and to bank receipts, with differences explained Usage and credit-based AI pricing widen the gap between annualised run rate and what is billed and collected Monthly ARR bridge rebuilt from invoices, reconciled to the management accounts and bank statements
CAC payback and sales efficiency Depends on ACV and motion; what matters is the trend and the split by channel AI search and AI sales tools change both how buyers find software and what it costs to win them Sales and marketing cost by channel against new ARR by cohort, 24 months
Customer concentration No single customer large enough to change the deal on its own Large customers adopting AI fastest are the likeliest to cut seats or demand outcome pricing ARR by customer for the top 20, with contract end dates and pricing basis

The SaaS Capital figures come from its 2025 B2B SaaS Retention Benchmarks, based on more than 1,000 private companies, mostly in the US. Companies define NRR in different ways, so ask for the calculation and recompute it from billing data. A big gap between the company's figure and yours is common enough that I check every time.

Regulation and standards that shape AI and technology in UK software

  • UK GDPR, the Data Protection Act 2018 and the Data (Use and Access) Act 2025. Most SaaS businesses are processors for their customers' data, and processors can be fined directly. The ICO's £3.07m fine on Advanced Computer Software Group in March 2025, after a ransomware attack on its health and care subsidiary affected 79,404 people, is the case every SaaS buyer should know. The new automated decision-making rules in section 80 of the Data (Use and Access) Act 2025, fully in force since 5 February 2026, matter for any product that makes significant decisions about people, and the ICO consulted on draft guidance on automated decision-making in 2026. Its guidance on AI and data protection is the reference for training on customer data and running DPIAs on AI features.
  • The EU AI Act. It reaches UK software companies with EU customers. Under the Digital Omnibus on AI, high-risk obligations for standalone (Annex III) systems now apply from 2 December 2027, while transparency duties for systems that interact with people or generate content have applied since August 2026. HR tech, credit scoring, education and some insurance products are the SaaS categories most likely to fall into high-risk.
  • DORA (EU Regulation 2022/2554). Applied since 17 January 2025 to EU financial entities, and it requires specific contract terms with their ICT third-party providers. A UK SaaS business selling into EU banks or insurers will be asked to accept audit rights, exit plans and incident reporting clauses. See the regulation text.
  • The Cyber Security and Resilience (Network and Information Systems) Bill. Introduced to Parliament on 12 November 2025, it had not received Royal Assent by late October 2026: it passed the Commons in June 2026 and its Lords report stage was scheduled for 26 October 2026 (UK Parliament, 2026), with most duties to follow in secondary legislation. It expands the NIS Regulations, and the government's policy statement proposes bringing managed service providers into scope, regulated by the ICO. SaaS businesses that also manage customers' IT should check whether they will be caught.
  • The National Security and Investment Act 2021. Buying more than 25% of a company active in one of the 17 mandatory sectors, which include artificial intelligence, data infrastructure and communications, needs government approval before completion, and a notifiable deal completed without it is void under section 13. The Cabinet Office's fifth annual report, published in July 2026, recorded 1,324 notifications in the year to March 2026, up 15% (Herbert Smith Freehills Kramer, 2026). In March 2026 the government confirmed it will narrow the AI schedule and widen data infrastructure to include certain cloud and managed service providers, through secondary legislation expected later in 2026 (Mishcon de Reya, 2026). Most SaaS deals won't be caught, but an AI or hosting business might be, so check in week one.
  • Cyber Essentials, ISO/IEC 27001 and SOC 2. Cyber Essentials is commonly required for UK public sector contracts. ISO 27001 and SOC 2 reports are what enterprise buyers ask for. Treat them as evidence of discipline, not proof of security. ISO/IEC 42001, the AI management system standard, is optional but gives AI governance a structure a buyer can check.
  • Open source licences. Not regulation, but the commonest legal finding in software diligence. AGPL in a hosted product and AI-generated code with untraced provenance both belong in the SPA conversation.

A 100-day plan for a newly acquired software business

The plan depends on what diligence found, but most software deals I've seen benefit from the same order.

Days 1 to 30: get the numbers you didn't have.

  • Rebuild gross margin with AI and cloud costs in cost of sales, by product and feature.
  • Pull seats per account for every customer over eight quarters and flag contraction.
  • Close anything critical from diligence: open pen test findings, missing IP assignments, admin access held by one person, open source licence conflicts.
  • Put retention arrangements around the two or three engineers who hold the core of the product.

Days 31 to 60: decide the pricing and AI position.

  • Choose which AI features to charge for, and how: add-on, credits, usage tier or outcome.
  • Run one pricing experiment with new customers and one renewal cohort, with a written success measure.
  • Put internal AI to work where it pays back fastest, usually support and engineering, and measure cost per ticket and lead time from a baseline.
  • Set up feature-level usage and cost telemetry if it doesn't exist. Without it, every later AI decision is a guess.

Days 61 to 100: build the evidence the exit buyer will ask for.

  • Ship or improve one AI feature that customers use weekly, with an eval set and a cost ceiling per call.
  • Write a one-page answer to "what happens when agents do the user's job?", with data, and test it with the board.
  • Add AI revenue, AI cost and seat movement to the monthly board pack.
  • Agree the data rights position with legal and update customer terms at the next renewal cycle.

An AI operating partner can carry this alongside management where the team lacks bandwidth. By month four the board should see AI in the numbers, not only in the roadmap.

Preparing a SaaS business for due diligence: an AI roadmap before sale

For software CEOs and CTOs 12 to 24 months from a process: the buyer will ask the questions in the table above, so ask them first.

Write your own disruption answer. One page: which parts of the product an AI agent could do, which customers would be tempted, and why they would stay. Back it with retention data and the exit interviews of customers who left. If the answer is uncomfortable, that's useful too: you have time to change the product or the pricing before someone else prices the risk.

Get AI into production where it can be measured. Buyers credit features with usage data and revenue, and they discount roadmaps. Two features with six months of weekly active users and a known cost per call are worth more in a sale than ten on a slide. I go into the evidence buyers credit in will buyers pay for your AI?.

Fix the pricing story before the buyer finds it. If seats are contracting in some accounts, show you've seen it and show what you're doing about it: an AI tier, a platform fee, usage pricing for new customers. A buyer who discovers seat contraction in the billing data you didn't mention will assume the worst about everything else.

Clean up the technical findings that cost little to fix. Open source licence conflicts, missing contractor IP assignments, model provider contracts on personal cards, a single engineer with every admin password, critical pen test findings. Each is cheap to fix now and expensive when a buyer prices it. Vendor technical due diligence covers the order to do this in, and our exit readiness checklist is a working list.

Put inference costs in cost of sales now. Restating them during diligence looks like a correction. Showing them from the start looks like control.

None of the older questions have gone away. Who owns the code, how releases work and what the security posture really is still decide a lot of software deals. AI adds a new set on top.

Red flags and green flags in software and SaaS deals

Red flags:

  • Most ARR is priced per seat for roles AI agents now do, and the plan assumes seat expansion.
  • Seat counts are falling in the largest accounts and nobody mentioned it.
  • AI costs sit outside cost of sales, or nobody can tell you the cost per call.
  • The product depends on one model provider with no eval set and no tested fallback.
  • "Proprietary data" with no contract clause allowing its use across customers.
  • AGPL code in the hosted product, or no policy on AI-generated code.
  • A shared vector store across tenants with no access filtering.
  • NRR that only works on the company's own definition.
  • AI features in the IM with no usage telemetry in the data room.

Green flags:

  • The product is the system of record for a process customers are accountable for.
  • Pricing already includes a platform fee, usage or AI tier, and renewals have been tested on it.
  • GRR above the median, holding across recent cohorts.
  • AI features with weekly active users, revenue and a known gross margin.
  • Contracted rights to aggregate customer data, and a product that improves with it.
  • Management that has written its own disruption answer and can defend it with data.

Case example: vertical SaaS diligence in nine days

We ran tech, product and AI diligence on a vertical SaaS target, taking nine days from scoping to investment committee. Two items from the resulting plan were delivered in the first quarter of ownership.

On a vertical SaaS deal the AI disruption question and the technical questions overlap. Whether the product is a system of record, whether its data is owned and usable, and whether the team can ship AI features at a sensible cost per call are one question asked three ways.

How we work with software investors and operators

For investors, we run technical due diligence and AI due diligence on software and SaaS targets, with an AI disruption risk review where the investment case depends on the product surviving agents, and VC technical due diligence for growth rounds. Every finding is written so the deal team can use it in price, structure or the 100-day plan. The AI disruption risk scorecard is a free first pass if you want to screen a target yourself.

For software CEOs and CTOs, we help prepare the evidence a buyer will ask for, from the disruption answer to the open source scan, and help get AI features into production with costs a buyer can see. If you have a deal or a sale coming up, a short conversation is usually enough to tell whether we can help.

How we help in software and saas

AI disruption risk review
Tests whether the product survives agents doing the user's job, with a written conclusion an IC can use.
MORE →
Technical due diligence
Code, architecture, open source licences, security and team, with every finding costed against price or the plan.
MORE →
AI due diligence
Checks what the AI features do in production, what they cost per call and who owns the data behind them.
MORE →
VC technical due diligence
A faster read for growth rounds: can this team and this codebase carry the next two years of the plan?
MORE →
Exit readiness
For software CEOs and CTOs 12 to 24 months from a sale: fix the cheap findings and evidence the AI roadmap.
MORE →
AI operating partner
Hands-on help after completion to reprice, ship AI features that earn money and protect gross margin.
MORE →
FREE RESOURCE · PDF + XLSX
Technical due diligence checklist
DOWNLOAD FREE →
FREE RESOURCE · PDF + XLSX
AI disruption risk scorecard
DOWNLOAD FREE →
FREE RESOURCE · PDF + XLSX
Technology exit readiness checklist
DOWNLOAD FREE →

Questions

Is AI killing SaaS?

No, but it is killing some SaaS business models. Products priced per seat for work that AI agents now do, with little proprietary data and shallow integrations, are losing pricing power. Systems of record, products embedded in regulated workflows and businesses with owned data are holding up better. Gartner estimates up to $234bn of enterprise application spend is exposed to agentic AI by 2030, about 20% of SaaS spending. That is a large shift, but most of the market remains.

What does SaaS due diligence cover in a private equity deal?

Commercial diligence tests the market and customers, financial diligence tests ARR quality and the numbers behind it, and technical diligence tests the code, architecture, security, open source licences and the engineering team. In 2026 most buyers add AI diligence: exposure to AI disruption, the real cost and usage of AI features, model provider dependency and data rights. The useful output ties each finding to price, deal structure or the 100-day plan.

How does AI affect per-seat SaaS pricing?

If a customer's team shrinks because AI does part of the work, a per-seat contract shrinks with it at renewal, even when the product is delivering more value. Vendors are moving towards hybrid pricing: a platform fee plus usage, credits or outcomes. In diligence, split revenue by pricing basis, find the customers whose seat counts are falling, and check whether the plan assumes seat growth that the customer's own AI adoption will reverse.

Is vertical SaaS safer from AI disruption than horizontal SaaS?

Often, but not automatically. Vertical SaaS that is the system of record for a regulated process, holds years of customer data and integrates with industry-specific systems is hard to displace. Vertical point tools that only generate documents, reports or schedules are as exposed as any horizontal product. Listed markets have treated vertical software harshly in 2026, so the useful work is separating the two inside each target.

What SaaS metrics matter most in due diligence now?

Gross revenue retention and net revenue retention by cohort and by pricing basis, gross margin with AI inference costs included, ARR split between seats, usage and platform fees, and the Rule of 40. SaaS Capital's 2025 survey of over 1,000 private B2B SaaS companies put median NRR at 101% and median GRR at 91%. Check the definitions behind the numbers, because companies calculate them differently.

How should a SaaS company prepare its AI roadmap before a sale?

Start 12 to 24 months out. Ship one or two AI features customers pay for or use measurably, track their cost per call against revenue, and write down which parts of your product an AI agent could replace and why customers would still stay. Fix open source licence conflicts, IP assignments and model provider contracts. Buyers credit production usage and margin data, and they discount roadmap slides.

Brannan Coady, founder of Anyway Ventures
Written by Brannan Coady
Founder of Anyway Ventures. Co-founder and CEO of Netsells for thirteen years, then Chief Product Officer, COO and CEO of YourParkingSpace and a member of the Flowbird Group executive committee through its sale to EasyPark (now Arrive). Investor and board member; now leads technology and AI due diligence and delivery for PE-backed and mid-market businesses. More about Brannan →

Related

ARTICLE · 8 MIN
Could a small team with AI coding tools rebuild this product?
ARTICLE · 9 MIN
Wrapper or moat? 12 questions that show whose AI it really is
ARTICLE · 12 MIN
AI disruption risk across a portfolio: a triage framework
OTHER SECTORSParkingManufacturingDistribution and logisticsProfessional servicesField services, FM and TIC

Working on a software and saas deal or business?

The first conversation is free and takes under an hour. If there's an engagement worth having you'll leave knowing its shape and its price. If there isn't, I'll say so.

Book 30 minutes brannan@anywayventures.com