ai-acquisition-premium-ppa-valuation-risk-2026

The Hidden Valuation Risk in AI-Driven M&A: Why “AI-Ready” Acquisition Premiums Are Creating PPA Nightmares in 2026

The Deal That Made Perfect Strategic Sense — Until the Auditor Asked One Question

The acquisition thesis was compelling. A strategic software company acquired a smaller, AI-native competitor for $85 million — a 14x ARR multiple that the deal team justified on three grounds: the target’s proprietary training data accumulated over six years of enterprise deployments, the AI research team of 23 engineers whose collective expertise had built a differentiated model, and the acquirer’s belief that buying rather than building this capability would compress a four-year competitive gap to zero.

The board approved it. The M&A advisors closed it. The integration team started on Day One.

Ninety days later, the acquirer’s Big Four auditor arrived with a single question that nobody on the deal team had adequately prepared for: “Under ASC 805, which of the three things you paid $85 million for are separately identifiable intangible assets — and which is goodwill?”

The answer to that question determines the opening balance sheet. The opening balance sheet determines the amortisation schedule. The amortisation schedule affects reported earnings for the next 3–15 years. And the goodwill residual — the amount left over after every identifiable intangible has been valued and separated — sits on the balance sheet until it is impaired.

For AI-driven acquisitions in 2026, that question is not being adequately answered before deals close. And the consequences of answering it incorrectly — in either direction — are showing up in financial statement restatements, auditor findings, and goodwill impairment charges that are surprising deal teams who thought the hard part was the deal, not the accounting.

This blog is for the CPA firm partners, M&A advisors, and corporate CFOs who are managing the PPA consequences of AI-driven acquisitions right now — or who will be managing them in the next 12 months as the 2026 deal surge documented in our M&A surge analysis continues to produce AI-premium transactions that have never been through a PPA before.

Why AI Acquisitions Break the Standard PPA Framework

Every PPA under ASC 805 follows the same analytical structure: identify the acquired assets and liabilities, measure each at fair value as of the acquisition date, allocate the purchase price across those measured values, and record the residual as goodwill. The framework is well-established, the methodology is codified in the AICPA Valuation Guide, and for most acquisition types — manufacturing companies, service businesses, SaaS platforms with standard customer and technology intangibles — experienced valuation teams apply it with reasonable confidence.

AI-driven acquisitions break this framework in three specific places.

Break Point 1: The Assets That Drive the Premium Are Often Not Separately Identifiable

The three things that typically justify an AI acquisition premium — proprietary data, AI research talent, and trained models — have radically different ASC 805 treatment:

In private equity acquisitions, several types of intangible assets are commonly valued, including customer relationships, trademarks, developed technology which encompasses software and proprietary processes, assembled workforce, and non-compete agreements.

But the assembled workforce — the AI research team whose collective expertise built the model — is explicitly excluded from separate recognition under ASC 805. An assembled workforce is defined in ASC 805-20-55-6 as an existing collection of employees that permits an acquirer to continue to operate an acquired business from the date of the acquisition. An assembled workforce is not an identifiable intangible asset that is to be separately recognised, and as such any value attributable to the assembled workforce is included in goodwill.

The acquirer subsumes into goodwill the value of an acquired intangible asset that is not identifiable as of the acquisition date. For example, an acquirer may attribute value to the existence of an assembled workforce, which is an existing collection of employees that permits the acquirer to continue to operate an acquired business from the acquisition date. An assembled workforce does not represent the intellectual capital of the skilled workforce — the often specialised knowledge and experience that employees of an acquiree bring to their jobs.

The practical consequence: if a significant portion of the $85 million AI acquisition price was paid for the research team, that value cannot be separately recognised on the opening balance sheet. It flows entirely into goodwill — which does not amortise under GAAP (for public companies) and is tested for impairment annually. When the team turns over — as AI talent inevitably does in a competitive market — the goodwill that represented their value does not automatically impair. But the business’s ability to maintain the model, iterate on the technology, and retain enterprise customers may deteriorate, which eventually triggers the impairment test that nobody wanted.

Break Point 2: The Proprietary Data Question Has No Standard Answer

The proprietary training dataset — accumulated over six years of enterprise deployments — is the asset that most AI deal teams believe is most defensible and most durable. The data is unique, it is not publicly available, it took years to accumulate, and it is the primary source of the model’s performance advantage.

The question of whether a proprietary dataset is a separately identifiable intangible asset under ASC 805 does not have a settled answer. The identifiability criteria require the asset to either arise from contractual or legal rights or be separable — capable of being sold, transferred, licensed, rented, or exchanged. Intangible assets that meet either of these criteria are considered identifiable and are separately recognised at fair value on the acquisition date.

For a proprietary training dataset, the analysis depends entirely on the facts:

If the data was collected under data licensing agreements with identifiable terms — and those agreements are transferable or assignable — the dataset may qualify as a contract-based intangible asset, separately recognisable under the contractual-legal criterion.

If the data was collected through user interactions with the target’s product — without specific contractual terms governing the data’s use or transferability — the separability criterion becomes the relevant test. Can the dataset be sold or licensed independently of the business? If the data is embedded in the model weights and cannot be meaningfully extracted and transferred without also transferring the model and the team that trained it, the separability test may not be met.

If the dataset cannot separately meet either criterion, its value flows to goodwill — not to a separately recognised intangible asset. For an acquisition where the data was the primary justification for the premium, this is a significant accounting problem: the asset that the acquirer most valued is the one that the accounting framework least accommodates.

Proprietary ML models, curated datasets, data privacy controls, and innovative monetisation models are now as important as revenue and growth rates. Traditional valuation frameworks while still relevant are being supplemented with AI-specific methodologies that better capture the unique value proposition of data-rich, IP-driven companies. But “supplemented with AI-specific methodologies” in a commercial context does not solve the ASC 805 identifiability problem — the accounting framework has not been updated for the AI era.

Break Point 3: The Developed Technology Useful Life Has No Historical Precedent

The AI model itself — the trained neural network, the inference engine, the fine-tuning infrastructure — is a technology-related intangible asset that can typically be separately recognised under the separability criterion. The model can be licensed, transferred, or deployed independently of the original team that built it. This is the cleanest identifiable intangible in most AI acquisitions.

The problem is the useful life. Under ASC 350, separately recognised intangible assets with finite useful lives are amortised over their useful life. The useful life is the period over which the asset is expected to contribute to the entity’s cash flows. For the AI model, this means: how long before this model is obsolete?

In a stable technology environment, the useful life of a software intangible is typically 3–7 years, based on the company’s technology refresh cycle. The analyst finds comparable acquisitions, looks at how long similar technology was maintained before replacement, and applies a useful life within the observed range.

For an AI model acquired in 2026, there is no meaningful historical precedent. The pace of AI model development — from GPT-3 to GPT-4 to Claude to Gemini, each making prior generations obsolete within 12–18 months for many applications — suggests that a 5-year or 7-year useful life for an AI model may be aggressive. But assigning a 2-year or 3-year useful life to a $30 million developed technology asset produces a $10–$15 million annual amortisation charge that materially affects reported earnings and that the deal team did not model when they justified the acquisition at 14x ARR.

AI vulnerability is a target’s exposure to AI-driven business model erosion across four axes: model dependency on third-party LLMs, weak data moat strength, agentic substitution risk, and AI talent concentration. Buyers in 2026 test these axes systematically and use findings to reprice or walk from deals. The same axes that determine AI vulnerability in diligence also determine the useful life of the developed technology in the PPA. A model with high third-party LLM dependency and low data moat strength is a model with a short useful life — because it can be replicated or replaced relatively quickly. A model with proprietary training data and proprietary fine-tuning infrastructure has a longer useful life — because replication requires assembling both the data and the infrastructure.

The useful life determination is not an administrative detail. At a $30M developed technology fair value, the difference between a 3-year and a 7-year useful life is a $6 million annual amortisation differential — which, over a 3-year period, produces an $18 million difference in reported GAAP earnings. The deal team that justified the acquisition on an EBITDA basis will not have modelled the GAAP amortisation consequence. The CFO who reports the post-acquisition earnings will not have a good answer when an analyst asks why GAAP earnings are so much lower than EBITDA.

The Six Most Common AI PPA Errors — With Specific Consequences

Error 1: Classifying the AI Team as “Assembled Workforce” and Routing to Goodwill Without Considering Employment Agreements

The assembled workforce exclusion under ASC 805 is unambiguous at the workforce level — the AI research team as a collective cannot be separately recognised. But individual employment agreements are a different matter. Although individual employees may have employment agreements with the acquiree, which may, at least theoretically, be separately recognised and measured, the entire assembled workforce does not have such a contract.

For AI acquisitions where the target has executed retention agreements with key researchers — multi-year employment contracts with substantial vesting schedules specifically tied to the acquisition — those contracts may qualify as separately identifiable intangible assets under the contractual-legal criterion, because they arise from contractual rights. The fair value of these agreements — the economic benefit to the acquirer of having the key personnel contractually obligated to remain — can be separately recognised and amortised over the contract term.

The error is treating all employee-related value as undifferentiated assembled workforce goodwill, without analysing whether specific individual employment agreements meet the contractual-legal criterion. For an AI acquisition where three lead researchers each have $3–$5 million multi-year retention packages tied to the close, the total contractual value of those agreements may be $9–$15 million — a separately recognisable intangible that reduces the goodwill balance and creates an amortising asset.

The consequence of the error: Goodwill is overstated by $9–$15 million. The annual goodwill impairment test carries a higher threshold. When those key researchers leave — as they frequently do within 24–36 months of an acquisition — the goodwill that represented their value does not impair automatically, but the business’s performance often deteriorates, eventually triggering the impairment test from the revenue and cash flow side rather than the carrying value side.

Error 2: The Data Asset Classification Mistake — Recognising Proprietary Data as Developed Technology

The most common single classification error in AI PPAs is treating the proprietary dataset and the trained model as a single developed technology intangible — recognising both under one line item, assigning a single fair value to the combined asset, and assigning a single useful life.

The problem: the dataset and the model are conceptually distinct assets with potentially different identifiability analysis and materially different useful lives.

The trained model — the neural network weights, the inference infrastructure, the deployment layer — has a useful life that is primarily determined by the pace of AI model development and the competitive pressure to upgrade.

The proprietary training dataset — the curated, labelled enterprise interaction data accumulated over six years — may have a substantially longer useful life, because the data itself does not become obsolete at the same rate as the model. The data can be used to train the next generation of models. The data is the durable asset; the model is the current expression of that data.

Combining them into a single developed technology intangible forces both to the same useful life — typically the shorter one (the model’s obsolescence horizon) — which accelerates the amortisation of the data asset and potentially understates its carrying value over time.

The correct approach is to analyse each component separately: can the dataset meet the identifiability criteria independently of the model? If yes, it is a separately recognised intangible with its own fair value and useful life. If no, it flows to goodwill. The model is separately recognised as developed technology with its own fair value and useful life.

The consequence of the error: Amortisation is accelerated on the data component, reducing reported GAAP earnings more aggressively than required. The impairment test for the intangible considers the combined asset’s recoverable value rather than each component’s contribution — which may trigger impairment of the combined asset even when the data component retains full value.

Error 3: Using the Cost Approach for the AI Model Without Adjusting for Technological Obsolescence

The relief-from-royalty method is the standard approach for valuing developed technology intangibles — it calculates the fair value as the present value of the royalties that would be paid to license the technology if it were not owned. For AI models, the relief-from-royalty method requires a royalty rate — which is typically benchmarked from observed technology licensing agreements.

Many AI PPA valuation teams default to the cost approach when comparable royalty rate data is thin — calculating the fair value of the AI model as the cost to reproduce it from scratch. The reproduction cost includes the compute cost of training the model, the data collection and labelling costs, the engineering hours to build the training infrastructure, and the fine-tuning and evaluation costs.

The reproduction cost approach is a valid method under the AICPA Valuation Guide for technology assets when the income and market approaches are not practicable. But it requires an obsolescence adjustment — a reduction in the reproduction cost to reflect the fact that the existing model may already be partially or fully obsolete relative to what could be built with current technology.

For an AI model trained in 2023 and being valued in a 2026 PPA, the technological progress in the intervening three years is substantial. A model that cost $15 million to train in 2023 might cost $3 million to replicate in 2026 with current technology — not because the model has become less valuable, but because the compute costs have declined and the training methodologies have improved. The fair value of the model is not its 2023 reproduction cost — it is its 2026 replacement cost adjusted for functional and technological obsolescence.

Failing to apply the obsolescence adjustment overstates the developed technology fair value, understates goodwill, and understates the annual amortisation required to bring the carrying value down to fair value over the useful life.

The consequence of the error: The developed technology intangible is overstated. Annual impairment testing — both the annual goodwill impairment test and any triggering events for the finite-lived intangible — is more likely to produce a charge because the carrying value was set too high relative to the asset’s true recoverable amount.

Error 4: The In-Process Research and Development Trap

For AI companies actively developing next-generation models at the time of acquisition — which describes virtually every active AI research company — the in-process research and development (IPR&D) classification is a critical PPA question that most deal teams handle incorrectly.

Under ASC 805, acquired in-process research and development is recognised as a separately identifiable intangible asset at fair value — even if the underlying research project has not yet reached technological feasibility. IPR&D is initially classified as an indefinite-lived intangible asset — it is not amortised, but is tested for impairment annually — until the project is either completed (at which point it becomes a finite-lived developed technology intangible) or abandoned (at which point it is written off).

For an AI acquisition where the target has an active project developing the next version of its core model — with a defined roadmap, a dedicated team, and documented milestones — that project is likely IPR&D. Its fair value is separately recognised on the opening balance sheet and not amortised until completion.

The error: treating the next-generation model development project as part of the general goodwill allocation, without separately identifying and valuing the specific IPR&D project. This understates the separately recognised intangibles, overstates goodwill, and misclassifies an asset that should be tested for impairment annually as an asset that never gets tested independently.

The consequence: when the IPR&D project is completed — and the next-generation model is deployed — the fair value of the completed technology should be transferred from IPR&D to developed technology and begin amortising. If the project was never separately recognised, this reclassification cannot occur, and the completed technology is buried in goodwill rather than amortised appropriately.

The consequence of the error: Goodwill is overstated by the fair value of the IPR&D project. The indefinite-lived intangible annual impairment test — which would specifically assess whether the IPR&D project has value — is never conducted. When the project is abandoned or pivoted, there is no IPR&D write-off — the loss is effectively deferred until a goodwill impairment test eventually captures it.

Error 5: The WARA Reconciliation That Never Gets Done — and Why It Matters Most for AI Deals

The Weighted Average Return on Assets reconciliation — the cross-check that verifies the PPA’s internal consistency by confirming that the blended return on all identified assets equals the overall WACC used in the acquisition analysis — is the single most important quality control step in any PPA. Test the results for reasonableness — by calculating the weighted average return on assets (WARA) for the post-PPA balance sheet. Logically, the unallocated goodwill should have the highest expected returns.

For AI acquisitions, the WARA reconciliation is more consequential than for any other deal type — because the asset mix is unconventional, the returns on individual assets are more speculative, and the proportion of goodwill is typically higher than in non-AI acquisitions.

The WARA reconciliation for an AI PPA must assign a required rate of return to each identified asset category: tangible assets (at a rate reflecting asset depreciation), customer relationships (at a rate reflecting customer retention risk), developed technology (at a rate reflecting technology obsolescence risk), IPR&D (at the highest required return, reflecting the speculative nature of the research outcome), and goodwill (at a rate above all other categories, reflecting the residual, non-identifiable nature of the value).

For most AI acquisitions, the developed technology and IPR&D components carry the highest required returns of any identified intangible — because they face the most rapid obsolescence risk. If the WARA reconciliation assigns modest required returns to these assets (treating them like stable customer relationships rather than high-risk technology assets), the WARA will not reconcile to the acquisition WACC — and the PPA is internally inconsistent.

The consequence of not doing the WARA: The auditor will either perform the reconciliation themselves or require the valuation team to perform it before signing off on the opening balance sheet. A WARA reconciliation failure discovered during audit review delays the quarterly filing, requires revisions to the PPA that may be material, and in some cases requires a restatement of provisional amounts that were already filed.

Error 6: Goodwill Allocation to the Wrong Reporting Unit — and the Impairment Test That Fails as a Result

Under ASC 350, goodwill must be allocated to the reporting unit that is expected to benefit from the combination. For AI acquisitions where the target’s technology is being integrated across multiple business lines of the acquirer — the AI model will be deployed in the acquirer’s CRM product, its analytics product, and its professional services offering — the question of which reporting unit receives the goodwill is not trivial.

If the goodwill is allocated to a single reporting unit — the product line that was the primary deal rationale — and the benefit of the AI capability is actually spread across multiple reporting units, the goodwill impairment test may fail for the designated reporting unit even when the AI investment is generating value elsewhere in the business.

A proper allocation of the purchase price paid and resulting goodwill to cash-generating units is vital and can prevent subsequent impairments. In the TMT sector, goodwill is a common feature, since it reflects expectations of future synergies and growth. A pre-PPA analysis identifies assets, synergies, and financial impacts pre-transaction, and avoids post-transaction surprises.

For AI acquisitions where the value is inherently cross-functional, the goodwill allocation methodology must follow the cash flows — allocating goodwill to the reporting units that will actually receive the economic benefits of the AI capability, in proportion to the expected benefit distribution. A goodwill allocation methodology that was developed for a focused product acquisition does not translate directly to a cross-functional AI capability acquisition.

The Emerging Framework — How to Value AI Intangibles Correctly

Given the absence of established precedent and the specific challenges described above, here is the emerging best practice framework for AI PPA valuation that is withstanding Big Four auditor scrutiny in 2026.

Step 1: The Pre-Close Asset Identification Session

Before the acquisition closes — ideally during the diligence period — convene a structured session with the deal team, the target’s management, and the valuation team to systematically map every value driver in the deal thesis to an ASC 805 asset category.

For each value driver, ask three questions: Does it arise from contractual or legal rights? Can it be separated and transferred independently? If neither — is it goodwill or an assembled workforce component?

The output of this session is a preliminary asset identification memo — documenting the asset mapping, the identifiability analysis for each component, and the preliminary classification. This memo becomes the foundation of the PPA and prevents the post-close discovery problem that occurs when the auditor asks the identifiability question for the first time.

Vertical AI companies building AI-native solutions for specific industries command some of the strongest multiples in the market. The logic is straightforward: these companies combine domain expertise, industry-specific data sets, and regulatory knowledge in ways that horizontal tools cannot easily replicate. That logic — domain expertise, proprietary data, regulatory knowledge — must be translated into specific ASC 805 asset categories before the acquisition closes. Domain expertise = goodwill (assembled workforce component). Proprietary data = potentially separable intangible if licensing agreements meet contractual-legal criterion, otherwise goodwill. Regulatory knowledge = potentially licence-based intangible if embodied in specific regulatory approvals or certifications.

Step 2: The Data Asset Due Diligence

For the proprietary dataset specifically — the asset that most frequently creates the identifiability question — a focused due diligence inquiry should be conducted pre-close:

How was the data collected? Through user interactions with the product (potentially non-contractual), through licensing agreements with third parties (potentially contractual), through paid data collection programmes (potentially contractual), or through proprietary research processes (potentially separable)?

What are the data rights? Does the target own the data outright, license it from third parties, or hold it subject to user privacy agreements that limit transferability?

Can the data be separated from the model? If the data is stored in a transferable format — independent of the model weights — it is more likely to meet the separability criterion than if it is embedded in the model in a way that cannot be extracted.

Are there existing data licensing agreements? If the target has licensed its data to third parties, those licensing arrangements are evidence of separability and provide a basis for the relief-from-royalty valuation approach.

The answers to these questions determine whether the data asset is separately recognisable — and therefore whether it reduces goodwill or contributes to it.

Step 3: The AI Model Useful Life Analysis

For the developed technology intangible — the trained model — the useful life determination requires a specific analysis that is distinct from the standard technology useful life methodology.

The standard methodology — observe how long comparable companies have maintained their technology before replacement — is not applicable to AI models, because there are no comparable companies with sufficient historical data. The AI era is too recent to produce the empirical useful life data that the standard methodology requires.

The alternative approach — a functional analysis of the specific model’s obsolescence risk — must address the four AI vulnerability axes described in Valutico’s 2026 framework: model dependency on third-party LLMs, weak data moat strength, agentic substitution risk, and AI talent concentration.

A model with high third-party LLM dependency (its performance is primarily a function of the underlying foundation model, not the target’s proprietary fine-tuning) has a short useful life — when the foundation model is superseded, the target’s model is superseded with it. This suggests a 2–3 year useful life.

A model with a strong proprietary data moat (its performance advantage comes from data that cannot be easily replicated) and low agentic substitution risk (it performs a task that AI agents cannot yet handle end-to-end) has a longer useful life. This might support a 5–7 year useful life.

The useful life analysis must be documented specifically for the acquired model — not borrowed from general technology useful life ranges. The documentation must survive an auditor who will ask: “How did you conclude 5 years when the AI model that preceded this one was superseded in 18 months?”

Step 4: The IPR&D Identification and Fair Value

For any active research project at the acquired company — a next-generation model in development, a new application domain being built on the core technology, a proprietary inference optimisation project — a formal IPR&D identification and fair value analysis is required.

The fair value of IPR&D is typically estimated using the income approach — the present value of the cash flows expected from the project upon completion, probability-weighted for the risk that the project will not reach completion. The probability of completion is the most important and most judgment-intensive input in the IPR&D valuation.

For an AI research project at an early stage of development — a next-generation model that has a defined architecture but has not yet completed training — a conservative probability of completion (40–60%) is appropriate. For a project that has completed the research phase and is in the implementation and testing stage, a higher probability (70–85%) may be supportable.

The IPR&D valuation must be documented with specific reference to the project’s current stage, the defined milestones, the remaining development timeline, and the basis for the probability of completion estimate. This documentation is the starting point for the annual indefinite-lived intangible impairment test.

Step 5: The Assembled Workforce Contributory Asset Charge

Even though the assembled workforce is not a separately recognised intangible asset, it functions as a contributory asset in the MPEEM valuation of the customer relationships, the developed technology, and any other earning intangibles in the PPA. The contributory asset charge for the assembled workforce — the return on and of the workforce required to generate the earnings attributed to other intangibles — must be calculated and deducted from the earnings stream in each MPEEM analysis.

For AI acquisitions, the assembled workforce is typically the largest contributing asset in the MPEEM — the research team’s ongoing contribution to model maintenance, iteration, and customer deployment is the primary driver of the intangible’s cash flow generation. An MPEEM that understates the assembled workforce contributory asset charge overstates the fair value of the customer relationships and developed technology intangibles — and understates goodwill.

The assembled workforce contributory asset charge requires an estimate of the fair value of the workforce — which, paradoxically, requires valuing an asset that is not separately recognised in order to charge for its contribution to assets that are. This is the analytical complexity that makes AI PPAs materially more difficult than standard technology company PPAs.

The Goodwill Impairment Timeline — When the AI Premium Becomes a Problem

The goodwill generated in a 2026 AI acquisition will not impair in 2026. The acquisition is fresh, the integration is underway, and the reporting unit’s fair value — reflecting the acquisition-date optimism about AI synergies — still supports the carrying value. The impairment risk accumulates over the 18–36 months following close.

The specific scenarios that trigger impairment for AI acquisition goodwill:

Scenario 1: Key talent departure. The AI research team that the acquirer paid a premium for begins to leave — attracted by competing opportunities, frustrated by integration friction, or simply completing their vesting schedules and exercising their options. The assembled workforce value — which was subsumed into goodwill — begins to erode. The model’s performance stagnates without the talent to iterate it. Revenue from AI-dependent products grows more slowly than the deal thesis projected. The goodwill impairment test catches this through the DCF: lower-than-projected revenue growth, updated projections, and a fair value that falls below carrying value.

Scenario 2: Technology supersession. A foundation model provider — OpenAI, Anthropic, Google — releases a capability update that renders the acquired model’s differentiation obsolete. Enterprise customers that were paying a premium for the acquired company’s AI capability find that the same capability is now available through their existing foundation model licences at effectively zero marginal cost. The revenue base erodes faster than the DCF model projected. The goodwill impairment test catches this through the market approach: comparable AI company multiples have compressed because the category has been commoditised, and the market approach indication falls below carrying value.

The effectiveness of goodwill impairment testing has been questioned, with concerns raised about insufficient goodwill impairment provisions. Data assets affect the provision for goodwill impairment — companies with stronger data asset positions show more stable goodwill impairment testing outcomes.

Scenario 3: Integration failure. The AI capability that was acquired for its cross-functional deployment potential proves harder to integrate than anticipated. The acquirer’s existing product architecture is incompatible with the target’s inference infrastructure. The deployment timeline extends from 12 months to 36 months. The synergies that justified the premium are delayed, and the goodwill impairment test in Year 2 post-close finds that the present value of the expected synergies — now pushed two years further out — no longer supports the carrying value.

For each of these scenarios, the impairment loss is a direct function of how much goodwill was created at the PPA stage. An AI PPA that correctly identified and valued the data asset as a separately recognisable intangible — reducing goodwill by $15–20 million — creates a meaningfully lower goodwill carrying value and meaningfully more headroom in the impairment test. An AI PPA that routed everything except the most obvious technology intangible to goodwill creates the maximum impairment exposure.

What CFOs and CPA Firms Should Do Before the Next AI Deal Closes

The AI PPA problem is not solvable after the fact. Once the acquisition closes and the opening balance sheet is filed, the PPA is set — revisions during the measurement period are limited to new information about conditions that existed at the acquisition date, not changes in judgment about the methodology. The time to get the AI PPA right is before the acquisition closes.

For Corporate CFOs

Engage the valuation team at signing, not at close. The pre-close asset identification session described above — mapping deal thesis value drivers to ASC 805 asset categories — is a 2-day engagement that prevents a 60-day post-close remediation. The cost differential is not marginal; it is the difference between a PPA that the auditor accepts at first review and one that requires three rounds of revision.

Budget for the AI model useful life analysis as a separate engagement component. Do not allow the valuation team to default to a 5-year or 7-year useful life without producing a specific functional analysis of the acquired model’s obsolescence risk. The useful life is the most consequential assumption in the AI PPA — and the one most likely to be challenged by the auditor.

Model the GAAP amortisation before the deal closes. The deal team that justified the acquisition on an EBITDA basis must understand the GAAP earnings impact of the PPA amortisation schedule — including the amortisation of developed technology, customer relationships, IPR&D upon completion, and any separately recognised employment agreement intangibles. If the GAAP earnings impact was not modelled before the acquisition was approved, the CFO will be explaining an unexpected earnings decline to the board for years after the close.

For CPA Firms

The AI PPA is not a standard technology company PPA with a different cover page. The assembled workforce exclusion, the data asset identifiability question, the AI model useful life determination, and the IPR&D classification are all novel methodological questions that require specific expertise — not the standard PPA workflow applied to a different sector.

For CPA firms whose AI acquisition clients are generating PPA requirements in 2026, the white-label outsourcing model — engaging Synpact’s valuation team to produce the intangible asset valuations, the WARA reconciliation, and the useful life documentation — is the capacity and expertise solution. The specific methodological questions in an AI PPA require valuation professionals who have worked through the assembled workforce analysis, the data asset separability question, and the AI model useful life determination for prior engagements — not ones who are encountering these questions for the first time in the context of a live audit.

→ Submit an AI Acquisition PPA Brief — Pre-Close Asset Identification and Opening Balance Sheet in 14 Business Days

The Regulatory and Standard-Setting Horizon — What Is Coming

The PPA methodology problems created by AI acquisitions are not going unnoticed by the accounting standard-setters.

FASB has an active agenda item on software cost accounting — including the treatment of AI model development costs under ASC 350-40 — that has direct implications for PPA methodology for acquired AI assets. The current guidance was written for traditional software development, not for the iterative, data-driven development process used in AI model training. The FASB project may produce guidance that specifically addresses the amortisation period and useful life determination for acquired AI models — but that guidance is not yet final, and the acquisitions being done in 2026 must be accounted for under the existing framework.

The AICPA’s Valuation Guide — the primary methodology reference for PPA valuation professionals — does not yet contain specific guidance for AI asset valuation. The guide was last significantly updated before the current AI era. The methodologies described in this blog — the data asset separability analysis, the AI model useful life functional framework, the IPR&D probability of completion methodology for AI research projects — represent current practitioner consensus, not codified guidance.

This gap between deal activity and accounting guidance is exactly the situation that creates the most post-close accounting problems. The acquisitions being done today, at AI premiums that are unprecedented in their composition and magnitude, will be the case studies in FASB’s eventual guidance — and some of them will be restatements.

The CFO and CPA firm partner who builds the methodological rigour described in this blog into their AI PPA process today is building the position that survives the eventual regulatory scrutiny. The one who routes everything to goodwill because the guidance is unclear is building the restatement risk.

The AI Premium Was Justified. The PPA Must Justify It Too.

The strategic rationale for AI-driven acquisitions in 2026 is real. Proprietary training data is genuinely valuable and genuinely difficult to replicate. AI research talent is genuinely scarce and genuinely consequential. Trained models with enterprise deployment records provide genuine competitive advantages that would take years to build organically.

The accounting framework does not automatically accommodate that rationale. The assembled workforce is not separately recognisable regardless of how valuable the researchers are. The proprietary dataset may or may not be separately recognisable depending on facts that most deal teams have not analysed. The trained model’s useful life may be 2 years or 7 years — and the difference is worth tens of millions of dollars in amortisation over the holding period.

Getting the AI PPA right requires engaging the methodology questions before the acquisition closes, not after. It requires a valuation team that has worked through the assembled workforce exclusion, the data asset identifiability analysis, the IPR&D classification, and the AI model useful life determination — not one encountering these questions for the first time in the context of a Big Four audit review.

The acquisitions being announced today will be filed on opening balance sheets within 90 days. The methodology choices made in the next 60 days will affect the acquirer’s financial statements for the next decade. Get them right the first time.

→ Submit an AI Acquisition PPA Brief or Book a Pre-Close Methodology Consultation — 14 Business Day Delivery

Related Reading on Synpact Blog: