# “Service as Software”: when the ability to deploy becomes the product

> Reference HTML page: https://leaid.ai/en/articles/service-as-software-capacite-deployer
> Language: en
> Author: Mamadou Waggeh, Founder of Leaid
> Date: 2026-08-31

PwC and Harvey, KPMG and ContractPodAi, Deloitte and Legora then Ironclad, OpenAI and its Frontier Alliances: in five years, alliances between AI vendors and consulting firms have become the norm in the legal market. Our analysis of some fifteen deals concluded between 2021 and August 2026 shows that these are not distribution agreements, but a new production architecture: the software produces, the consultancy governs, the client buys a result. This model will not lead to the full automation of law — it recomposes the service into five layers, with support becoming a productive component. Provided the risks that the press releases gloss over are addressed.

## Introduction

The first phase of generative AI applied to law asked a simple question: would legal professionals use these tools? Three years on, that question is settled — and the problem has moved. The issue is no longer giving access to a tool, but turning a technical capability into lasting results — embedded in processes, compatible with confidentiality, verifiable. Yet the tools are everywhere; the results far less so. In 2025, 88% of respondents to McKinsey’s global survey reported regular use of AI in at least one function, but only 39% reported any effect, however limited, on EBIT. The same year, BCG estimated that 5% of companies were capturing significant value at scale, while 60% saw only modest gains or none at all. The technology keeps advancing; the transformation of the organisations that must put it to work does not keep pace.

The legal market is giving this problem an increasingly visible answer: the alliance. PwC teamed up with Harvey as early as March 2023, then with ContractPodAi. KPMG built a managed legal services offering on the same platform. Deloitte Legal is supporting Legora from pilot to scale, and Deloitte has since allied with Ironclad across the full contract lifecycle. In France, Lefebvre Dalloz, Xelya and Anomia combine reliable content, practice software and continuous training. Model providers are following the same trajectory: Frontier Alliances and a deployment company at OpenAI, centres of excellence and large-scale certifications at Anthropic, a partner ecosystem around Gemini Enterprise at Google Cloud.

Our reading: these alliances are not a sales channel but a production infrastructure. As models become commoditised and platform features converge, the advantage shifts to the ability to choose the right problems, integrate with existing systems, organise human oversight, drive adoption, measure — and take on a growing share of the delivered result. The partnership becomes part of the product. That is the point of this analysis: the ability to deploy is itself becoming the product.

## 1. “Service as Software”: what is being sold changes

SaaS rests on a clear separation. The vendor supplies a tool accessible by subscription, billed per seat, by volume or by feature; the client remains responsible for organising the teams and producing the result. Traditional consulting follows the opposite logic: the client buys expertise, effort and judgement, priced by time, seniority or fixed fee — technology improves the firm’s productivity but remains invisible in the offering. Between these two poles sit managed services, where a provider operates a process on a lasting basis, and “software plus services”, where support eases deployment without changing what is sold.

The term “Service as Software” marks a sharper break: the provider no longer sells only the tool with which the work is done; it sells the work done — or a bounded result — a substantial share of which is produced by software. Foundation Capital describes this shift as a transfer of responsibility: the client no longer has to use the platform to reach its goal; the provider answers for the delivered result. Sequoia contrasts the copilot, which sells the tool to the professional, with the autopilot, which sells the work itself to the client. Bessemer observes that AI now reaches the production layer of the service, no longer just its interface. Born of investment literature more than academic research, the notion matters for four shifts: what is sold (from access to result), operational responsibility (from the client to the provider or the consortium), pricing (from seats and hours to matters, volume, value) and growth (less and less proportional to headcount).

In law, the full autopilot runs up against the very nature of the work. A standard clause, a consistency check, an extraction of obligations can be turned into repeatable outputs; a litigation strategy, a risk assessment, a sensitive negotiation or advice engaging professional liability resist it. The plausible model is therefore not the erasure of human service, but its recomposition: AI industrialises the repeatable segments; the firm defines the rules, provides context, supervises the exceptions, owns the judgement and guarantees the relationship. Legal “Service as Software” is a hybrid system of governed production — not autonomous software.

## 2. Law, prime terrain under constraints

A large share of legal work rests on documents, rules, precedents and repetitive processes: reading, comparing, drafting and tracking contracts; mapping regulations and translating them into operational obligations; structuring bodies of evidence and chronologies. These characteristics make many sub-tasks automatable: law thus ranks [among the first vertical markets targeted by generative models](/en/articles/geants-ia-marche-droit-verticale).

Efficiency there, however, is neither uniform nor synonymous with quality. In the randomised trial by Choi, Monahan and Schwarcz published in the Minnesota Law Review (2024), time savings on tasks typical of junior lawyers range from 11.8% to 32.1%, but the average improvement in quality remains small and uneven. An experiment by Nielsen and co-authors (Journal of Empirical Legal Studies, 2024) shows that AI-generated highlighting cuts processing time by 30% without degrading quality — while a simple summary adds nothing. The precise design of the assistance matters as much as the general power of the model.

The study of 758 BCG consultants by Dell’Acqua and co-authors pins down the limit. For tasks inside the model’s capability frontier, users complete 12.2% more tasks, work 25.1% faster and produce higher-quality answers; for a task just beyond it, they are 19 percentage points less likely to reach the right conclusion. This frontier is “jagged”: two apparently similar tasks can sit on opposite sides of the reliability zone.

In law, this jaggedness is not a mere production defect: it is a professional risk. An error can compromise proceedings, expose confidential information, create a conflict of interest or distort the client’s decision. The CCBE reminds practitioners that the use of generative AI engages confidentiality, competence, independence, transparency towards the client and the verification of outputs; the ABA insists on supervision and reasonable billing. The need for support is therefore not a temporary digital-skills deficit: it is structural. Someone has to determine where automation is legitimate, how it is controlled and who answers for the result.

## 3. 2021-2026: three generations of alliances

The recent history of these partnerships reads as three generations. The first predates the public rise of generative AI and concerns contract systems: when Accenture deployed Icertis in its own operations — nearly 3,000 professionals across 46 countries — before formalising a joint offering, the firm was not bringing a contacts book; it was bringing proof of use. This “client zero” pattern would become the norm. The second generation, that of generative assistants, sees software enter the very production of consulting: mass equipping of internal teams at PwC with Harvey, a consulting service powered by Leah, managed legal services at KPMG — the software is no longer installed, it is operated, under human supervision.

The third generation, visible since 2025, targets the complete operating model — reshaping relationships, skills, data and workflows — and the model providers are generalising the movement. OpenAI’s Frontier Alliances entrust firms with strategy, workflow redesign, integration and change management; the creation of the OpenAI Deployment Company, with the acquisition of Tomoro, internalises deployment engineering; Anthropic is building centres of excellence and large-scale certifications with Deloitte then Accenture — 15,000 then 30,000 professionals targeted, engineers embedded with clients. Summize’s acquisition of teams and assets from the consulting firm InnoLaw Group says the same thing in reverse: when the deployment layer becomes strategic, it is brought in-house.

### First generation · 2021-2022 — Deploy in-house, prove through use

- **2021 · Accenture × Icertis**: internal deployment to 3,000 professionals, then a joint CLM integration offering — the “client zero” precedes the market launch.
- **2022 · Bain × OpenAI**: the first major consulting × model provider alliance, expanded in 2024 into a centre of excellence and sector solutions.

### Second generation · 2023-2024 — Co-producing the service

- **2023 · PwC × Harvey**: equipping 4,000 professionals in more than a hundred countries, commercialisation, customisation — the firm becomes user, integrator and co-designer.
- **2024 · PwC × ContractPodAi**: a new consulting service powered by Leah — software enters the production of consulting.
- **2024 · KPMG × ContractPodAi**: managed legal services with human supervision — from licence to operated process.
- **2024 · PwC Germany × Aleph Alpha**: the creance.ai joint venture, initially focused on DORA compliance — joint creation of a vertical asset.

### Third generation · 2025-2026 — Transforming the operating model

- **2025 · Deloitte Legal × Legora**: from pilot to lasting adoption, with preconfigured workflows — the operating model becomes the object of the engagement.
- **2025 · Thomson Reuters × Icertis × Accenture**: content, contract platform, orchestration — a multi-player value chain.
- **2025 · Anthropic × Deloitte, then × Accenture**: centres of excellence, tens of thousands of certified professionals, engineers embedded with clients.
- **2026 · Deloitte × Ironclad**: end-to-end agentic contract lifecycle — agents, controls, data, governance, usage.
- **2026 · OpenAI — Frontier Alliances, then Deployment Company**: deployment capability becomes a function of the model provider itself.
- **2026 · Summize × InnoLaw Group**: acquisition of consulting teams and assets — adoption is internalised as a differentiator.

*Source: official announcements by the organisations cited. The descriptions refer to announced structures, not audited performance.*

## 4. A five-layer value architecture

These alliances assemble five complementary layers. Intelligence capability, supplied by general-purpose or specialised models. Legal knowledge — sources, taxonomies, precedents, document templates, business rules — which grounds the answers. Workflow, that is, the system in which the work is triggered, enriched, validated, traced and handed over. Adoption and transformation: diagnosis, integration, training, changing roles, measuring value. Responsibility, finally, where a professional or a legal function owns oversight, escalation, the relationship and the final decision.

A legaltech can master the first three layers without possessing the legitimacy, the operational presence or the transformation capability of the last two; a firm can hold the client relationship and the domain expertise without controlling the technology or the content. The alliance reduces this discontinuity: the client buys a coherent trajectory rather than a set of components it would have to assemble itself.

Three degrees of integration — the equipped firm, which builds its own competence on the tool; the integrator firm, with methods, certifications and teams dedicated to client deployments; co-production, where the partners build a joint offering or managed service and share part of the economic responsibility for the result. The joint venture and the acquisition are its equity variants: they appear when the learning, the usage data and the methods become too strategic to remain at the periphery.

The “client zero” is central to this architecture. By deploying the solution internally first, the firm turns an abstract alliance into proof of use: product limits, playbooks, friction points, references. This experience reduces the information asymmetry vis-à-vis the client and gives the vendor a demanding learning ground. It is credible, however, only if the results, the failures and the conditions of success are documented — not converted into marketing copy.

## 5. Support is a strong market expectation

The idea that a good tool naturally ends up adopted confuses availability with transformation. Thomson Reuters’ Future of Professionals Report 2026 illustrates the point: 74% of the professionals surveyed use AI several times a week, but 41% lack a tool designed for professional work and grounded in verified content, and 34% report using tools not authorised by their organisation. Above all, where an identified AI strategy exists, 66% consider that AI meets or exceeds value-creation expectations, against 22% in organisations without an active strategy. This is a self-reported correlation, not causal proof — but the gap speaks to the weight of the organisational framework.

Value is built through operations that technology does not perform on its own. Translating an economic objective into precise use cases: cutting contracting lead time is an objective; “deploying a chatbot” is not. [Redesigning the workflow](/en/articles/ia-juridique-maitrise-du-workflow) — inputs, automated steps, controls, exceptions, responsibilities: McKinsey observes that workflow redesign is, of the attributes studied, the one most strongly associated with EBIT impact — while only 21% of organisations using generative AI reported in early 2025 that they had fundamentally redesigned certain processes. Preparing the data and the integrations, because a legal AI cut off from the DMS, the CLM, the CRM and the validation circuits produces no lasting value. Governing: authorised information, verification, audit logs, escalation, incidents. Managing change: role-based training, communities of practice, office hours, incentives. Measuring, finally: useful adoption, cycle times, quality, coverage, risks avoided.

Two recent studies measure the effect of this support directly. In Chien and Kim’s study of legal aid professionals (Loyola of Los Angeles Law Review, 2025), a randomly selected subgroup received, on top of access to the tools, “concierge”-style support — shared use cases, office hours, assistance. It achieved significantly better results on self-reported productivity, satisfaction, perceived quality and frequency of use; in total, 90% of pilot participants reported a productivity gain. At the scale of an entire system, the 2026 experiment with 1,559 judges across 118 Pakistani jurisdictions by Ash, Mehmood and Goessmann reaches the same conclusion: the assistant paired with targeted training produces stronger and more persistent adoption than generic training; the most exposed districts resolve more cases — an estimated 6.3% increase at the median level of exposure — with no degradation in the quality indicators used, and usage shifts towards the tasks where the model is most reliable.

Consulting thus performs a conversion function: it turns a probabilistic capability into a governed system of work. This function directly produces the quality of the result — it reduces unsuitable uses, accelerates learning, organises responsibility, makes the gains repeatable. In a “Service as Software” model, it must be designed, priced and measured as a component of the offering, not as an ancillary cost meant to compensate for the product’s shortcomings.

## 6. Legaltech, firm, client: what each gains, and on what conditions

For the legaltech, the benefit is first commercial: the firm brings trust, sector knowledge, access to decision-makers and a local presence that are hard to replicate — in regulated environments, this legitimacy reduces perceived risk and shortens the time from signature to value. It is then informational: engagements produce data on errors, exceptions, control points and the real conditions of adoption which feed, within an appropriate contractual and ethical framework, the evaluations, the workflow libraries and the product. Bessemer sees here the advantage of AI-native services: every verifiable unit of work is a learning signal. In law, this advantage rests on capitalising methods, error categories and escalation rules — not on appropriating clients’ confidential data. The condition: preserving the software economics. Too much customisation or too broad a responsibility turns the vendor into a low-leverage services company; the right target is not to sell more days, but to convert each engagement into reusable components — connectors, playbooks, evaluations.

For the firm, the alliance is a path to productisation: selling a diagnostic, a preconfigured workflow, continuous contract review or a compliance capability rather than a succession of hours — keeping the judgement and the relationship, increasing the number of matters served per professional. The commercial transition is the hard part. Simon-Kucher’s 2026 study of 182 executives of services companies shows that 90% of companies are active in AI or plan to be, but that monetisation is moving more slowly than production: 30% of the most advanced companies use subscription pricing, against 14% of the sample, and only 16% have a fully differentiated allocation of technology costs. The conflict with time-based billing is real: when twenty hours become two, mechanically keeping the hourly model either cuts revenue or captures the entire gain without making it visible to the client — neither option is sustainable. The ABA reminds lawyers, moreover, that fees must remain consistent with the time actually spent — and that a lawyer does not bill the client for learning the tool.

For the client, the main gain is reduced coordination risk: instead of separately selecting model, platform, integrator, trainer and validator, it buys an orchestrated trajectory — provided the contract leaves no grey areas — and can align the price with a useful unit: contract processed, matter analysed, lead time cut, level of coverage. The most credible pricing will be hybrid:

That is where the essential shift lies: the unit of value is no longer the hour or the seat, but reliable processing capacity, coverage, decision speed and risk control. Part of the productivity gain must be shared with the client; the rest funds oversight, security, improvement and responsibility.

## 7. Eight principles for designing an alliance that holds

Eight design principles emerge from this analysis — a working framework, not a recipe: each alliance will have to translate them into its own context and governance.

- **Start from a business result, never from a tool**: contracting lead time, share of matters left unhandled, cost of a review, error frequency — measurable starting points, which guard against the proliferation of pilots with no owner and no value trajectory.
- **Break the workflow down and make the judgement gates explicit**: for each sub-task, decide whether AI proposes, executes, checks or assists; define the uncertainty thresholds and escalation triggers before deployment. This is the operational translation of the jagged technological frontier.
- **Make the firm a “client zero” with a duty of proof**: internal deployment must produce reference data, evaluation protocols and documented incidents — and distinguish the technical gain, the adoption gain and the economic gain.
- **Contract for legible end-to-end responsibility**: an operational RACI covering model provider, legaltech, firm, in-house legal department, IT and users; output validation, service levels, model changes, incident procedures.
- **Build an adoption system, not a training course**: role-based pathways, a use-case library, office hours, ambassadors, indicators of useful usage. Article 4 of the EU AI Act, applicable since February 2025, makes AI literacy an organisational obligation.
- **Preserve control over data and reversibility**: location, secondary uses, subcontractors, rights over configurations, export mechanisms; separate data, content, workflow and model as far as possible; test reversibility during the contract, not at its end.
- **Align the price with an observable unit of value**: subscription for availability, usage for volume, fixed fee for a defined output, performance for a measurable result — without stacking licence, consulting days and integration fees without justification.
- **Govern the partnership as a living product**: a shared backlog, versions, recurring evaluations, documented decisions on model or scope changes; and an allocation of assets — shared, specific to each partner, belonging to the client — decided before the value crystallises in the prompts, the tests and the workflows.

## 8. Why France and Europe are the natural terrain for this model

France and Europe combine the conditions that make this model necessary: structured regulated professions, representative institutions, [strong requirements of sovereignty and confidentiality](/en/articles/ia-generative-secteurs-regules-confiance), rich legal corpora — and a fabric of firms, notarial offices and SMEs that do not always have in-house teams able to integrate AI on their own. These characteristics increase the need for trusted intermediation and make a strategy based solely on the horizontal diffusion of a general-purpose model hard to credit.

Two recent architectures illustrate this. The partnership between Lefebvre Dalloz, Xelya and Anomia combines a reliable legal knowledge base, the software environment firms use daily and initial then continuous training: the value lies not in one more interface, but in integrating content and production into the practice system, with guarantees on hosting and the non-reuse of data. The choice made by the Conseil supérieur du notariat — Mistral AI and Scaleway, announced in July 2026 — follows a more institutional logic: providing a controlled European environment, organising acculturation and collectively framing usage, without touching the notary’s mission or responsibility.

These examples open a space for specialised consulting firms, smaller than the big integrators but able to bring together domain understanding, regulation, institutions, decision circuits and execution. Their role is not to resell a platform: it is to operate adoption — selecting the use cases, structuring the partnerships, building the governance, connecting the players in the ecosystem and driving deployment through to observable results. In regulated markets, this knowledge of the terrain is a more defensible advantage than generic configuration skills. On one condition: preserving independence and technological plurality — a transparent diagnostic method, comparable criteria, an open architecture, and the ability to say that a tool is not the right fit. The European market could thus see the emergence not only of legaltechs closer to the service, but also of consulting firms closer to the product.

## Conclusion

The multiplication of alliances between legaltechs and consulting firms is not a peripheral phenomenon. It responds to the fundamental constraint of enterprise AI: technical capability advances faster than the workflows, responsibilities and skills that turn it into results. In law, this constraint is compounded by confidentiality, verification, independence and professional liability. “Service as Software” provides the lens — provided it is not confused with total automation. The credible trajectory is that of a productised, governed service: software produces a growing share of the work; the firm designs the process, organises adoption, supervises the exceptions and owns the judgement; the client buys a result that is faster, better covered, more measurable.

Our conviction, in short: as models become commoditised, the ability to deploy is itself becoming the product. The best-placed players will be neither those who alone hold the best model, nor those who line up the most consultants, but those able to assemble technology, knowledge, workflow, adoption and responsibility into a coherent, proven, improvable offering — without diluting the judgement, trust and independence on which legal service is founded.

---

Markdown version of https://leaid.ai/en/articles/service-as-software-capacite-deployer — Leaid · https://leaid.ai · https://leaid.ai/llms.txt