Skip to main content

BUILD VS BUY

Build is possible but rarely within reach...

Before 2026, the people you turned to for putting software in place were:

  • your product or tech teams,
  • vendors of the publisher, consultancy, or AI-agency kind.

We believe these players are capable enough for simple needs, but that they are held back by their organizational and technical debt, in general.

The gap with academic research clearly shows that companies have therefore not been able to innovate in AI these past years.

A reality check, people and hiring

We are betting that consultancies and IT services firms have not had the time needed to acquire the expertise, given the acceleration that has just taken place. Their tech teams, like your in-house teams, are tied up on the problems of the moment, which often blocks the long training this subject requires.

A lack of real innovation

LinkedIn conversations show a 2–3 year gap between what it is possible to do and what people claim to be working on. So it is with the prevalence of systems over models (multi-hop, Stanford, March 2023), and with using the terminal to generate code on the fly, the famous harness (ADAS, Facebook, August 2024).

BUILD VS BUY

Buying is now more effective and less expensive...

We believe AI will be outsourced to specialists. That realization will rest on this simple argument: these new players can combine:

  • deep expertise in AI, software engineering and systems, for any kind of need,
  • a higher production capacity, through a generic, unified architecture,

and therefore offer you a tailored service, more effective, faster and less expensive than those of the incumbents — most often generic and hard to evolve — who lack either expertise or organizational flexibility.

An AI studio, for example, has its own generic AI-native application and infrastructure framework (a little like a high-code n8n framework). Run by a small team for all of our clients, this tool opens a new horizontal business model and lets us offer you:

  • an MVP of your system, free, within a week (a production-ready demo, once we receive a dataset),
  • a deployment on GCP infrastructure, 1–3 weeks later (after your feedback, the iterations, and model training),
  • corrective and evolutionary maintenance,
  • a price that decreases with the number of products ordered.

We usually run an internal audit when the need is not specified from the outset.

BUILD VS BUY

Why are pure players less expensive?

Conway's law, and the recent history of microservices, often adopted by organizations for the wrong reasons, shows the advantage reserved for AI pure players who control their methods from the start, with a small headcount.

We implement a modular-monolith software architecture, going beyond the residual overheads our competitors still face because of their history.

  • a small team, full technical ownership, every client. Separation and genericity of project code. Less communication overhead,
  • stack: clean architecture, no dependency on heavy AI frameworks, or on complicated APIs and syntax,
  • a global systems view, translated into our design infrastructure for every project, tech and industrial (real-time, critical, model-checking & AI), as a SaaS/dashboard, conversational or A2A/MCP version. Every improvement benefits all products already in production,
  • minimal DevOps. Generic IaC focused on serverless and scalable VMs (no container orchestration). Less CI/CD overhead, through the absence of conflict and full ownership on every maintenance action.

Below, we set out how we go beyond these residual overheads of the classic modular monolith, guaranteeing you lower costs than historical players (with the performance of bespoke work on top) and fast access to high-value systems alongside your current means.

AI-native modular monolith (our solution)

Technical and organizational complexity

🟠 limiting🔴 very limiting🟢 not limiting

Microservices

  • A substantial effort on deployment infrastructure (DevOps). Container orchestration, born of microservices, suited organizations that were not very agile, but well funded, and split into autonomous services (Conway's law).

Modular monolith

  • The monolith brings back the "single delivery train", where the whole company depends on the same pipeline. The continuous-integration queue (CI/CD) becomes the organization's bottleneck.
  • Teams remain subject to Conway's law: human and technical staffing leads to compromises. For example, the container-orchestration infrastructure is sometimes kept when moving back to a monolith architecture.
  • The popularity of frameworks such as langchain in Python, though suboptimal, shows that technical debt is a real issue in AI, another brake on real innovation.

AI-native modular monolith (our solution)

  • Projects are generic; they differ only in their business building blocks. This segmentation makes corrective and evolutionary maintenance easier for the teams, who are competent on every client project. The CI/CD philosophy is therefore minimal, and comes down to versioning (no conflict, full ownership on every action, on these business perimeters only).
  • A studio sees serverless and VMs as sufficient for modern AI problems. Our cloud solutions therefore do without container orchestration from the start.
  • Our application framework is built on DSPy (Stanford), which is more effective and more ergonomic than other popular frameworks that have severe limitations (API dependency, syntax).

Bill

🟠 limiting🔴 very limiting🟢 not limiting

Microservices

  • Very high. It requires permanent Kubernetes clusters and a service mesh to orchestrate the network, which creates a large fixed minimum cost even with no traffic.

Modular monolith

  • High to moderate. The Kubernetes infrastructure is often kept, but constrained so it can run a single large container.
  • Incumbents have not yet reduced their organization to small technical cores limited to shared, high-value methods. That extra cost is necessarily passed on to your bill, whatever the narrative about new pricing models.

AI-native modular monolith (our solution)

  • Right-sized. Two managed primitives, matched to the kind of load.
  • Serverless (Cloud Run): Scale-to-zero when latency is acceptable, or scale-to-one that is not billed at rest (CPU allocated only to the request). This makes it possible to run dozens of client projects with an infrastructure cost at rest close to zero.
  • Dedicated VM (Compute Engine): Scale-to-one that is billed, to guarantee immediate responsiveness. Reserved for projects with models hosted on GPU (confidentiality requirements, or intensive volume that is more cost-effective than API tokens). The VM can host the whole application or only its AI module (a hybrid architecture that gives granularity back, for the price of a single extra deployment).
  • A concentration of tech and systems expertise, and no technical or organizational debt: our systems are therefore designed and run with a small team, as expected of AI pure players.

Time-to-Market & Project Bootstrapping

🟠 limiting🔴 very limiting🟢 not limiting

Microservices

  • Slow (3 to 6 weeks for the first production-ready component).
  • Every new service means creating a repository, configuring network routing, defining API contracts, provisioning resources with Terraform, and integrating the service into the service mesh and the CI/CD.

Modular monolith

  • Moderate (1 to 3 weeks of preparation).
  • The initial start of a project is faster than with microservices, but the team usually has no reusable end-to-end method, expressive enough, for designing a new AI system. Communication, infrastructure and stack overheads still mean several weeks of startup, or longer depending on the size of the organization.
  • Beyond pricing that will not come down quickly, especially on complex systems, your commercial or consulting relationship will remain frustrating, built around a story meant to justify these costs, made of risk and complexity.

AI-native modular monolith (our solution)

  • Immediate (1 to 2 days for a deployed prototype).
  • Thanks to the studio's reusable generic template, client projects start on a ready-made base. From the first hour, the effort is focused on business logic and AI components. The scalable deployment files are reused as well.
  • If an AI studio's pricing is clear and reasonable from the commercial proposal onward, questions of technical and regulatory expertise, audit, maintenance, SLA commitments and, above all, of making any bespoke system available quickly, remain ahead of what indebted historical players deliver. A studio also offers you tailored SLAs and insurance, but that should not get in the way of product quality and delivery!

Our studio has six years of experience in language models and has also built systems such as :

  • Deep Research & Advanced RAG : Complex document analysis engines, multi-source research, and high-precision semantic extraction

  • Health Tech & Media Intelligence : Ingestion and structuring pipelines for medical and publishing sector data

  • Algorithmic Trading & Finance Agent : Predictive analysis models, time-series processing, and financial decision support

  • Propale Agent : Agent that drafts a commercial proposal automatically after a call,

  • Excell-killer Agent : It was time to move forward on how you present your data, conversationally !

So what MVP would you put in front of your CEO or sales team next week ? Find many anticipated use cases for agentic industry under the Use cases tab !

See also

REFLECTIONS & USE CASES

Market fit

Topics often framed as blockers: adoption, complexity, strategy. Our approach

Use cases

Studies, market analyses, and many use cases to inform your thinking, by sector or function

THREE NEEDS

AI for your processes

Audit your business, outside perspective on ambition and what is possible. Agentic AI aligned with regulation and your business reality

AI for your customers

Customer-facing IT: let customers query your data directly. Support 3.0, conversational, online presence

AI in your product

Software vendor ? Integrate agentic AI in your product to the highest standards, with fast execution