Skip to main content

Perello Consulting

Not a promise. A method tested in production every day.

By Laurent Perello, founder of Perello Consulting — web pioneer for over 25 years, AI operator in production since 2024. Last updated: 13 April 2026.

The Perello Consulting AI methodology holds to one constraint: no workflow is delivered to a client until it has proven itself in production, on our own infrastructure. Four stages — build, prove, systematize, duplicate — with a fifth added to complete the cycle: transmit. All of it operated by one human, seven AI orchestrators, and their teams of specialized agents.

A framework tested on ourselves first

We build for ourselves before delivering for others. The method is not a theoretical framework borrowed from a white paper or a certification: it is the documented extract of what has withstood, on our side, thirty-six consecutive days of production. As of 13 April 2026, the firm's public journal counts thirty-six dated entries, continuously verifiable at perfectaiagent.xyz.

This choice is not an affectation. It is an answer to a shared observation: a significant share of digital transformation projects fail not through technological shortcomings, but through the absence of real operational practice on the part of those who prescribe them. France's DGE, in its AI self-assessment framework published with Bpifrance and France Num (francenum.gouv.fr, autodiag IA, consulted April 2026), prescribes a structured evaluation by capability. We apply that grid to ourselves before offering it to any client.

A firm that sells a method it does not operate is selling a document. A firm that operates its method and then transmits it is selling a capability that runs. The gap between the two does not show in the brochure. It shows the day delivery is required.

Building for yourself first also means accepting that theory is never enough. The real impact of an automation can only be measured under real conditions, with teams who live with it daily — not in a controlled setting. We took this requirement at face value: our own workflows are our first clients. What breaks on our side will never be deployed on yours. What holds on our side is, by definition, ready to be transmitted.

The flywheel in five stages

The method unfolds in five linked phases. Each produces a concrete, dated, verifiable output. None can be skipped without breaking the next.

Build

To build is to deliver a system that works under real conditions. Not a mockup, not a demo prototype, not a specification document: an operable workflow, versioned, deployed. The output is a usable artefact from the moment the phase ends.

On 10 April 2026, a complete design configurator was delivered in a single working day across our internal portfolio: twenty-four files, two thousand eight hundred lines of code, a scope a traditional agency would spread across three weeks. That figure is archived in the public journal for that day. It is not presented as a record; it is the ordinary cadence of an infrastructure built for that density.

Building at this pace is only possible because nothing is rebuilt from scratch. The components exist. The agents know what to produce. The orchestrators coordinate without meetings. The marginal cost of a new deliverable tends to fall with each iteration. This pattern is documented in the AI productivity studies published by the OECD (oecd.ai, AI Policy Observatory) and referenced by France Stratégie in its work on AI's impact on labour (strategie.gouv.fr).

Prove

To prove is to measure what the system produces and compare that figure to what it cost before. The proof is not qualitative. It is numbered, dated, reproducible. Without this stage, nothing that follows holds.

The discipline of measurement takes two forms on our side. First, the public journal: each day, what was delivered, what broke, what was corrected, documented with a date. Then the calculation in time equivalent: on 3 April 2026, fourteen articles were drafted, reviewed, and published in one session — a volume a traditional editorial team covers in three full weeks. On 7 April, five hundred and thirty-two documents were migrated in a single pass, without manual intervention.

These figures are verifiable in the public commits of the VantagePeers repository and in the dated diary entries. Productivity measured in production is the only indicator that holds over time. Everything else belongs to communications.

Systematize

To systematize is to extract the mechanics of the first success so they can be replicated without depending on the founder, a team member, or the particular context of the first deployment. A documented process that another person — or another agent — can operate without reinventing.

In practice: every workflow that has held for several weeks is converted into a deliverable. The code is commented, documentation is written, error cases are catalogued, environment variables are listed. Knowledge leaves its author's head and enters a transmissible artefact. This is the phase where the speed of future implementation is decided.

The CNIL's public methodology for documenting AI processing (cnil.fr) imposes a similar discipline: map the flows, trace the decisions, name those responsible. We draw on it for business documentation: every deliverable deployed at a client site is accompanied by its processing map, aligned with that framework.

Duplicate

To duplicate is to deploy the same model in a new context — another team, another division, another client — with an implementation time that decreases with each iteration. This is the stage where the initial investment becomes a multiplier.

A dated example: Palmarès Digital Auto, the first client project to have moved through our full delivery cycle, is in production. The firm's internal workflows were adapted to the automotive business context, redeployed, documented, and handed over. What would have taken a team starting from zero several months of scoping, we delivered in weeks — because the foundation was already running.

The EU AI Act (Regulation 2024/1689, eur-lex.europa.eu) governs the reuse of AI systems across contexts, requiring compliance documentation that travels with the deliverable. Our duplication architecture respects this framework by design: every deployment carries its documentation, its scope, and its guarantees.

Transmit

To transmit is to open what is mature so others can benefit from it without us. VantagePeers, the inter-agent coordination infrastructure we built for our own teams, is published as open source on npm under an MIT licence. The code is public, versioned, installable by any organisation that needs it — including our competitors.

This choice is not altruistic. It is economic. A method kept secret dies with its author; a method transmitted enters a common language and installs its vocabulary there. Transparency accelerates iteration: feedback from the community flows back, patterns stabilize, the method gains robustness without our having to control everything. ANSSI, in its security recommendations for AI systems (cyber.gouv.fr), encourages this public traceability as a vector of trust. We make it an operational principle.

The three-tier architecture

Three distinct levels, with explicit boundaries. Not a human team with AI tools. Not an AI replacing a team. A three-tier architecture — named, documented, visible in the public journal.

Tier 1: Laurent Perello (human). Sets priorities, validates deliveries, arbitrates strategic choices, meets clients. The only human on the Team. No longer writing code day to day: he thinks, decides, transmits. Each morning he reads the state of the system, assigns missions, resolves what is worth resolving. Each evening he checks what has been delivered.

Tier 2: seven AI orchestrators. Each manages a business unit with its own named scope:

  • Pi: strategy and meta-architecture (ElPi Corp, primary internal orchestrator).
  • Tau: product and interfaces (VantageStarter, VantageOS).
  • Phi: narrative and content (Perfect AI Agent, diary, articles).
  • Sigma: infrastructure and coordination (VantagePeers).
  • Omega: client delivery and catalogue (VantageRegistry, VantageOS Team).
  • Zeta: external outreach (open source contributions, pull requests on third-party repositories).
  • Eta: quality review (all pull requests entering the codebase).

Tier 3: specialized agents. Each orchestrator commands a team of atomic agents: front-end developers, blog writers, copywriters, reviewers, scrapers, analysts. An agent does not make strategic decisions. It executes a well-defined task, reports to its orchestrator, and waits for the next one.

The distinction between the three tiers is not cosmetic. It ensures that every decision is made at the right level, every execution is assigned to the right resource, and the system holds over time without depending on any single person. It is the opposite of a bottleneck: it is a resilient architecture.

What makes the method verifiable

Transparency is not a communications argument. It is an operational constraint we have imposed on ourselves so that it remains a discipline.

A daily public journal. Since 8 March 2026, one dated entry every day at perfectaiagent.xyz. As of 13 April 2026, thirty-six consecutive entries — including days when something broke, including days when a hypothesis was revised. A case study selects; a journal records.

A public codebase. VantagePeers, our coordination infrastructure, is published on npm under an MIT licence. Anyone can read the code, install it, fork it, critique it. This exposure holds us to a production quality that no private codebase ever guarantees.

Dated figures. Every data point we present carries its date and its place of verification. Twenty-four files delivered in one day on 10 April 2026. Fourteen articles drafted and published in one session on 3 April 2026. Five hundred and thirty-two documents migrated in a single pass on 7 April 2026. Fifty GitHub missions coordinated in twenty-four hours in April 2026. All of these are verifiable in the public portfolio commits and in the diary.

A public method. The page you are reading is itself the method's artefact. It is dated, versioned, revisable. Any executive who wants to compare what we say to what we do has all the pieces available to them.

This approach aligns with the CNIL's framework on AI and data protection (cnil.fr, AI sandbox), which recommends accessible documentation and public tracking of processing activities. Everything can be verified. That is the constraint we have chosen. Not a posture.

Why our costs and timelines are lower

Because there are no meetings to decide whether to hold another one. Because there is no human team to coordinate: the orchestrators communicate directly via VantagePeers, without intermediaries, without minutes to formalise. Because seven intelligences work in parallel, including nights and weekends.

The economic architecture rests on three measurable factors. First, parallelization: seven orchestrators move simultaneously across seven distinct scopes, without waiting for one to finish before another begins. France Stratégie's studies on AI's impact on labour (strategie.gouv.fr) document a significant recovery of weekly hours on automatable cognitive tasks. On our side, that recovery is absorbed by higher production density — not by free time.

Next, the absence of human coordination costs. Coordination costs represent one of the main friction items in mid-sized organisations. A firm with seven orchestrators coordinated by an automated protocol eliminates that line item entirely.

Finally, the duplication pattern. Each iteration costs less than the previous one because the reused component is amortized across multiple deliveries. A concrete counterexample: where a traditional agency invoices three weeks of scoping and six weeks of execution for a complete design configurator, we delivered the same scope in one day on 10 April 2026. The price follows the cost structure. The timeline follows the execution structure. Neither rests on a promise. Both can be checked in the commits.

The European Commission, in the recitals of Regulation 2024/1689 on AI (eur-lex.europa.eu), notes that the genuine democratisation of AI in business requires industrialised execution systems — not repeated proof-of-concept cycles. Our cost structure translates that logic into client pricing.

What you receive

A quantified audit, at no cost, delivered within one week. The report is structured by intervention priority, deployment complexity, and expected impact, actionable from the day after the meeting — with or without continuing the engagement. You describe what in your operation consumes time without producing value; we identify what can be entrusted to a team of agents and hand you the diagnosis in writing.

Beyond the audit, if you decide to move forward, you receive:

  • Teams of specialized agents deployed in your context. Drawn from our internal infrastructure, adjusted to your business, documented. Not built from scratch at your expense.
  • The source code of what runs on your side. It belongs to you. Another provider can take over maintenance: the documentation is designed for that.
  • Process documentation. Transmissible to your internal teams. What runs remains readable, modifiable, auditable.
  • Monthly measured follow-up. Production figures, costs before and after, variances against the trajectory committed in the audit report.

What you do not receive: an extended onboarding, a team of consultants to coordinate across your calendar, a dependency on our continuous presence. The Team builds so the system holds without us. That is the only architecture that lasts.

An anonymized example

Accounting firm, 22 staff, eastern France. Brief taken on 12 March 2026, audit report delivered 14 March. The diagnosis identifies 143 hours per month spent on manual entry of client statements. Proposal: an OCR extraction workflow with human validation. The system is built and tested on our infrastructure over two weeks before client deployment. Production launch on 2 April 2026. Thirty-day review: 121 hours recovered — 85% of the target. The remaining 22 hours correspond to edge cases where human validation remains faster than correcting a workflow. The workflow was installed on the client's own infrastructure, documented in their internal wiki, and handed over to their IT lead. No intervention since.

About the author

Laurent Perello leads Perello Consulting, an independent AI automation firm for SMEs. After 25 years building products for the web, he now orchestrates seven AI agents, which he runs alone, with a production journal published daily at perfectaiagent.xyz. He publishes his methodologies and his rates online so every executive can decide with full information.

Frequently asked questions

What is the difference between an AI agent and an AI workflow?

An agent is a program that makes decisions within a defined scope: reading an email, generating a report, reviewing a commercial proposal. A workflow is the sequence of multiple agents, tools, and human checkpoints that produces a complete business output. You do not manage individual agents: you request a result — a monthly report, a competitive brief, a commercial proposal — and the workflow executes it end to end. The agent is a component; the workflow is the delivery.

What does a typical deployment cost?

The audit is free. Deployments vary by scope: a few thousand euros for an atomic workflow, several tens of thousands for a full operational overhaul. The price is always presented alongside the measurable gain: time recovered per month, costs avoided, payback period. No quote is issued until the audit report has quantified those elements. You decide knowing the payback before committing to anything. For market context, the Bpifrance Diag Data IA (diag.bpifrance.fr) offers a useful reference.

Does your organisation need to be a technology company to work with us?

No. The majority of the executives we work with have no internal technical team and have no wish to recruit one. The firm addresses precisely those non-technical executives: the workflows delivered operate without client-side code intervention. Technical complexity stays with us. You receive a system that runs, readable documentation, and a single point of contact for adjustments. Your team does not need to learn a new discipline to use what we deliver.

What happens to the workflow if the firm were to close?

The source code of delivered workflows belongs to you. The method is public and documented on this site. VantagePeers, the coordination infrastructure, is open source under an MIT licence. You are not locked to a single provider. Another firm can take over maintenance: the documentation is designed for that. We deliberately choose this non-captive architecture. Trust comes from reversibility, not from lock-in. That is a commitment few firms are willing to hold.

How do you guarantee data confidentiality?

Client data does not leave the infrastructure you have authorised. The firm operates within the CNIL framework (cnil.fr) and the EU AI Act (Regulation 2024/1689, eur-lex.europa.eu). ANSSI's recommendations for AI systems (cyber.gouv.fr) are applied throughout. Each workflow is accompanied by a data processing map and an explicit contractual clause. For regulated sectors — healthcare, finance, legal — sovereign hosting is available.

What is the difference between you and a large integrator such as Capgemini?

A large integrator deploys an existing platform, adapting it to the client. We operate our own infrastructure, then duplicate it. The practical difference is speed and coupling: where a major integrator scopes over months and invoices by the day-rate, we deliver in weeks and invoice by the deliverable. We are not a direct competitor to large integrators: we occupy a segment their economic model makes unprofitable — executives of SMEs and divisions who want a system that runs, not a transformation programme.

Does the method work without Laurent Perello directly involved?

Yes. That is precisely the point. The public journal demonstrates that the system delivers every day, including when the founder is not watching. The seven orchestrators execute their missions continuously; Laurent arbitrates, sets priorities, meets clients. The firm is built not to depend on one person. A founder who is a bottleneck cannot deliver the quality promised at the scale promised. That is the only architecture that holds.

Are you Qualiopi-certified for the training component?

Qualiopi certification is being obtained through a white-label partnership. Once active, the training delivered as part of deployments becomes eligible for funding through OPCOs and the CPF. The institutional framework for professional training is outlined on the France compétences site and in the DGE guides (entreprises.gouv.fr). We will keep you informed of the date. In the meantime, technical deployments are unaffected: Qualiopi applies only to the pedagogical component.

Start with an audit

You describe what in your operation consumes time without producing value — what slows your deliveries, accumulates without being addressed, ties your teams to tasks no system has yet taken over. We identify what can be entrusted to a team of specialized agents. We quantify the time recovered and the costs avoided. We hand you a structured report — intervention priority, deployment complexity, expected impact — actionable the day after the meeting, with or without us.

The audit is the firm's only point of entry. No price is proposed before this diagnosis. No engagement is made on intuition. The method begins where the quantification begins.

Request your AI audit: free, one week, quantified report, no commitment. You can also meet the Team or read the daily journal.

Orchestrator: Alpha — Perello Consulting | 2026-04-16