Your business processes automated by a tested AI agent
A plugin brings together skills, which describe your procedures to the agent, hooks, which automatically block the actions you have forbidden, and MCP connectors, which link the agent to your software. Your teams get a reliable business tool.
Why ad hoc LLM use does not hold up in production
No business guardrails
Models respond outside the agreed scope, without validation or an audit trail. The result is inconsistent answers, declining quality and legal exposure.
No access to your systems
Without connections to your systems, AI has no access to your CRM, databases or internal tools. Your teams copy and paste. The model lacks the context.
No tests
No guarantee that a model update or prompt change will not break today’s workflow. Regressions are discovered in production.
No method
Without a method, the project becomes impossible to maintain, hand over or develop further. It depends entirely on the person who wrote it: their departure is enough to freeze it.
What a Colombani.ai agent contains
Skills
Business expertise encoded as reusable modules that can be versioned: sales processes, accounting standards, clinical protocols, legal rules.
Hooks
Business rules applied automatically at each step: input validation, structured logging, guardrails, audit trail.
MCP
Controlled access to internal systems (CRM, ERP, document repository, business APIs) without exposing credentials to the user or the model.
The right model for your case
Claude, GPT, Mistral or a model installed on your machines, chosen at scoping according to the task, the cost and the sensitivity of the data. The best practices published by the chosen vendor are applied: instruction structure, tool use, agent loops, context management.
Tests and benchmarks
Reproducible evaluation suite. Comparison against a baseline (a human process or a competing tool). Metrics: accuracy, latency, cost.
Docs + monitoring
Technical documentation for your teams. Usage dashboards. Regression alerts. Any drift by the agent is flagged to you before your teams notice it.
Agentic systems architect method, in five phases
1. Architecture
Workflow mapping with the business team. Decision points and success metrics identified. Deliverable: functional specification + KPIs.
2. Design
Selection of required skills, hooks and MCP. Choice of the best-suited model, whatever the vendor. Effort and risk estimation. Deliverable: technical design.
3. Build
Continuously tested code. Strict compliance with the design patterns recommended by the chosen model’s vendor. Code review at every step. Deliverable: working plugin + test suite.
4. Benchmark
Evaluation on real cases against a baseline (a human process or an existing tool). Metrics: accuracy, time, cost. Iteration until the defined quality threshold is reached. Deliverable: a benchmark report.
5. Deployment
Pilot rollout to one team. User training. Monitoring set up. Deliverable: plugin in production + documentation + 30 days of support.
Example plugins already designed using this methodology
Sales Ops Plugin
Inbound lead qualification: email reading, scoring, a sales briefing before the meeting, CRM updates. Stack: CRM-aware skills + validation hooks + HubSpot or Salesforce MCP.
Legal Review Plugin
Contract document analysis: extraction of sensitive clauses, GDPR and AI Act checks, a risk report. Stack: legal skills + source citation hooks + document repository MCP.
Month-end Close Plugin
Month-end reconciliation: multi-source aggregation, anomaly detection, report generation. Stack: accounting standards skills + control hooks + ERP MCP.
Medical Coding Plugin
ICD-10 code generation from clinical notes: extraction, suggestion, validation. Stack: medical nomenclature skills + traceability hooks + EHR MCP.
Six commitments kept on every plugin
Anthropic and OpenAI certifications
A method built on the best practices published by the vendors, led by a founder who holds the Claude Certified Architect Professional certification, issued by Anthropic, and nine practitioner certifications issued by OpenAI.
Qualiopi
Quality framework applied to engineering processes. Traceable deliverables, decisions and evaluations.
Tests and benchmarks
Every plugin ships with its evaluation suite. The result is reproducible on every run, and quality is measured.
Source code delivered
Code, configurations, documentation and the repository belong to you, with no subscription or recurring licence fees for the plugin.
Monitoring included
Usage dashboards and regression alerts. You know if the plugin drifts before users report it.
Optional maintenance
Monthly subscription for enhancements and monitoring of model updates. You retain control of the schedule, with continued support after delivery.
Frequently asked questions
How does this differ from conventional AI development? +
A conventional AI development project delivers a script or an API. A Colombani.ai plugin delivers a structured system (skills + hooks + MCP), tested, documented and aligned with the chosen vendor’s best practices, with its evaluation suite and monitoring. A script often stops working at the first model update; a tested plugin adapts.
How long until a plugin is operational? +
From 4 to 12 weeks, depending on workflow complexity and the number of systems to integrate. A simple use case (lead qualification) can be in production in 4 weeks. A plugin spanning multiple systems with complex integrations takes 8 to 12 weeks.
Which systems can be integrated through MCP? +
Any system with an API (REST, GraphQL, SOAP). Connectors available for Salesforce, HubSpot, Notion, Google Workspace, Microsoft 365, GitHub, Slack and SQL or NoSQL databases. Custom connectors possible for internal systems.
What happens when the model evolves? +
The plugin is versioned. Every model update, or change of vendor, is tested against the evaluation suite before it is deployed. If you subscribe to maintenance, Colombani.ai monitors model updates and handles migrations. Otherwise, the documentation explains how to do this in-house.
Can we use a local model for sensitive data? +
Yes. Plugins work with models from the main vendors (Claude, GPT, Mistral) as well as local models (Mistral, Qwen via Ollama), and can switch models without being rewritten. The choice is made during scoping, based on data sensitivity. See the sovereign AI consulting offer for architectures that run entirely locally.
How do you ensure the plugin stays reliable over time? +
A test suite runs on every change. Monitoring raises an alert when usage or quality drifts. Documentation lets your team maintain the plugin without depending on Colombani.ai.
Got a project in mind?
Describe your situation. The first 30-minute call is free. Get a frank response within one business day.