01
Diagnostic
Find where AI pays, before buying a tool.
You get
- A map of your processes and data
- Opportunities ranked by value, feasibility, data readiness and risk
- One pilot chosen, with go or no-go criteria written in advance
Applied AI. Designed for real operations.
Tracee helps organizations identify where AI can remove operational friction, design the right solution, and integrate it into existing processes with the controls required for daily use.
For
Organizations, SMEs first, whose daily work runs through documents, requests and business systems.
You get
A ranked use-case register, a written design, a tested prototype and a controlled path into your systems.
How we work together
Design-and-integration mandates with a decision gate at every stage.
What we offer.
Four services. Each can be bought on its own, and each ends with something you keep. Most clients start with the diagnostic.
02
Workflows and agents
AI designed into the process, with a person reviewing what matters.
You get
- A step-by-step workflow specification, failure handling included
- A prototype tested on labelled cases
- The integration design for your existing systems
03
Deployment and data
Decide where models run, what providers keep and who operates them.
You get
- A map of every data flow to and from a model
- Deployment patterns compared for your data and capacity
- Supplier terms checked: retention, training use, location
04
Method and governance
Validate, integrate and hand over with controls in place.
You get
- A control matrix with an owner for every control
- A test report against criteria fixed before testing
- Runbook, trained reviewers and monitoring handed over
Four problems we are asked to solve.
Most requests start from one of these. None of them is solved by choosing a better model.
Where is AI actually worth investing?
Every demo looks useful. The real question is which processes have the volume, the data and the tolerance for error that make AI pay for its upkeep, and which ones ordinary software would handle better.
Our documents and data are scattered.
Contracts on a shared drive, invoices in email, three versions of the same procedure. Before any model, someone has to decide which sources count, who owns them and who may see them.
The prototype never reached operations.
A demo on twenty clean examples is not a process. Daily use needs exception handling, a reviewer, logs, a way back, and a named person accountable for the output.
We are not sure the integration is safe.
Where the data goes, what the provider keeps, which network path it takes and what happens when an output is wrong are design choices. Each one has a cost and an owner.
Where AI pays off.
Typical use cases we assess. The method is the same in every sector; the use cases come from your own processes.
Finance and treasury
- Bank reconciliation and matching
- Cash forecasting
- Exposure and limit monitoring
Compliance and risk
- KYC and onboarding file review
- Alert triage for transaction monitoring
- Regulatory watch and summaries
Operations
- Invoice, contract and document extraction
- Demand forecasting and stock planning
- Scheduling and reporting
Customer service
- Request triage and routing
- Draft answers for agent approval
- Knowledge base upkeep
Sales and marketing
- Market and competitor analysis
- Content drafting and editorial calendars
- CRM updates
Teams and knowledge
- Internal search across documents
- Drafting, translation and summaries
- Training on AI tools
Agentic AI workflows, under human control.
An agentic workflow lets an AI agent carry out steps on its own inside your tools. The design question is never whether the agent is clever enough; it is where it stops and a person decides. How we design them.
01
The agent prepares
Gathers documents, checks them against rules, extracts what is needed and drafts a proposal, within permissions you set.
02
A person decides
Anything with a real consequence goes to a review queue, with the evidence and the agent's uncertainty shown.
03
Everything is recorded
Inputs, versions, proposals, decisions and overrides are logged, so any outcome can be explained and reversed.
Three application patterns, shown as designs.
These are illustrative patterns that recur across sectors. They show how we design; they are not completed client deployments.
Document extraction with human validation
For invoices, delivery notes and forms. The model handles only what rules cannot, and every value it proposes can be traced to the page it came from.
- Document received and fingerprinted
- Rules check format, duplicates, attachments
- Model extracts fields on unknown layouts
- Doubtful fields flagged with a reason
- Reviewer approves or corrects
- Export to the business system, logged
Internal knowledge retrieval with sources
For procedures, policies and product documentation. Answers carry their sources, and when the corpus holds no answer the system says so.
- Question asked by an employee
- Access rights filter the corpus
- Passages retrieved from curated sources
- Model drafts an answer citing them
- User checks the cited passages
- Gaps logged for the content owner
Workflow triage with an approval queue
For customer requests, internal tickets and supplier queries. Nothing with a real consequence leaves the queue without a person.
- Request arrives on any channel
- Rules route the obvious cases
- Model classifies the rest, drafts a reply
- Approval queue, ordered by risk
- Action taken in the business system
- Outcome recorded for evaluation
How delivery works.
Four stages, each closed by a decision you make on evidence. The method page lists the inputs, work, outputs and gate of each one.
-
Stage 01
Diagnose.
Processes, data and baseline mapped. Opportunities scored. One pilot chosen, or none.
-
Stage 02
Architect.
Workflow, controls, deployment pattern and suppliers specified before anything is built.
-
Stage 03
Validate.
A prototype measured on labelled cases against criteria written in advance.
-
Stage 04
Integrate and enable.
Connected to your systems with your IT team or partner, reviewers trained, monitoring handed over.
What you receive.
Working documents your teams keep using after we leave, not a slide deck.
Use-case prioritization
Opportunities ranked by business value, feasibility, data readiness and risk, with the evidence behind each score.
Design specification
Each workflow step, its rules, where AI is used and where it is not, the review points and the integration contracts.
Data and deployment assessment
Sources, owners and sensitivity of the data, and the deployment pattern that fits them, with provider terms checked.
Prototype and test plan
A working prototype, the labelled evaluation set it was measured on, and the results against the agreed criteria.
Control matrix
Access, review, logging, versioning and rollback controls, each mapped to a risk and assigned to an owner.
Scope, stated plainly
We design, specify, prototype and support integration. Hosting, operation or a managed service are scoped separately, with the owner named.
Financing in France
Your AI project can be financed.
French companies with up to 49 employees and more than three years of activity can finance an AI integration project with Bpifrance's Prêt Boost Intelligence Artificielle, an unsecured loan for AI adoption: automation, predictive analytics, content generation, and team training.
The loan is granted by Bpifrance to your company, and the application goes through your chartered accountant. We provide the scoped project description and budget that support it.
tracee is an independent provider and is not affiliated with Bpifrance. Loan terms are set by Bpifrance and may change; check the current conditions on its site.
See the Prêt Boost IA on Bpifrance€5,000 to €100,000
Loan amount.
Up to 49 employees
Company created more than three years ago.
No guarantee
No personal or company guarantee required.
3 to 5 years
Repayment term, with up to 12 months before the first repayment.
Technical guides.
Original, dated and sourced. They explain how we reason before you ever talk to us. All guides are in Insights.
Guide A
Choosing an AI deployment posture: residency, retention and isolation are three different questions
Guide B
Data before models: most AI pilots fail on ownership, not accuracy
Guide C
From AI pilot to production: a vendor-neutral reference pattern and its acceptance tests
Tool
Deployment Explorer: screen five deployment patterns against your data and your capacity
Who does the work.
The practice is founder-led. Bassel Assaad is an engineer with a doctorate in IT and systems from Sorbonne University. Through tracee, Bassel works on payment infrastructure and regulated financial products, where every flow has to be traced, reconciled and explained. The Applied AI practice brings that discipline to operational AI: explicit data boundaries, review points and evidence.
Tracee is independent. It resells no model, platform or license and takes no referral fees, so supplier comparisons are documented on their merits. Full founder profile.
Frequently asked
Common questions about AI services.
What is an agentic AI workflow?
A process in which an AI agent carries out steps on its own, such as gathering documents, checking them and drafting a decision, while a person approves the result at points you define. Unlike a chatbot, the agent works inside your tools, under rules you set.
Where should we start?
With the diagnostic. It maps one or a few processes, measures a baseline and ends with a written decision on one pilot, or a reasoned recommendation not to use AI there.
Do you work with companies outside finance?
Yes. Our deepest experience is in regulated finance, where controls are strictest, and we apply the same method to any company that wants AI in its daily operations, from SMEs to large groups.
Which AI tools do you use?
We are independent: no software to sell, no referral fees. We recommend what fits your data, budget and constraints, whether an off-the-shelf tool, a configured platform or a custom agent built on a leading model.
How do you handle data protection?
It is part of the design: where data is processed, who can access it, what the model is allowed to see and what is logged. We document the data flows, processors and retention your data protection officer or counsel needs for the GDPR records and to assess EU AI Act obligations.
Can our AI project be financed?
In France, companies with up to 49 employees and more than three years of activity can apply for Bpifrance's Prêt Boost IA, from €5,000 to €100,000, without guarantee. We provide the scoped project description and budget for the application.
Start with a conversation.
Thirty minutes, no slide deck, no obligation. Tell us which work you want AI to take on, and we will tell you whether we can help, including when AI is not the right tool.