Scope orchestrates four AI models, each assigned to the task it does best. What follows explains the pipeline, the model choices, and the cases the tool deliberately does not handle.
01
The 5-step pipeline
A chronological, readable, traceable flow. You stay in control at every step: AI proposes, you decide.
01
Brief
2 min
You import a discovery transcript, an RFP PDF, a recap email or handwritten notes. Audio, PDF, DOCX and plain text are accepted.
02
Extraction
60 s
Scope automatically identifies objectives, constraints, stakeholders, deliverables and grey areas. Every piece of information is traced back to its source in the original document.
03
Scoping
10 min
You validate the extraction and answer 3 to 7 prioritised clarification questions (blocking, major, minor). Grey areas become traced assumptions.
04
Estimation
5 min
Breakdown into work packages, roles and person-days with optimistic, realistic and pessimistic ranges. Your daily rates per role and tech are applied automatically.
05
Proposal
20 min
9-section scoping note + estimation exported as PDF and DOCX, signable via DocuSeal. You edit in block mode before sending to the client.
Product UI shown in French — Scope is published from France.
getscope.dev/projets/veolia/cadrage
Octovia
Sections
Contexte
Objectifs
Périmètre
Livrables
RACI
Estimation
Planning
Risques
Note de cadrage · Phase 1
RACI · Plateforme data Veolia
La gouvernance projet est tripartite — sponsor métier, DSI, équipe Octovia. Les arbitrages sont collégiaux sur les sujets d'interopérabilité.
SUGGESTION IA
Le DSI valide les choix techniques.La DSI Aurélie Vasseur (R) arbitre les choix d'architecture, sur recommandation conjointe de l'architecte solution Octovia (A) et du Tech Lead côté client (C).
AccepterRejeterbasé sur 12 cadrages similaires
Activité
R
A
C
I
Architecture cible
AV
ML
TL
DPO
Migration connecteurs
ML
TL
AV
DSI
POC IA prédictive
DS
ML
AV
Métier
Sources
transcript_kickoff.txt
RFP_veolia_2026.pdf
arch_diagram.png
Suggestions IA
Manque PIA RGPD dans la matrice — j'ajoute la ligne ?
Step 03 — block editor with AI suggestions you can accept in one click.
02
Why 4 different models?
No model is good at everything. Rather than a single giant model that does everything passably, the pipeline assigns the right model to each task.
This multi-model approach significantly reduces inference cost vs a monolithic GPT-5, while raising quality step by step. Each model stays in its sweet spot, and a provider can be swapped without breaking the rest of the pipeline.
Step
Model
Why
Region
Extraction
GPT-4o mini (OpenAI)
Very fast and cheap on long contexts, ideal for parsing 45-minute transcripts or 60-page RFPs. Fallback: Claude Haiku 4.5.
Via OpenRouter passthrough · OpenAI operates from the United States; transfers are governed by EU Standard Contractual Clauses and the contractual data_collection=deny directive.
Clarification
GPT-4o mini (OpenAI)
Excellent at rephrasing a business question in formal French, and prioritising what is blocking the scoping. Fallback: Claude Haiku 4.5.
Via OpenRouter passthrough · OpenAI operates from the United States; transfers are governed by EU Standard Contractual Clauses and the contractual data_collection=deny directive.
Scoping note generation
Claude Sonnet 4.6 (Anthropic)
Native FR professional writing quality, robust handling of structured schemas, long-context reasoning to assemble a coherent scoping note. Fallback: GPT-4o.
Via OpenRouter passthrough · Anthropic operates from the United States; transfers are governed by EU Standard Contractual Clauses and the contractual data_collection=deny directive. EU-resident routing options (Bedrock eu-west-3, Mistral Large) are under evaluation.
Person-day estimation
Claude Sonnet 4.6 (Anthropic)
Strong quantitative reasoning on breakdowns by work package and role, stable schema-driven output, self-check on ranges. Fallback: GPT-4o.
Via OpenRouter passthrough · Anthropic operates from the United States; transfers are governed by EU Standard Contractual Clauses and the contractual data_collection=deny directive.
Product UI shown in French — Scope is published from France.
getscope.dev/projets/veolia/chiffrage
Charge
142 j/h
127–170
Valeur HT
168 k€
marge 18 %
Confiance IA
82 %
12 cas similaires
Charge par profilcalibré
Profil
TJM
min
j/h
max
Distribution
ATArchitecte
1 450 €
10
12
16
TLTech Lead
1 200 €
24
28
32
DEData Eng. senior
1 050 €
38
42
50
DSData Scientist
1 100 €
20
24
30
DEVData Eng.
850 €
18
22
28
DEVOPSDevOps
950 €
8
10
14
Step 04 — person-day estimation with confidence ranges per role.
03
Compared with the alternatives
Plenty of tools can produce clean text. Very few understand the reality of a French ESN pre-sales cycle.
Criterion
Scope
ChatGPT alone
Notion AI
Human consultancy
Brief → proposal lead time
4 hours
1 to 2 days
1 to 2 days
2 to 3 weeks
French ESN knowledge
Native (rates, roles, RACI)
None
Generic
Varies
Calibrated person-day estimates
Yes, per role and tech
No
No
Yes, manual
Signable SOW templates
Yes, eIDAS
No
No
Yes, manual
Audit log and compliance
Native GDPR, EU-hosted
GDPR risk (US)
GDPR risk (US)
Depends on firm
Cost per proposal
Included in subscription
Marginal but human time
Marginal but human time
€5,000 to €15,000
A senior consultant scopes a project in 2 to 3 weeks, for €5,000 to €15,000 of internal cost. Scope brings you to the same deliverable in a few hours of effective work, with estimation ranges that no project manager would ever have had time to produce by hand.
04
How the AI is trained
The question every CISO asks first. The answer fits in one line.
Your data is never used to train the AI models.
OpenRouter contractually relays the data_collection=deny directive to the underlying providers actually used by Scope (OpenAI GPT-4o and GPT-4o mini, Anthropic Claude Sonnet 4.6 and Haiku 4.5), prohibiting the use of your prompts or generated outputs to train or fine-tune their models. Provider-side logs are purged within 30 days at most.
An Enterprise opt-out of outbound AI calls (manual-only editing) can be contracted on demand. An EU-resident routing option (Bedrock eu-west-3 for Claude, Mistral Large hosted in Paris) is under internal validation and will be offered as an Enterprise option as soon as FR quality benchmarks are conclusive. As of today, no self-service "sovereign AI mode" is yet enabled.
So how does Scope improve?
Prompts, scoping templates and the estimation reference will evolve from an internal anonymised dataset, built exclusively with pilot partners once the pilot programme is launched (post-PMF). No customer data is used for training.
Better to be clear now about the limits than to disappoint you in month three.
Where Scope saves you time
Structured formatting, professional FR writing, estimation calibration, RACI and planning, translating a technical brief into a proposal readable by a non-tech buyer.
Where your judgement remains essential
Assessing political risk on the client side, commercial negotiation, strategic architecture choices, validation of sharp technical assumptions, and final margin arbitrage.
Cases Scope does not handle
Public tender responses in BOAMP / TED format requiring 80+ page technical memoranda, projects above €10M with multi-package crosscutting concerns, audit or due-diligence engagements.
Scope does not replace your sales expertise. The goal is to give you back the eight hours of formatting and calibration that slow down every pre-sales cycle, so you can put them where they matter: in discovery and negotiation.
Convinced by the method?
Run Scope on your next brief
14-day free trial, no credit card required. Your application data is hosted in the EU and transfers to our AI sub-processors are governed by EU Standard Contractual Clauses with a contractual no-training clause.