[01 · WHAT WE SEE]
Every tool shows up.
Not one is declared.
Monday, 9 am. What an ordinary SME doesn't see.IT knows none of them.
TOOLS0 / 9Representative tools. An illustration, not a measurement.
Three tensions, one job.
[01]
BUSINESS
Teams kit themselves out, often for good reason, and IT finds out later.
[02]
FINANCE
Nobody sees the real cost of AI, because tokens, APIs, licences and agents are scattered across supplier invoices and expense claims.
[03]
REGULATOR
Auditors want proof of compliance quickly, in their own format. Usually it doesn't exist, because nobody produced it along the way.
WHAT'S AT STAKE
over40%
of agentic AI projects will be cancelled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls.
We read those three reasons (cost, value, risk) as the blind spots a control tower has to cover.
Your controls were never built for agents.
- Identity[01]
An agent has no badge, and it appears in neither the directory nor the access review.
- Pace[02]
An agent calls a model at every step, day and night, faster than any human review can follow.
- Cost[03]
The agent is paid for by the token rather than the licence, so the budget comes to light on the invoice.
- Data[04]
An agent reads and sends files; a contract or a payroll file can leave in a single request.
- Proof[05]
The agent decides without anyone signing; ask ‘who approved this?’ and the answer is ‘the model’.
An agent can act on its own, provided a flight plan frames it and a tower on the ground keeps its log.
What we see
02 · THE CONTROL TOWER
One tower for all your agents.
Ascenzia is the platform for creating and steering the uses of AI in the company. The studio builds agents, the tower watches them, and we want leadership, IT and legal on one screen.
Ascenzia Sentinel
A creation studio. A control tower.
01 /Create
Ascenzia Studio
02 /Steer
Ascenzia Sentinel
03 /Choose
Multi-model gateway
Every new agent sharpens the picture rather than blurring it.
One agent keeps to its rule, all logged.
One screen, four ways to read it.
[01]
EXECUTIVE
You track a capped budget in euros for each agent, with no converting from tokens, and the company sets the rule.
[02]
CIO
You spot Shadow AI and set a policy for each agent (approved models, guardrails, budget). If you prefer, you keep your own API keys.
[03]
EMPLOYEE
They use a single access point, the gateway picks the model, and they know their cap before the first request.
[04]
COMPLIANCE
The execution log and costs in euros keep themselves, and the export is ready the day the audit asks for it.
REAL-WORLD CASE · A TOWN HALL IN SEINE-SAINT-DENIS
12
steps automated with an agent in production.
7
town hall departments involved.
The process runs over several weeks. The agent sends reminders in Microsoft 365 and has AI review the documents, but a person signs off at the key stages, by design. It prepares the municipal council meetings under the control of Ascenzia Sentinel.
[03 · UNDER THE HOOD]
Maxime de Tourtier
President
Laurent Fontaine
Chief Product & Technology Officer
Two founders with 25 years of IT experience each, on systems that cannot afford to fail, and a company working in France and Switzerland.
Your data stays home.Your AI, no strings.
[CHOOSE · MULTI-MODEL GATEWAY]
Claude, Mistral, Gemini, GPT (OpenAI), Llama, Grok. You set the catalogue and switch models when you decide to, without touching your agents; if one goes down, the gateway fails over.
Your agents
HOSTING
Platform, memory and database in France (Paris region) or in Switzerland with Hidora. For each model, you can see where inference runs.
METADATA ONLY
Guardrails read each request in transit, but we keep only the verdict (status, duration, token count), never the prompt or the response.
CUSTOMER-MANAGED KEYS
Your keys, and the models you approve.
ISOLATED EXECUTION
An agent runs in an isolated environment and reaches only the connectors you open to it.
COMPLIANCE
The platform helps you comply with DORA, NIS2, the AI Act, the Cloud Act and GDPR.
Steering also means spending less.
The gateway sends each task to the model that is good enough for it. On simple tasks, the cost per token is ten times lower. After optimisation, document-sorting agents use 60% fewer tokens.