David Acremann
Co-founder · Director, Professional Services
Leads the forward deployed engineers who wire agents into your real systems and stay with you until go-live.
// Agents · hosted in Europe
Agentic AI, finally within reach of your organisation.
Tessellion gives your organisation agents that do real work, ask before they act, and keep a record of every run. You set them up by configuration. The runtime already exists.
- ✓ Triage 46 mails
- ✓ Quote for Martin SA
- ✓ RFP answer
- ✓ Grant watch
// Platform
An agent is assembled, not custom-built
A custom agent is usually a large bespoke project. On Tessellion every part of it is a setting on a runtime that already exists, so you start small and change one part at a time.
- 01
Role
Start from a template and describe the job in plain words.
- 02
Model
Pick the model per task: OpenAI, Anthropic, Mistral, Gemini, OVHcloud or Scaleway, on your own models or on our private models.
- 03
Knowledge
Your documents and memory, searched when the job needs them.
- 04
Tools
MCP servers and any API, connected without code, OAuth included.
- 05
Guardrails
Workspace rules teams can tighten but never loosen, plus approvals.
- 06
Triggers
Email, Slack, schedules, webhooks, voice, or a chat on your website.
- Role
- Sorts the town hall's inbox, drafts replies
- Model
- mistral-large · eu
- Knowledge
- procedures.pdf · faq
- Tools
- imap · register · calendar
- Guardrails
- no email sent unapproved
- Triggers
- every 15 min · inbox
- Budget
- €40 / month
6model providers, 2 of them European clouds
6channels for approvals and digests
6interface languages
0lines of code to set up an agent
// Features
What makes it work on a Monday morning
Demos are easy. These are the parts that let an agent run every day, inside an organisation that answers to someone.
Where people reach it
Talk to it
Call an agent and hear it answer, in the language you pick. Approve an action by voice.
On your website
One script puts an agent on your site or in your app, for visitors or signed-in users.
Approvals where you work
Questions and approvals reach Teams, Slack, Telegram, Discord, email or a webhook.
From your own assistant
Reach your agents and workflows from Claude, ChatGPT or Le Chat.
Who stays in control
Approvals you can edit
See the exact action in plain rows, change any field, then approve or refuse with a reason.
Changes are proposals
When an AI reconfigures an agent, the change waits as a diff until a person applies it.
One-click stop
Halt every agent and workflow in the workspace at once while you look into it.
Hand-off rules
Decide which agents may pass work to which, and how long a chain may get.
How you know it works
Test before you change
Real requests are replayed against a new prompt, tool or model and graded pass, partial or fail.
Try a cheaper model
A candidate model runs on the test set and is promoted only if it clears the bar you set.
Automatic diagnosis
Runs are analysed into findings, with evidence, that tell you what to fix first.
Value you can defend
The dashboard counts outcomes a person accepted, next to what they cost.
What it costs, and when it fails
Spending caps
Limits per workspace, agent or user, per run, day or month, with a warning first.
Every tool call priced
Paid tools are priced per run, not just model tokens.
Retry from the failed step
A failed run picks up where it stopped, without asking for the same approval twice.
Traces to your tools
Send each run to Langfuse, Braintrust, Arize Phoenix or any OpenTelemetry tool.
- trigger Email from Martin SA: “quote for 40 desks”
- model Read request, plan 3 steps
- tool crm.lookup → customer since 2021
- tool pricing.quote → 40 × €189
- guard Discount above 10% needs approval
- you Claire edited discount 15% → 12%, approved
- done Quote sent, filed in CRM
cost€0.04
// On the record
Every run leaves a record you can read
Prompts, tool calls, approvals and cost, step by step. When something goes wrong you follow what happened instead of guessing, and the result comes back as a table, a diff, a map or a chart rather than a wall of text.
- Each step timed and costed
- The rule that applied, and why
- Results as tables you can export to CSV
// Workflows
Fixed steps where you want them, judgement where you need it
Some work should run the same way every time. Workflows do that: signed webhooks, schedules, traffic caps, export and import. A workflow can call an agent for the step that needs reading or writing, and an agent can start a workflow.
- Webhook
- Validate
- Agent
- Approval
// Your model, your key
Pay the provider directly. Change it when the market moves.
Your API key lives in your workspace and your agents call the provider you chose, with no markup on tokens. Pick a model per task, and keep it in Europe when you need to. Or skip the key: we also run models privately on our own infrastructure in Europe, so your prompts never go to a third-party provider.
- Tessellion private models
- OpenAI
- Anthropic
- Mistral
- Google Gemini
- OVHcloud
- Scaleway
- Any OpenAI-compatible endpoint
// Use cases
Whatever your line of work
What organisations put agents on first. Yours is probably close to one of these.
Sales
Lead qualification, quotes, RFP answers, meeting preparation and follow-up.
Strategy & marketing
Competitive and technology watch, campaign drafts, social content, visuals.
Operations & service
Customer support, back-office questions in plain language, incident reports.
Finance, legal & compliance
Expense and mail sorting, payment chasing, regulatory watch, clause review.
Human resources
Job ads, CV screening, interview summaries, onboarding help.
Procurement
Supplier quote comparison, price and raw-material watch, specifications.
Audit & expertise
Audit preparation, report drafting, diagnostic assistants.
IT & development
Internal support, documentation, code review, questions over your systems.
Learning
Course creation and upkeep, diagrams, training videos, virtual coaches.
// How we work
Small steps, in production
No twelve-month programme. Something works within weeks, or we tell you it doesn't.
01 1–2 weeks
Assess
We map your workflows, systems and data, then pick the two or three tasks where an agent pays for itself first.
02 4–6 weeks
Ship one agent
One task, wired into your real systems, tested before go-live, with a person approving what leaves.
03 ongoing
Train and hand over
More agents, more integrations, and your team trained to run them. We are meant to become optional.
// Founders
The people who built it, and run it
Tessellion is not a product we resell. We wrote it, we operate it, and we know exactly where its edges are.
Georges Coudray
Co-founder · Sales Director
Leads business development. Your first contact to find the tasks where an agent pays for itself.
Jean-Claude Antonio
Co-founder · CTO
15+ years building and leading engineering teams across startups, scale-ups and enterprises. Has designed several software platforms and taken them into production, from architecture to operations.
// Contact
Tell us what is slowing you down
A short call is usually enough to tell whether you have an agent use case, an IT problem, or neither. We will say so either way.
Or write to contact@tessellion.com