Videos
Our analyses of what makes an artificial intelligence project hold — or fail — inside a company: the conditions, the costs, the responsibilities. Each video starts from a concrete case and ends with what remains to be decided.
In-depth
AI agents in production: the 3 conditions
7 min 32
What keeps an agent from leaving the pilot phase is almost never the model. This video sets out the three conditions to write down before production — explicit permissions, complete logging, human validation — and applies them to a concrete case: following up on pending quotes.
Watch on YouTube
AI agent: 3 cost lines the quote leaves out
7 min 16
An agent's price is not in the licence. Operations, data quality and skills: three lines that recur every month and decide the real cost. This video explains how to budget them honestly.
Watch on YouTube
Your agent executes. Who authorised it?
7 min 11
An assistant answers; an agent writes into your systems. That difference changes accountability, scope and traceability. This video shows what an agent actually commits, and under which conditions an organisation can carry it.
Watch on YouTube
The model is not the first question
6 min 33
Most projects start with “which model”. This video proposes the reverse order: the data, the use case, the named owner — and then the model, if one is needed.
Watch on YouTube
Your three data questions: quality, access, rights
6 min 04
A table can be correct while the view the agent reads is no longer current, and nobody sees it. This video sets out the three questions to settle before any model choice — freshness at the level of the object actually queried, written permissions, a named owner — then says plainly what they do not cover: running cost, team skills and the model.
Watch on YouTube
AI agent: 4 rights nobody wrote down
6 min 23
A service account created for a forgotten report keeps reading the database, and the agent that reuses it inherits rights nobody can enumerate. This video contrasts two ways of opening access, describes the three levels actually used and gives the four lines that bring an agent back to its scope: the source, the columns, the level, the name of the person authorising. It also says what that list does not cover — the quality of the data read, the correctness of the answer, the need for a review.
Watch on YouTube
AI agent: the 4 log lines your auditor will demand
7 min 56
A technical log proves that a query ran, not on whose behalf or under which authorisation. This video separates the two logs, sets out the three questions an audit will ask and gives the four lines that make the trace demonstrable: the timestamp, the agent's identity, the exact object touched, the decision with the name of whoever validated it. It then covers what makes a log hold — append-only writing, a decided retention period, an export readable without the tool — and says plainly what it does not prove.
Watch on YouTube
AI agent: 3 actions your agent will never approve itself
7 min 15
What changes on the day the agent is right is not accountability: it is the time spent validating. This video separates reliability, measured on a sample, from responsibility, which rests on a dated act. It names the three families of actions that keep a signature — payment, commitment, sending to a third party — gives the four conditions that make the arrangement hold, then says plainly what validation does not fix.
Watch on YouTube
On-prem LLM or API: 3 things decide
7 min 13
On-premises does not close the question of what leaves: it moves it. The video names the three real exits — the prompt, the document excerpt, the corrections that make the model useful — then compares what is comparable over three years. It ends on independence where it is actually decided: in the update, and in what you take with you when you leave.
Watch on YouTube
Data sovereignty: 3 questions, not a slogan
6 min 26
Data sovereignty is almost always confused with where data is hosted. This video separates the two words that dossiers merge, hosting and being sovereign, then asks the three questions that decide: where the data is stored, who can read it, and which law applies. It gives the six lines the contract must carry, names the three recurring confusions, the three-step method, and what sovereignty does not guarantee.
Watch on YouTube
AI total cost of ownership: 3 cost lines, not a price
6 min 35
An agent platform’s price is signed once; its cost is paid every month. This video separates the two words, then prices the three lines a quote never carries — operations, data, skills — with four costing rules and a checklist to run against an existing quote, line by line. It also says what a total cost does not cover: means can be counted, the value produced is only observed in use.
Watch on YouTube
A proof of concept that decides, not one that demos
6 min 25
A pilot that ends in a demo decides nothing. This video sets out the other arrangement: the five pieces written before the first line of code, three signatures — IT, business, security — and the order of the five steps, from written scoping to a signed decision. It also says what a pilot cannot settle: adoption, the value produced, and the running cost over time.
Watch on YouTube
7 roles, 6 refusals: where an AI agent acts
6 min 43
An assistant proposes; an agent writes into the system. The video separates those two verbs, sets four written conditions — an output a reviewer can check in a minute, a bounded write scope, a write that can be corrected or must be signed, a named trace — then applies them to thirteen common functions. Seven pass, from customer service to reporting; six steps, from granting credit to answering a regulator, stay human, and the video says why.
Watch on YouTube
CIMA reporting: what AI prepares, what stays signed
6 min 50
An insurance report is often signed before it is explained. The video names the four pieces an variance must leave behind — a dated source, the calculation and its version, the reconciliation, the explanatory sentence — then follows one variance from detection to filing, saying at each step whether the agent acts or stops. It sets out what an agent does not do, the three signatures that stay human, and what the method does not say.
Watch on YouTube
Telecoms: the cause before the complaint
6 min 51
A telecom operator often learns about an outage from the customer's call. The video separates detection, which compares a measurement with a threshold, from diagnosis, which looks for the cause across the equipment, the link, the configuration and the ticket, then crosses four signals already present in the information system. It follows one anomaly through five steps, says at each step whether the agent acts or stops, names the three roles accountable for service, and closes on what the method does not say.
Watch on YouTube
Utilities: the unpaid bill starts at the meter reading
7 min 29
At a water or power utility, meter reading and dunning sit in two departments that never talk to each other, and one assumes the other without ever checking it. The video names four discrepancies the data already holds — zero consumption on an active meter, an estimate carried forward, a connection still drawing on a closed contract, a payment received but not allocated — then follows a dunning case through five steps, saying at each step whether the agent acts or stops. It closes on what the agent never decides: correcting an index, granting a payment plan, cutting supply, qualifying fraud.
Watch on YouTube
Import CEMAC: 5 documents, 3 gaps that stop the file
6 min 10
In a CEMAC port, an import file rarely stalls on the vessel. The video takes apart the five documents that make it up — bill of lading, commercial invoice, packing list, certificate of origin, import declaration — and names the three gaps that stop it: names, weights, dates. It then follows one file through five steps up to filing with the customs broker, saying at each step whether the agent acts, flags, or stops.
Watch on YouTubeShort formats
The split that protects your data
46 s
The choice is split by data, not by doctrine, and it costs two chains to operate.
What you take with you the day you leave
51 s
Dependence changes sides: an installed model with nobody able to reinstall it is not an asset.
What an AI agent is not allowed to touch
38 s
Scope is not a good intention: it is a written list.
What is not logged should not run
41 s
The auditor's question always comes: who, what, when, on which data.
Human validation is not temporary
38 s
On a sensitive action, a person's approval does not fade with time.
Three costs an AI agent quote leaves out
56 s
Operations, data, skills: what comes back every month.
No figure holds without a scope
44 s
A number without a written scope is an opinion, not a budget.
An assistant writes an email. An agent writes in your ERP.
50 s
At the end there is a write — with an author, a time and a consequence.
What you pay every month after go-live
53 s
Operations and skills do not stop at acceptance.
Four stages, only one is about the model
50 s
Scoping, data, industrialisation — the model comes last.
The project budget does not cover year one
53 s
What costs is what continues after signature.
What the agent reads is not current
48 s
The table is right, the view is frozen: the agent answers without nuance.
A service account reads everything
48 s
Permissions are not settled in the technical layer but in a written list.
What the log gives, what the audit asks for
56 s
Four boxes the technical log never fills, and that cannot be filled after the fact.
Three mistakes that cost you the audit
55 s
Logging errors only, a shared account, no review: the auditor will challenge the successes first.
What a signature must contain before it validates
50 s
An agreement in principle is not a validation: it takes the case file, the decision and the name.
What turns a human validation into a formality
51 s
Validating without the file, without a deputy, or in a peak period: three ways to keep the form and lose the decision.
Hosting your data is not data sovereignty
55 s
Where data is hosted answers an invoice question, not the question of who can open it or which law applies.
Three confusions about data sovereignty
38 s
A flag, a label, a local server: three reassuring answers to the wrong question.
A total cost can be counted. Value cannot
40 s
Costing tells you the means, not the outcome: a properly budgeted platform can end up serving nobody.
The pilot that demos decides nothing
51 s
Two pilots look alike on screen: the one that shows and the one that settles. A pilot that cannot conclude no concludes nothing.
IT, business, security: three signatures
52 s
Three signatures rather than a consensus: IT on data and operations, the business on usage, security on access and logging. Each one can stop the pilot.
An assistant proposes, an agent writes
49 s
An assistant hands back a draft; an agent writes into your system and leaves a named trace. What decides the function is whether the output can be checked in a minute, not the model.
HR, production, reporting: the condition that holds
52 s
These three functions pass the test, on one written condition: screening a person out, stopping a line and publishing a report stay human.
What an agent never signs
48 s
An agent can reconcile, draft and propose the explanation of a variance; it does not choose the reserving method, does not correct an entry silently, signs nothing and does not speak to the supervisor.
Three signatures a tool cannot replace
50 s
Technical direction, accountable officer, compliance: three signatures already in the organisation chart, which an agent project makes explicit instead of routing around.
Your alarms see the symptom, not the cause
54 s
Monitoring fires on a threshold and names the symptom; diagnosis looks for the cause across the equipment, the link, the change and the ticket, and that is the work an agent can hold, provided it cites its pieces.
An agent warns; it does not correct
51 s
On network logs, an agent does not declare the root cause, changes no configuration, closes no incident and does not speak to the customer: it gathers the pieces and alerts the person on call.
Meter reading and dunning never talk
53 s
The meter reading records consumption, dunning claims a sum, and dunning never opens the reading that produced it — that is where the error reaches production.
Three roles, one file
58 s
Billing holds the indexes, collections signs the notice, the security officer controls what the agent reads: three roles already exist; what is missing is the written rule saying who decides what.
The bill of lading says one thing, the invoice another
52 s
Three gaps are enough to stop an import file — the name, the weight, the date — and none takes particular intelligence to see, only to be looked at before filing.
What the agent never signs
46 s
Tariff classification, customs value, duty payment, release: an agent that prepares an import file signs nothing, and that is exactly what makes it usable in production.
Videos are hosted on YouTube and played from YouTube: no copy is stored on this site.