AI Intelligence & Technology Monitoring
How We Build an Artificial Intelligence Watch
A watch is only worth what its method is worth. Ours is written down, applied in the same order on every assignment, and open to your review at each step.
Artificial intelligence produces an unusual amount of noise: preprints presented as products, demonstrations that never reach general availability, and version numbers that hide changes in behaviour. A method is what separates the part of that flow which affects your organisation from the part which does not.
The sequence below is the same whether the assignment is a single study or a watch running over several quarters. What changes is the depth of each step, the number of sources retained and the rhythm of delivery, all of which are fixed in writing with you beforehand.
Seven steps, in this order
No step is skipped. The early ones cost time and save far more of it later, because they prevent a watch from being built on the wrong question.
-
Understanding the need
We start with the decisions in front of you and the people who will read the deliverable. A watch designed for a technical team and one designed for an executive committee are not the same watch, even on an identical subject.
- The decisions the monitoring has to support, and their timing
- The readers of each deliverable and their level of technical detail
- What is already followed internally, so we do not duplicate it
-
Defining the scope
The scope is written: subjects included, subjects deliberately excluded, markets, languages, frequency and formats. Exclusions matter as much as inclusions; they are what keeps a watch readable when a subject becomes noisy.
- A written document validated before any collection begins
- Explicit exclusions to avoid a watch that expands without limit
- Priority levels distinguishing what triggers an alert from what waits
-
Selecting sources
Sources are chosen for the subject, not taken from a standing list. On artificial intelligence, that mixes technical publications, official registers, regulatory texts, vendor documentation and specialised professional press.
- Primary sources preferred over commentary whenever they exist
- Each source assessed for reliability, update rhythm and bias
- The list reviewed during the assignment as the subject moves
-
Collecting and qualifying
Collection is continuous; qualification is what makes it usable. Duplicates, recycled announcements, unverifiable claims and items outside the agreed scope are removed before anything reaches analysis.
- Systematic checking of dates, versions and original publication
- Announcements separated from general availability of a product
- Items that cannot be verified are flagged as such, not silently dropped
-
Human analysis supported by digital tools
Tools handle volume, deduplication and tracking. The reading, the comparison and the judgement of what matters for your organisation stay human, and every analysis is attributable.
- Automation used for collection, never for the final assessment
- Comparisons made against criteria defined with you, not generic scores
- A stated level of confidence when the available evidence is thin
-
Producing alerts, syntheses or reports
The deliverable is written for its reader, in the agreed format and rhythm. Each element carries its source and its date, so it can be verified or reused internally without further research.
- Sourced and dated material throughout every deliverable
- A short summary at the top for readers who will not read further
- A clear separation between what is established and what is our reading
-
Adjusting the set-up
We review what proved useful, what was ignored and what was missing. Sources, filters, frequency and formats are then adjusted, because a watch that never changes stops matching its subject.
- A review point planned from the outset, not improvised
- Sources that produce little are replaced rather than kept out of habit
- Frequency raised or lowered according to how the subject actually moves
Ten delivery formats
Formats are combined in a single assignment. A fast channel carries what cannot wait; a slower one carries the analysis that needs perspective.
Immediate alert
Sent within hours when an event affects your scope: a deprecation announcement, a change of usage terms, a security disclosure or a regulatory decision that moves a deadline forward.
Daily summary
A short daily digest for teams working on a fast-moving subject, listing what appeared, what is confirmed and what remains rumour, without commentary that would slow the reading.
Weekly bulletin
The most common rhythm. A structured review of the week by theme, with the changes that matter for your projects and an explicit note on what did not move.
Monthly report
A longer document with the perspective a weekly rhythm cannot give: trends confirmed or contradicted, comparisons over time, and the questions your governance bodies are likely to raise.
One-off study
A documented answer to a single question, delivered once. Used when a decision has a date and a continuous watch would not change the outcome.
Benchmark
A structured comparison of models, tools or providers against criteria agreed with you, stating the test conditions, the sources used and what the comparison does not establish.
Mapping
A readable picture of an ecosystem: developers, integrators, funders and the dependencies between them, so that a supplier choice is made with its full context visible.
Tracking table
A maintained table following defined items over time: model versions, licence conditions, regulatory milestones or internal projects, updated at the agreed frequency.
Custom research
A targeted search on a precise point, when a team is blocked on a question that would take them days and takes us a shorter and better-sourced path.
Strategic support
Regular working sessions with your team to discuss the material delivered, prepare a committee or challenge an option before it becomes a commitment.
Five parameters set with you
These five choices define the assignment. They are written into the scope document and revisited at the review point.
- Frequency
- From an immediate alert to a quarterly report. The right frequency is the one your organisation can act on; anything faster produces material nobody reads.
- Languages
- English and French as standard, for sources and for deliverables. Other languages are added when a market, a research community or a regulator justifies it.
- Sources
- Selected subject by subject and reviewed during the assignment. Primary sources are preferred; commentary is used only when it adds analysis that the original does not carry.
- Geographical areas
- The markets and jurisdictions to follow, which on this subject rarely match: a technology can be developed in one region and regulated in another that concerns you more.
- Delivery channel
- Electronic mail, a shared document, a structured file for your own tools, or a working session. The channel is chosen for how the material will actually be used.
Our quality rules
- Every element is sourced and dated, so that any reader can go back to the original.
- An announcement is never reported as an available product until availability is verified.
- What we could not verify is stated as unverified rather than left out silently.
- Comparisons state their criteria, their conditions and what they do not allow you to conclude.
- Our reading is separated from the facts, so you can disagree with the reading and keep the facts.
Questions about our method
Do you use artificial intelligence to produce the watch itself?
Digital tools, including automated ones, are used for collection, deduplication and tracking volumes no team could read manually. The qualification of sources, the comparison and the final analysis are human work, and each deliverable is attributable to an analyst who can defend it.
How do you avoid delivering a simple press review?
By filtering against a scope agreed in advance and by reporting what changed rather than what was published. A great deal of what appears on this subject repeats itself. A useful deliverable is often shorter than the flow it comes from, and says explicitly when nothing significant moved.
Can you work from our internal documents?
Yes, when it makes the watch more precise: an architecture note, a supplier contract or a project brief helps target the sources that matter. Internal material stays confidential, is used only for your assignment and is never reused elsewhere.
What happens if a subject turns out to be less active than expected?
We tell you, and we propose a change rather than filling deliverables with marginal items. That may mean lowering the frequency, widening the scope to an adjacent subject, or switching from a continuous watch to a periodic report on the same question.
Tell us what you need to monitor
Describe your subject, your markets and your decision timeline. We study every request individually before proposing a monitoring set-up.
Every request is reviewed confidentially. No commitment is required to discuss a scope of work.