Business Intelligence & AI
Summary
TLDRLe script présente une nouvelle méthode d'analyse des entreprises, développée par Alis, qui utilise l'intelligence artificielle neuro-symbolique pour la prise de décision. Après cinq années de recherche et développement, cette approche permet de détecter des risques et des opportunités de manière automatisée, en identifiant des signaux d'alerte précurseurs. L'IA spot sept fois plus d'insights que l'exploration humaine, aidant ainsi les analystes à comprendre les causes profondes des problèmes et à élaborer des plans d'action correctifs. L'objectif principal est de repérer des opportunités même dans les unités commerciales en difficulté et des points d'amélioration chez les unités réussies. La solution d'Alis vise à résoudre les problèmes des outils de reporting actuels, en offrant une vision consolidée des données et en mettant en avant l'importance de l'analyse déconsolidée et de la dynamique des données pour les entreprises.
Takeaways
- 🚀 Une nouvelle méthode d'analyse des entreprises est présentée, basée sur une intelligence artificielle neuro-symbolique pour la prise de décision.
- 🔍 L'intelligence artificielle permet d'automatiser la détection des risques et des opportunités, ainsi que la découverte de signaux d'alerte précurseurs.
- 🌟 L'objectif principal est de repérer des opportunités même dans les unités commerciales en difficulté et des points d'amélioration dans les unités réussies.
- 📊 Les outils actuels de reporting sont critiqués pour ne pas offrir d'analyse multidimensionnelle ou de déconsolidation des données.
- 🤖 L'intelligence artificielle générative est combinée avec l'IA symbolique et les graphes de connaissances (KGE) pour identifier la cause profonde des problèmes.
- 🔑 L'algorithme de priorisation Mark of Shame est utilisé pour classer les problèmes business en fonction de leur impact et d'autres critères.
- 🧠 Une solution basée sur l'IA est proposée pour aider les analystes à construire des plans d'action, plutôt que de simplement générer des rapports.
- 📈 Des trackers spécialisés sont utilisés pour analyser les données financières d'entreprise de manière systématique et multidimensionnelle.
- 🗣️ Une interface de speech-to-text est envisagée pour simplifier l'interaction avec l'application et aider à identifier les problèmes non résolus dans les rapports.
- 🔄 L'IA est utilisée pour upskill les analystes et les aider à améliorer leur expertise en matière d'analyse financière.
- 🎯 L'objectif de la solution est de devenir un outil d'analyse complémentaire aux méthodes de reporting traditionnelles, offrant une vision plus large et une aide à la prise de décision.
Q & A
Quel est le nouveau modèle d'analyse des entreprises présenté dans le script ?
-Le nouveau modèle d'analyse des entreprises présenté dans le script est la méthode d'analyse des risques et des opportunités. Elle utilise une intelligence artificielle neuro-symbolique pour la prise de décision, permettant de découvrir automatiquement les signaux d'alerte précoce dans les données de l'entreprise.
Quelle est la durée de recherche et développement pour cette nouvelle méthode d'analyse ?
-La durée de recherche et développement pour cette nouvelle méthode d'analyse est de cinq ans.
Quelles sont les antécédents de l'orateur en matière de science des données ?
-L'orateur a travaillé dans le domaine de la science des données pendant cinq ans avant de rejoindre Alis. Avant cela, il était Directeur Financier (CFO) dans le secteur public en France. Il a également publié beaucoup de contenu lié à la science des données sur LinkedIn et Medium.
Comment l'intelligence artificielle peut-elle aider dans l'analyse des performances d'entreprise ?
-L'intelligence artificielle permet aux organisations d'automatiser la détection et la découverte des risques et des opportunités, ce que l'on appelle les signaux d'alerte précoce. En utilisant cette méthode, on peut identifier sept fois plus d'informations pertinentes que par l'exploration humaine.
Quel est le rôle d'un analyste dans le cadre de cette nouvelle méthode d'analyse ?
-Le rôle d'un analyste est de comprendre la cause fondamentale du problème. Grâce à l'intelligence artificielle, les analystes peuvent automatiquement détecter les anomalies dans les données de l'entreprise, puis trouver la cause racine et bâtir un plan d'action correctif.
Quelles sont les principales limitations des outils de reporting actuels ?
-Les outils de reporting actuels sont principalement basés sur une analyse consolidée des données, ce qui masque les anomalies et les insights pertinents. Ils ne permettent pas une analyse multi-dimensionnelle systématique et ne répondent pas aux questions sur la conversion d'insights en plans d'action.
Pourquoi les entreprises ne détectent-elles pas proactiquement les anomalies avec les logiciels de business intelligence actuels ?
-Les logiciels de business intelligence actuels ne sont pas conçus pour détecter proactiquement les anomalies. Ils laissent aux analystes la responsabilité de faire une analyse multi-dimensionnelle manuelle, ce qui est impossible pour l'homme lorsqu'il s'agit de données à plusieurs dimensions.
Quelle est la solution proposée par Alis pour améliorer la détection d'insights ?
-Alis propose une solution qui utilise de l'intelligence artificielle, combinant l'IA générative et l'IA symbolique sous forme de graphes KGE, pour trouver la cause racine et bâtir un plan d'action correctif. Cette méthode permet de repérer jusqu'à sept fois plus d'insights qu'avec les méthodes traditionnelles de reporting manuel.
Comment la solution d'Alis aide-t-elle les entreprises à réagir plus rapidement que leurs concurrents ?
-La solution d'Alis permet de détecter les anomalies et les risques plus tôt que les concurrents, grâce à une analyse multi-dimensionnelle systématique et automatisée. Elle offre également des insights sur les actions correctives à prendre, permettant aux entreprises de réagir rapidement et de s'adapter à la situation.
Quelle est la vision globale de la solution proposée par Alis pour l'analyse des entreprises ?
-La vision globale de la solution proposée par Alis est de trouver des opportunités même dans les unités d'affaires en difficulté et des points d'amélioration dans les unités prospères. Elle vise à aider les entreprises à répondre aux questions critiques et à suivre les KPIs pertinents grâce à une analyse plus approfondie et à une prise de décision plus éclairée.
Quelle est la méthode proposée par Alis pour améliorer la qualité des données et la pertinence des insights ?
-Alis utilise une approche de traitement par lots pour détecter les insights chaque mois en utilisant des données financières de qualité supérieure. Ils ont également développé un algorithme de priorisation pour gérer le flot d'alertes et pour identifier les insights les plus importants en fonction de la stratégie de l'entreprise et de la situation urgente.
Outlines
📝 Présentation de la nouvelle méthode d'analyse d'entreprise
Anthony, le responsable des opérations d'Alis, présente une nouvelle méthode d'analyse d'entreprise basée sur l'intelligence artificielle. Après cinq années de recherche et de développement, cette méthode, appelée "reporting des risques et des opportunités", permet de détecter automatiquement les risques et les opportunités, ou les signes d'alerte précoce, dans les données d'une entreprise. Cette approche permet de découvrir sept fois plus d'informations que l'exploration humaine. Le rôle de l'analyste est alors de comprendre la cause profonde des problèmes. L'intelligence artificielle, en utilisant des techniques génératives et symboliques, aide à identifier la cause racine et à élaborer un plan d'action correctif.
🔍 Problèmes avec les outils de reporting actuels
Le script souligne les problèmes avec les outils de reporting actuels, qui sont principalement basés sur une analyse consolidée de données. Cela masque les anomalies et les insights importants qui ne sont pas visibles dans l'analyse consolidée. Les outils actuels ne permettent pas une analyse multidimensionnelle systématique, ce qui est essentiel pour identifier les problèmes. Les outils de reporting ne fournissent pas non plus de KPIs sur le nombre d'insights identifiés et convertis en plans d'action. L'intelligence artificielle peut jouer un rôle crucial pour résoudre ces problèmes en détectant proactivement les anomalies et en aidant à répondre aux questions critiques.
🤖 Solution Alis pour la détection d'insights
Alis propose une solution qui utilise l'intelligence artificielle pour une analyse multidimensionnelle des données. Leur outil peut identifier jusqu'à 85% de plus d'insights que les méthodes traditionnelles de reporting manuel. Cependant, l'afflux d'insights peut être overwhelming, nécessitant une algorithme de priorisation pour aider les utilisateurs à identifier les insights les plus importants. Alis a développé un algorithme de priorisation qui prend en compte la stratégie de l'entreprise, l'urgence et la faisabilité pour aider les utilisateurs à agir efficacement.
💡 Amélioration du rôle de l'analyste avec l'IA
L'IA permet de libérer les analystes de la tâche de générer des rapports et de les aider à trouver des informations pertinentes. Les analystes peuvent ainsi se concentrer sur la résolution des problèmes et l'élaboration de plans d'action. Alis utilise une combinaison d'IA générative et d'IA symbolique pour aider les analystes à comprendre les causes profondes des problèmes et à construire des plans d'action. L'IA sert à recommander des actions et à poser des questions, aidant ainsi les analystes à améliorer leur expertise et à prendre des décisions plus éclairées.
🌐 Vision de l'IA dans l'analyse d'entreprise
La présentation d'Anthony met en évidence la vision d'Alis sur l'IA dans l'analyse d'entreprise. Il souligne que l'IA peut aider à détecter des problèmes de manière systématique et à recommander des actions, mais que les humains sont mieux placés pour prendre des décisions. L'IA peut également aider à améliorer l'expertise des analystes, en particulier dans les entreprises où il y a un taux élevé de rotation. Le but d'Alis est de créer une application simple et facile à utiliser qui automatise la génération de rapports et permet aux analystes de se concentrer sur la résolution des problèmes.
Mindmap
Keywords
💡Intelligence artificielle neuro-symbolique
💡Détection d'anomalies
💡Graphes de connaissances (Knowledge Graphs)
💡Analyse multidimensionnelle
💡Priorisation des alertes
💡Consolidation des analyses
💡IA générative
💡Détection proactive
💡Rapports de risques et opportunités
💡Analyse granulaire
Highlights
提出了一种新的商业分析方法,通过多年的研究和开发,专注于风险和机会报告。
使用神经符号人工智能进行决策,通过自动化检测公司数据仓库中的异常,发现风险和机会的早期警告信号。
与人类探索相比,该方法能够发现7倍多的洞察力,转变分析师的角色,更注重理解问题的根本原因。
利用生成型人工智能和符号型人工智能的结合,通过知识图谱(KGE图)来发现问题的根本原因。
该方法不仅适用于问题业务单元寻找机会,也适用于成功业务单元的改进点分析。
提出问题:为什么2023年的商业智能软件还不能主动检测异常?强调人工智能在大规模并行处理中的作用。
指出当前报告工具的问题,如合并分析导致的信息补偿和隐藏,以及缺乏动态技术分析。
强调需要系统性多维分析,以及人工智能在自动执行此类分析中的必要性。
讨论了当前报告世界的状况,指出传统报告工具只能提供全局视角,而无法发现被隐藏的洞察力。
提出了解决方案,使用机器学习和监督学习进行异常检测,并结合生成型人工智能和知识图谱。
解释了如何通过人工智能帮助用户构建行动计划,而不是完全自动化决策过程。
讨论了数据质量问题对报告的影响,以及如何通过工具解决这些问题。
强调了人工智能在增强分析师技能和专业知识方面的应用,特别是在人员流动率高的领域。
提出了一种新的人工智能应用方法,即通过生成型人工智能辅助分析师提出问题和建议行动。
讨论了人工智能在决策支持中的角色,强调人类在最终决策中的重要性。
提出了一种新的报告工具,旨在通过人工智能自动化和系统化地提供风险和机会的早期警告。
强调了人工智能在帮助用户理解和应对复杂业务问题中的潜力。
讨论了如何将人工智能与业务分析和项目管理相结合,以提高决策效率。
Transcripts
started
recing hey Anthony welcome welcome to
this event and happy you could make it
um thank you so um I'll let you start
okay so today I would like to talk about
um a new method uh to do business
analysis uh at Alis we have been working
on it uh for many years five years of
research and development and uh it is a
new way to do um we call it risk and
opportunities
reporting uh to to talk more about me so
I am a chief IE officer at
Alis uh I have been uh doing data
science for five years uh before that I
was a CFO so Chief Financial Officer in
the public sector in France and I have
been uh uh publishing many data science
related content over LinkedIn and medium
and uh since then I it has been two
years that I I have joined Alis on this
project on this
adventure uh so uh what we are
using we are using neuros symbolic
artificial intelligence for decision
making uh what we have understood
is
that uh for for to do a business
performance analysis the artificial
intelligence uh allow an organization to
automate uh the uh detection the
discovering of risk and opportunities
what we call early warning signs and by
doing it with method that we use we spot
seven s times more insights than by
human exploration so the question is no
longer where to look
at it
is the why question the analyst and the
artificial intelligence sport Insight
needs to understand the root cause of
the problem so it is a new role for them
because because we use artificial
intelligence to automatically detect
anomalies within uh the companies data
warehouse uh the the new role for the
analyst is to find the root cause and
for that what we are
leveraging are artificial intelligence
in form of generative EI combined with
symbolic EI in form of kge graph uh so
it is it is to find the root cause and
on you you you spot the root cause uh
all method allow also to help the
analyst uh build a corrective action
plan so it is the old question or to
react quicker than your competitors and
to to to to act on those uh root cause
uh that uh our method allow to to detect
so our
approach its main goal is to find
Opportunities even in a trouble business
unit okay and
Improvement points in success successful
ones uh because even your best uh
business units uh have Improvement
points and I will show how how we are
doing
that okay what I ask our clients is
this uh why in
2023 no business intelligence software
is able to proactively detect
anomalies why is it is it the the end
user the analyst
to
manually uh that he has to manually uh
do a
multi-dimensional to do an an
analysis why we don't use artificial
intelligence with massively uh parallel
processing to detect proactively
anomalies this is the first question and
usually what I ask also to clients is
this from your reporting uh we can take
for example September do you know how
many insights were identified and
generally the answer is no because this
is uh a kpi it is not a kpi that is uh
um analyzed by
companies and even imagine if we we we
we we we respond to this question yes
uh how many of those insights did you
convert into corrective action plans and
we can go even further uh from those
corrective action action plans what is
the trend curve of Correction rate of
insight into action plans those those
question should be the uh kpi of any
Reporting System yet currently in many
companies they are not uh answered they
are not followed and it should be the
the main goal of any reporting tool
currently and we think at Al thisis that
artificial intelligence uh is a parament
is essential to answer those question
and help a business answer those
question and follow these uh main C
Capi so what is the main problem of
current reporting tools
the main issue is
Consolidated
analysis all current
reporting are based are built on
Consolidated analysis of
data why is it a problem I will show you
a simple example right here we have two
continents okay uh what you can see at
the solidated view is that continent two
it's in a worse situation than continent
one but but if you DEC consolidate okay
what you can see is that for continent
one that seem better off there is an
anomaly that is compensated by other
countries for this continent this is the
problem of consolidation because
consolidation you have multiple uh
mathematic IAL
phenomenon like Simpson Paradox
like mix when you mix there is a
compensation between negative value and
positive value but even in positive
value you have a compensation uh if you
have an extreme positive value and other
value are smaller in comparison this is
Extreme positive value will be hidden
will be masked by Consolidated uh data
by Consolidated value value this is this
this seem a simple problem but it is not
because right now every um reporting
tool uh um uh consolidates data and
don't have a a comprehensive and
systematic way to
do uh a deconsolidated
analysis and it is also true in time de
in time dimension
right here I am showing you an example
where um I am analyzing the current year
and I compared it to the previous year
so I am in 23 and I am comparing 22 I am
studying what we call in b b business
finance the year to date so I am
analyzing the
situation uh we are in October so until
until October and I and I comparing the
situation from the pre previous year so
right here in my reporting what I
observe is that I am better off in
203 comparing to 22 but this
Consolidated time analysis uh behind
that you can have multiple
situation
okay it might be that the situ the
current situation it explained by an
anomaly a positive anomaly uh uh right
here you can see an outlier a quarter
outlier that is explaining the current
better off situation of my current here
to date but it can also be that uh my my
my positive trend uh is degrading is is
decreasing okay but you can Al also have
a a tendency reversal so you wear better
off you were improving during uh until
the the Third quarter and after you are
degrading those multiple and I I only
took a three example but there is many
many configuration uh all those example
show that you need to
deconsolidate even time analysis so you
need to add Dynamic technical analysis
within your reporting to analyze the
trends uh and the dynamic of your data
so and another issue that is uh I am
stacking stacking up problems so for you
to understand what is the current issue
of
reporting another issue is that to have
a an Insight detection solution you need
to do a systematic multi-dimensional
analysis okay I will show you an example
right here I am taking the client uh
dimension in line I have product
but every of those line are
Consolidated okay so here the situation
appear appear to be good okay all all
are positive all is fine okay I am
taking the now the product dimensions in
every line of my client dimension for
the product Dimension all are
Consolidated and
positive but what
happen if I combine the two with those
same data
here with only two Dimension it is it is
only to illustrate the necessity of
multi-dimensional analysis with only two
Dimension I can see I can spot problems
and now imagine imagine to do that with
a four four five six seven eighth
Dimension uh right now every most of
multinational companies big company have
built uh olap cubes a multi-dimensional
data warehouse so they are able to do
those multi-dimensional analysis but it
is impossible for human Okay human can
can only
do
uh up to three dimension on a spre
spreadsheet spreadsheet so so you need
artificial intelligence to do
this in a systematic and automatic
manner
yet not many many companies not many
solution propose that okay and yet yet
it is necessary to sport
insights okay right now what is the
situation of reporting of the reporting
world what current reporting tool can
see is this it is a global vision a
reporting a Consolidated View Vision
what we call a top
approach so what they spot is what we
call flush head flush head are rotten
Branch they are problem that are so
important that they are coming
up the they are affecting the
Consolidated
view okay but when you spot them as a
Consolidated level it is too late
because what you you should have done is
to spot
them when they first appear before they
they were so big that they were
affecting the Consolidated View and what
those reporting doing Consolidated
analysis are not showing you what they
can't show
you are pocket of insights and golden
nuggets what we call them it is meth
methaphor metaphor but those are insight
that are impossible to spot at a
Consolidated view why because there is
there is compens compensation effect
they are they are concealed within your
reporting you can't spot them as a
Consolidated
View and be because because you don't
look at the at the most granular optimal
level of data analysis you can't spot
those insights and what we found with
our solution with with our experimental
uh uh work is that
85% of insights are currently not found
by traditional mean of reporting by
manual mean of
reporting when you use artificial
intelligence to do that in a systematic
way you find up to seven time more
insights more anomalies within uh for
example a business reporting uh for a
large corporation
so it is like we we are comparing it to
a Iceberg of urance I don't know if you
know the the management auor that told
about that but it is like uh you are
uring most of the problem that you have
okay
so how can you tell that the approach
that we are proposing and and I think we
are not the only one to propose um to uh
to propose a solution of inside
detection but how can you tell that our
solution is better than others
okay uh there there are three
criteria like I said I think that you
need a multi-dimensional analysis
because no not no human and and no human
are able to do it and you need
artificial intelligence to do that in a
systematic way because you need a
multi-dimensional analysis you need to
compare product geography a time uh
region uh Etc to spot anomalies uh to to
detect early warning sites to act on
them quick and and and quicker than your
competition but what happen when you do
that is that you are overflowed by alert
okay because you are sporting seven time
more Insight the user will be
overflowing by alerts and anomalies and
what you need to
introduce is some sort of priorization
so you need to build up a priorization
algorithm what what what what will be
within this priorization
algorithm uh are for example the
strategy of the
company uh some notion of urgency some
notion of um feasability
some notion of whatever we built we
built with an University uh a
priorization algorithm that we can
personalize to the company Focus or
strategy because you need to tell the
user with this uh what are insight that
are uh I high priority level what are
the most impactful insights
and what you do what you need also to do
you need to to find the most optimal
granular level of analysis and it is uh
more simple to say it than to find it so
you need a
method you need to find the right level
of analysis that you can
compare across time okay because if you
go too low to too um too below too below
certain
fesal you won't be able to compare it
over time so you need to find a perfect
granular level to do your
analysis and artificial intelligence can
help also with
that
so so those what I I just explained
those are the problem of current
reporting what we have building what we
have built at Alesis is the solution to
to solve those problem and to um to to
give tools to a company to have an early
warning side detection
tool
so this in this uh now I will talk about
our data quality
approach right now our solution is based
on the use case for business finance so
we use uh Finance
data to detect
anomalies so it is a a batch a batch
processing uh uh tool right now for to
detect Insight every month because
Finance data is every month uh and why
why we we chose this this this approach
it's because Finance data uh generally
is the mo is the most uh is um its
quality is better uh because there there
are lot of um regulation around Finance
data because you got the taxpayer it
wants it's money you want Auditors
Auditors want certain uh certification
around run Finance data shareholders
also want a good quality of Finance data
because they they want to know where to
put their money okay
so for the what for the first part of
detec detecting inside we use Finance
data and we found Insight what our tool
is
proposing is to find the causality of
what Insight because Insight are just
symptoms okay there will be anomalies
within uh data warehouse uh so an
anomalies of
symptoms is not possibly a problem you
need to find uh the causality the why
for you to know if you need to act or no
so right here we are helping the analyst
in in his new role because this is a new
role for analyst most analyst in large
corporation are only doing the first
part the what but the what like I just
said we automate it with artificial
intelligence so we need the user in in
his in his new role and I we think that
it it new role is is in more it has more
value much more value than the first one
the second one is much more complicated
and and right here what we use is
generative
a but not just only generative a we have
built rack system so retrial augmented
generation that is a system that use use
data to help the Genera VI generate its
content and we where we get our data
it's from operational data so the
problem with operational
data is that you you you got quality
issues some talk about data swamp in
many companies because you got e genous
data quality inconsistency in accuracy
uh you got because if you take also
external data you you will got biased
one-sided aliden data you are in a VCA
World okay VCA it means a complex word
in certain uh so generative VII will be
here only to to to give you
methodologies to give you uh what is
currently
available it will we are here in using
artifici icial intelligence to to build
a collective intelligence so it is to
help the user build an action plan and
not build for them the action plans
because here it is too complicated to
automate so artificial intelligence is
only able to help the user find
information and what we like to do is
that here we are more in the data
science size and here in this part we
are more in a decision science s and
decision science sites you got a
psychology you got uh information that
are not present in the data in the in
the in the internet in the data way
house in instructure data
Etc one question are you going deeper
into the algorithms so I can ask my
question here uh I I will uh in the
subse slide I I will go deeper yes so I
would be
patient I am doing uh in the step by
step uh
manner okay so what we use so I would
like to to show you our stack of
artificial
intelligence so we use machine learning
supervis and nonsupervised uh machine
learning uh for the detection sit okay A
Time series um we use Al EJ boost uh to
predict values from uh
the the detection in the detection part
of our
solution like I said we also use
generative VI so what we do but I I will
show
you right here we combine generate VI
with chge graph a chge graph that are
built uh on operational data from our
clients and we also use chge graph uh to
inform uh the generate VI of some sort
of methodologies
uh for me it is an alternative of
fine-tuning i i i i much more better
build U an effective sge graph uh that
gives uh that can uh when you have an an
ontology or operational data than fine
tuning a model because fine tune uh you
need with time you need to to retrain
your model Knowledge Graph is more
easier to up toate to to prove over time
and what we have also we have uh some
parts of reinforcement learning the
reinforcement learning part it's more
for the super priorization the
priorization algorithm that we use and I
what I will show I will show the actual
app just after this
line This is the data science side okay
and what we have on
top are symbolic artificial
intelligence
because our prediction from machine
learning supervis
nonsupervised use in the detection sites
a rule engineering expert system so we
combined a connectionist with a symbolic
artificial intelligence and I will show
you just after this
slide uh how how we do we we did that so
and the second part is decision
science we have a natural business
analysis explainable intelligence that
explain what is happening in orap but in
why this insight has been prioritized
Etc and we have what I have just said a
virtual collective intelligence what we
call a process intelligence that helps
the user build his action plan so right
here we have an escalation a hierarchy
of
explainability because those part of at
the bottom are not explain are not that
much
explainable and more you you come up
this uh line more they are more
explainable so we use symbolic EI to uh
to have an
explainability within our our stack of
artificial intelligence and I will show
you how how it it happen in the
solution so what are what what are the
symbolic EI that we use we call them
trackers okay trackers are specialized
analysis so like I just explained
earlier we have the maintain the M of
data what we do is that we explod this m
to have a a small granular small
granular slides of data and for each
slides of data we do a systematic
multi-dimensional analysis and to do
this systematic multi-dimensional an
analysis we
use some sort of rules because uh you
need to know where to look at and
currently our use case is uh is built
around business finance so our our
trackers are are here to simulate best
practice SOA a state-of-the-art practice
for business finance
analysis and what happen is that we have
two types of trackers we have trackers
that are mathematical and business
mathematical are here to spot a
mathematical anomalies okay for example
a gross revenue increasing
significantly above his objective okay
it is a threshold that we put that is
best practice to analyze corporate data
and it will it will spot this threshold
automatically within granular data
analysis but we get so business trackers
business tracker are more uh uh um um
specialized rule for to do corporate
finance right here I I put an example of
gross revenue increasing plus net
revenue decreasing it is an an anomaly
okay if your gross revenue is increasing
and that your net revenue is decreasing
is that you have a problem okay so we
spot how much automatically and and we
have 100 uh just here I just shown you
two example but we have hundred of those
trackers and we can with our client
build more
okay imagine that our client have some
sort of strategy or some sort of
specific problem we can build them with
them those type of
trackers so these are the symbolic uh
part
alongside Knowledge
Graph
so what I have just said until now seem
very complicated okay but what we aim
for is link
reporting we want our app to be in one
page
okay with two
parts we want a part that shows the risk
and opportunities found by artificial
intelligence right here in the first
part you got overview so we we show the
user what it it is used to see
Consolidated
data okay generally in pobi for example
reports you got Consolidated data and
the user can drilled on okay that's what
happen in most reporting here we show
the Consolidated data but below it we
show where to act okay what are the top
insights and right here you got a
list so you got a
list what are the most important
important
Insight individually and for each of
those
insights there there were there were a
multi-dimensional analysis and and those
are granular slice of data that were
systematically analyze and compare and
you got on the same line you got a train
analysis okay so we we do a dynamic
analysis we we show the user uh where
this product for example is going in the
next month we show a risk this is the
risk of not attaining the budgets the
budget Target okay we show the influence
of outlier
if this this uh result was explained by
outlier and we show
effects effects are uh the mix of price
and volume okay uh something can be
explain uh by an evolution in price but
also by an evolution in volume so we
show that in this visualization so this
is the first part of our application the
second part it is interpretability part
what we use we
use all fashion natural language
processing combined with Genera VI to
explain for each of those insights so I
have just to click on it and I will get
a full explanation of why this insight
has been
prioritized I will get also a
comparision with other product
etc etc for the analyst to to know why
this insect has been prioritized and to
be explainable because we are using
artificial
intelligence so what I am telling you is
that our
solution don't mean to erase the current
weight to do standard
reporting what we are building here is a
boost is a a complex mentary is a more
effective way to do traditional
reporting and it is complimentary in in
in many
ways so if we compare the uh the ability
to to give an overall picture A big
picture our reporting is not able to
give a big picture even even if we show
the Consolidated view but standard
reporting is more able to show uh an
overall picture or big picture okay
does the standard reporting do a
complete analysis no I I've just just
shown you uh standard reporting don't do
a multi-dimensional analysis don't de
consolidate so our solution is able to
detect early oneing sight uh to detect
uh problems when they happen and the
fastest fastest way possible and they
they do it it do it uh in a systematic
way and with a priorization
uh algorithm to do
that is standard reporting a guidance on
where to act no standard reporting in
most company are
Forest of reports and you don't know
where to act you need to do manual
digging it's like
Fishing manual
digging uh and human uh to do for
example a drill down approach to
deconsolidate manually it takes time and
it is a guessing it is an arbitrary way
of analyzing data what we do is that we
automate and systematically with
parallel uh processing algorithm with
Cuda calculus we automate the systematic
way of the aaging multi-dimension
uh the
analysis so in term of dimensional
analyze with with standard reporting you
can go to up to
three
dimension more you can't show it
visually you can't show it in a in a
pobi in whatever in a in in an
orac you can't show more than three
dimension with our solution this is why
we don't show visualization in this way
we we we do
unlimited amount of Dimension at the
same time in the
meantime as many as the management
account accounting for example for you
are use case currently
carries but what is useful with current
reporting is that it give
you where to
demobilize because on you spot uh an
Insight what high the high level
management will will want to do is to
move resources on those insights to dep
prioritize maybe some other project that
are not
uh a PRI priority thanks to our analysis
so this reporting can help you know
where to act in this regard where to
demobilize demobilize resources
so this
is the second part on you spot
Insight what we are building is Genera
VI to help the user build an action plan
so I have my insight okay I the Deep
detect we call it the first module
detect an insight and I want to build an
action plan with deep root this is the
second module that we we got and what we
we did is that we bu we built a
methodology to analyze a
problem okay what we are
spotting we do in a in a multi multi
multistep way a methodology to analyze
this problem and we are combining behind
the H A generate T VI with Knowledge
Graph to do that so the artificial
intelligence is here to to ask questions
to spot information available but also
to spot information that are necessary
but but they are not available so what
we what we would like to do with that is
to induce some sort of gumba walk some
sort of questioning some sort of inquiry
to prepare the analyst to do a business
plan and to show that at the actual
meeting that will decide prioritize
project so we are making a bridge
between the we reporting world and the
uh project management World in most
companies those two words don't talk
with each
others so we are building a two a bridge
between these two
words and we are building also a process
intelligence in the app uh you can uh uh
you have a view uh of what artificial
intelligence you use which which Step
because we are simulating what what I
didn't explain is that we are simulating
sparing sparring partners so I am
simulating a marketing director I am
simulating a financial director so
obviously each of those artificial
intelligence will have access to the
specific data of his domain and will be
able to help the user uh build some sort
of process intelligence Maybe in the
risk opportunity scope you you talk with
the the the
marketing generative AI uh and after the
context change when you analyze the con
that has been changing uh with the um
Supply uh uh Supply manager you want to
go back to the risk opportunity scope
and to talk with the uh Chief Financial
Officer okay what we are building here
is not to automate we are here to help
the user find information and build
something okay and I and we found that
generate VII is not here to help user uh
do for example text
tosql uh help them ask question to Data
Warehouse they are here to help to to to
ask to to have methodology to help them
get methodologies to help them find
information uh in this way they are far
more
effective so that's was it my
presentation I have lots of questions
yes no problem because
um how did you train the model yes did
you train a model or are you
using standard kind
of for for the generative AI part okay
uh for the model currently I am using um
because I I have no choice because I am
leveraging conage graph I am using uh
three 3.5 five uh model of open ey I am
I am uh it is a necessity because it is
the only model that is is not that much
costly and is able to quaring in um in
Cipher
but but what I I want uh and what I am
looking for currently in term of
development is to leverage for example
zier zier
7B or that sort of Open Source model to
do that but currently I didn't see in in
the research papers uh any evidence of
such capabilities I didn't have time to
find two an open source model of this oh
okay but then for the inside you're
using anomal detection model right
mostly inside no because what we are
using like like I said we are using uh a
layer of symbolic artificial
intelligence okay we are using
effectively uh a Time series to uh
calculate uh Trends and uh data
scientist will tell me that I don't have
enough data uh to do that but we have
some technique of data augment
augmentation mhm
uh and our solution overall is an
anomaly detection but we don't use uh an
anomaly detection uh we have an an
inhouse uh algorithm that is proprietary
we are doing a patency pat patency yes
claim we have a mark of Shain that is
doing the priorization of
insight so the priorization that you saw
behind the Ood is it is a complex Mark
of Shain to prioritize business problems
okay and when you said we we using I
mean this is interesting long long time
ago when I started studying artificial
intelligence it was still called expert
system yes that's why that's why we use
it yes and to see that expert system are
coming back after probably 30 years
there are I think right now uh recently
a stud a study showed that generate tvi
is more is more able to do text tosql
thanks toled graph where thege graph
have a database Shema in it so I think
we we are seeing am and I I am writing a
lot about it we are seeing uh some sort
of neuros symbolic I call neuros
symbolic system that are emerging
because the problem with a pure deep
learning approach is that deep learning
is not
interpretable uh or or it is very very
much difficult to interpret yeah and I I
believe that in most use case you want
you
need interpretability an
interpretation because artificial
intelligence its role is to if you want
artificial intelligence to be
prescriptive to prescribe uh
action you need a collaboration between
what the are the output of the
artificial intelligence and the human
and to have this collaboration you need
some layer of
interpretation
so and
uh but okay when you go to different
clients right then you might have
different rules business to you
different data quality you have
different semantics you have different
context you have different everything
right so yes so SCE you you very there
are two question in your
question for for data quality what
happens is that our tool is so granular
is so multi-dimensional that at
first what you are getting are lot of
data quality issues we are spotting
anomalies that are not based on a
business problem but anomalies that are
based on uh on the quality issue and and
we are we we help the client deal with
that because we are expert my company
has been working on on on this type of
reporting tool for almost 20 years and
we are expert in dealing with those
types of data quality issues so we help
them resolve their quality issues so it
is the first St of this is the first it
is the first tape and what you told me
is
that our current use case is Corporate
Finance so we uh like I said with
symbolic AI we built methodologies to
analyze Corporate Finance but even even
for this use case we can add specific
trackers okay currently we are working
with a large multidimens multinational
corporation in France and we are adding
uh specific trackers that are relevant
of what what we are doing right now with
their data so we can personalize our our
tool because uh the Deep learn the
machine learning part is is uh is
instrument to the symbolic part
the the main engine are symbolic the the
machine learning part is instrumental so
it doesn't it's not the main focus of
symbolic layer at the top you show
because our true Innovation if I I I I
put it blankly it's not an artificial
intelligence Innovation it is a
methodology method of De consolidating
multi-dimensional uh but it is only
possible with methodology
by uh parallel processing massively
parallel processing and with artificial
intelligence to do it in a systematic
and automated
way anyway this is a dream right to be
able
to um to get early warnings to get early
insights to take actions to make
decisions this has been a dream since
I'm do since I'm doing technology by the
way so it's not new
it is a dream but we talk about um my my
CEO talk about with high level figures
in the big
Corporation currently uh reporting tools
don't deal with Consolidated problem
because they don't have a
solution they they simply don't have a
solution or they just used uh
statistical kind of methods right to
find ins it doesn't yeah but then yeah
then you still need to like you said to
deconsolidate right now with a classical
statistical method there there are nine
nine methods to deconsolidate you can
use PCA Dimension reduction etc etc the
problem with those method is that uh you
can't do it
systematically and you need an a data
scientist to to to ask himself how here
there is maybe a problem and I I will do
uh some sort of analysis but it is too
late because when it SP the problem it's
taking too much time yeah it's taking
too much time yes so our our solution I
won't reveal or we we did
it obviously but but we we take a a
surprising
approach a Brute Force approach of
course I approach interesting I don't
know if there are more questions but
um so it's quite interesting because we
see the application of not only
generative AI but also standard machine
running uh you you spoke about xgb XG
boost right so which is kind of standard
machine
learning for database for symb combining
quite a lot of different yes and for one
page because what we want
is the Simplicity of it we want
something
U to be funny uh we are we are
incorporating in the app a speech speech
to interface but I can talk to the
application to show me what I want to
see yeah we want something that is
simple to understand because we want to
focus the Analyst job and they there
might be hundreds of them in in our
client to to focus on building a Sol
solution to the problem that has been
spot and not to build reporting because
the situation
currently most of the time of of of the
job day of those
analy are building
reporting dealing with data quality and
exact exactly we want to automate that
and we think that artificial
intelligence needs to be used to do that
because it is possible to do that in a
systematic way and we are not here to to
to to allow the user to ask any question
he wants right now all of lot of Builder
are speaking about text tosql unlimited
data on demand
Etc we think that artificial
intelligence is much better better than
than us to find problems in a systematic
way and to help making decisions based
on the experience yes but but to make
decision we think that the human it's
it's is it's more better to find
problems the EI is more better to make
decision the human is more better but it
can be helped with artificial intellig
but you can recommending you can
recommend actions
exactly you can also ask questions for
example you can have the we see it as
questions right it is interesting
because we see a generative as some sort
of recommendation engine it is here to
it is here to to recommend action but
not to make action for the user yes a
colleague of mine was showing me just
this morning um just a quick uh quick
and dirty app about reporting by the way
and then somebody was writing a report
and then the app was asking questions
that report was not answering right so
for example I don't know you have a
project stus report and then you forget
to men a Target date you forget to
mention this that then the the board of
the assistant will spot this and say
okay what about when do you expect this
to happen you need a Target blah blah
blah so this kind of thing is really
augmenting the intelligence yes because
I think gener is is really good to to to
to process a lot of data and to infly to
to to next token predict what is missing
uh what
what's present and yes it's really good
it's really good to do that like like
like you just just said I use it a lot
with u um with tabular with tabular data
frame and I I I ask it I ask it what is
missing from this
uh yes yes it's very good and we are
using it for that we are using it what
we are doing is that you you are using
it here in your solution for operational
data but not only for unstructural for
structured information but also for
unstructured information we are putting
in graph methodologies a way of doing
things a way of doing a problem a
particular problem Etc ET to help the
user to to upskill the level of
expertise of those analyst because uh
for our use case we have a problem of
turnover for those position and lots of
analyst don't stay a lot of time within
the company so we are helping them also
with their for their
expertise Mercy thank you very much
thank you thank you it was very
interesting good discussion hope
everybody enjoyed so I will I will
upload the video on YouTube and give you
the links on LinkedIn and every
everywhere possible thank you
Patrick and enjoy nice to talk to you
and enjoy the weekend
speak
Voir Plus de Vidéos Connexes
Projet Tada, une révolution crypto pour l’IA ?!
What is AI? - AI Basics
La data au service des clients, la data pour les PME
AI in Healthcare: The Next Frontier | Leonardo Castorina | TEDxUniversityofEdinburgh
Skills-Based Organization: What Works, And Twitter vs. Threads
Tout comprendre sur les modèles ARVALIS, au cœur des Outils d’Aide à la Décision - ARVALIS.fr
5.0 / 5 (0 votes)