100 Ways LLMs can Boost Your Business

LLMs are not just about chat. All professions can reap productivity increase on various tasks. The only limit is your imagination!

LLMs certainly are a breakthrough in terms of natural language processing. However the real spark that turned to world mad is ChatGPT. Before it, you could still use GPT-3, but few people outside of specialists did. It’s when the chat form factor appeared that the general public started to realize the power of LLMs.

Unfortunately, chat — or at least passing as intelligent humans — is not the main strength of this technology, which is rather a sort of elaborate parser/translator. As such, there is a million ways you could integrate a LLM into your business at different levels, optimizing 10% of someone’s job here and there.

To prove this point, today we’ll explore 100 use cases that stand besides the stereotypical uses of LLMs to imagine what you could truly do in a wide range of industries, provided a bit of brain juice and a few lines of code.

Development and Project Management

Automated compliance checks in code or documents

Any company beyond a few dozen employees ends up forced to draft policies, processes and rules that must be followed. Some of those require a big picture thinking, but some of them are precise checkpoints that can be easily checked in text-based outputs: source code, contracts, commercial propositions, etc. A series of robots could entirely make sure that the bulk of policies is indeed applied throughout the company.

Programming language conversion

As said in introduction, LLMs are great at translating. But while this works amazingly for human-to-human languages, it also works quite well for programming languages. Typically, you can take any API vendor documentation in any language, get the example snippets and convert them into your current language. This also works within a given programming language to replace a specific library by another one that has equivalent features but different structure.

Detect bug reports from user reviews

It becomes easy to apply Linus’s law: “given enough eyeballs, all bugs are shallow”. If your product is meeting a certain level of success, people will inexorably start complaining about their frustractions online: through social media, app store reviews and so forth. Using a LLM, you can parse the whole lot of those reviews to detect if any of them actually describe a potential bug that you should care about.

Validate business strategies against doctrine

It is no secret that I am a fan of Wardley Maps. The only issue being: the source material is very long and complex. A potential use for LLMs (especially long context ones) is to be able to assist you in the map creation but most of all to check that your predictions and projections are actually taking into account all the rules from the 800 pages of the book.

R&D progress audit

It is always tedious to document R&D as the nature of it asks to iterate rapidly between various experiments. However if you were to centralize all your results in a semi-formal way, you can imagine have a LLM take over this reporting process and generate exact day-by-day reports of who did what, what are the conclusions and what are the next things being tested. Extremely convenient in the case of grant justification as well.

Task break-down and planning

Why are developers always so late? Sometimes, it’s simply unforseeable problems popping up, but most of the time — and especially for juniors — it’s because they fail to decompose the tasks that they have not already done. If you never did something, your brain will probably ignore all the sub-tasks that you will have to accomplish. A LLM could be a good help to break down a given task until all the steps and dependencies are clear.

Natural language programming

Instead of having to code a specific behavior from a software component (email filter, automation platform, data ingestion platform, etc), you could simply specify what you want in plain human language and have it transformed into code under the hood.

Drive processes (CRM, issue tracking, etc)

Having a system read all your emails, messages and so forth will definitely be a privacy challenge, but on the other hand this could enable automatically reporting status updates and changes to CRMs, issue trackers and so forth. For example you could analyze the Git history to move an issue’s status (alongside with comments explaining what happened). Or track commercial emails to automatically report on a lead’s status.

Run end-to-end tests of applications written in natural language

Isn’t that so fucking annoying to write front-end tests? This could change with appropriate use of LLMs. They could not only write tests for you but — and most importantly — they could also heal existing tests to adapt for code changes.

Visually test applications

LLMs can have vision capabilities. As such, they are able to do something more smart than a pixel-perfect validation. They could check two images and tell you if there are significant differences. Look at a web page and tell you about obvious issues (text overflowing, alignment problems, etc).

Log analysis to detect abnormal behaviors

Server logs are usually very long files that you keep to be able to diagnostic a particular issue if it happens, but when it’s about knowing what happens in real time then it becomes more complicated. Log monitoring tools exist but they are limited by the fact that logs are extremely diverse and unexpected. Instead, LLMs could be used to read all logs in real time and raise alerts when needed.

Threat modeling assistance

How do you secure a product? Nothing can be considered secure in the absolute, best practices are only good as long as they fit your needs. That is why you need to model your threat, which basically comes down to finding the weakest link in all the components holding the product’s security and figuring which might break easily enough for the prize to be worthy of the effort. This requires to imagine a full dependency map of everything related to the product, which a LLM could help enumerate.

Open source issue qualification

Open source projects historically always had issue with bug reports and feature requests, that are often done in a terribly unclear way. A robot could on the other hand be able to assist people in doing their report, until the produced description is clear enough for all parties.

E-commerce

Image-based search

Classical e-commerce facetted search requires detailed product description with a structured model of the product’s characteristics. And while for a stick of RAM this might be kind of easy, for some fields like clothing for example it’s already harder to categorize everything. On the other hand you could be asking questions about images, like “I want a pair of blue jeans with a contrasting seam” and the search engine could smartly filters images on this unexpected characteristic.

Mix-and-match assistant

Imagine that you find your perfect pair of pants but you are looking for a shirt to go with it. A LLM would be able to understand the level of formality, the color and the style of those pants and then to find a matching shirt for it. Let’s note that it’s a different concept from “recommended products” that exist today: here we consider the user’s explicit intention. This works for all kinds of products: cosmetics, food, tools, etc.

Organize products from raw pictures and spec sheet

Imagine that you are building a e-commerce in which you have raw material for each product that are pictures and PDF datasheets. You could have AI take care of creating categories, structured product characteristics and product description completely automatically, only leaving humans for review.

Product composition decoder

Imagine that you are lactose-intolerant and are looking to buy food. Or your skin has some specific alergies to chemicals. It would be interesting to be able to ask those questions to the e-commerce directly, which will decode tricky product compositions for you. Or even better, state in your profile which components you wish to avoid and the system will automatically put a warning tag on all corresponding products, along with a warning before checking out.

Product suggestion

You are redecorating your terrace and you need to figure what to put there. Send a picture of it to your furniture store and have directly matching suggestions displayed to you. It also works for various cases where the user could state a problem: “my computer is too slow”, “I need to water my tomatoes”, etc.

Visual audit of second-hand products

Since LLMs are able to view and to follow instructions, second-hand platforms could specifically ask to visually check known defects on pictures. This could help the user into qualifying their own product, as well as highlight important checkpoints to customers.

Price suggestion for second-hand platforms

In the same vein of being able to analyze products visually, you could as well automatically compare a given product to similar products sold in the past and suggest from there a fair price completely automatically.

Extract and categorize pain points from online reviews

Online reviews are a trove of user feedback for some products sold beyond a certain scale. Using a LLM to systematially parse them can be an interesting way to find out defects, use cases, quantify perception through time, etc.

Faceted search

Most e-commerce websites have what you call faceted search. It’s those filters on the left that allow you to refine a listing by some characteristics whether it’s size, color or anything else. Sometimes the experience is great but sometimes it is also not super smart. A way to improve the experience wouuld be to have a search bar that lets you specify in natural language the filters that you want to apply and then let the AI translate into the right request. No more awkward clicking, scrolling and waiting for page load again and again.

Entertainment

Drive NPCs — basically, Westworld

The Westworld show was pretty good — at least season 1 — at showing us instinctively what AI could accomplish for us and how it could do it. Give structured scenarios to NPCs and let actual player interact with them. LLMs can entirely be used to generate dialogues, figure the steps to stay on the scenario, etc. Potentially very exciting amusement parks in perspective, but also of course video games.

Generate backstories and character sheets for RPGs

If you are a RPG player — like D&D and the sort — you probably know that getting your character off the ground can be a lengthy process. Generate a backstory, specs, etc. It’s hours spent doing administrative procedures instead of playing. Instead you could just prompt the basic concept of the character and have it all generated in an instant.

Assist users learning how to play a game

Board games are always hard to understand the first time you play them. You could however imagine that an LLM-boosted agent could understand those rules and help beginners to play: explain what happened, let them know of potential moves, etc.

Answer questions about movies halfway

Sometimes in the middle of a movie you are just lost at what happened. However platforms already have a lot of data a LLM could exploit, like subtitles for example. Using this, you could imagine asking Netflix to clarify specific plot points and have the system check the transcript of the movie thus far to help you understand.

Image culling and storytelling

Ever came back from holidays with thousands of pictures that you never actually sort through? LLMs would be good at making a consistent story and picking the top X pictures to tell it in your album.

Book, podcast, etc. length or style adjustment

I personally hate reading fiction. For some reason, I’ve been devouring wikis and theories from GoT, LOTR and so forth but never actually managing to finish the books. It’s too long and too indirect. What if book — and podcast, news, etc — platforms let you adjust the length and style of what you are reading? 50-pages version of GoT? 2 minutes, to-the-point version of a 20 minutes podcast? The ability to further explore topics that piqued interest? Lots of people already watch movies in 1.5x, this would only be a logical next step.

Trope analysis and novelty factor

Star Wars episode IV is always a good example of how contextual movies need to be. Watch it in 1977 and it will fucking blow your mind. On the other hand I’ve shown it recently to a friend that was like “oh come on and then it’s going to be his father? how fucking original”. If you want to make a movie or an article entertaining, it must be composed of a good mixture of things that people are used to, spiked with an edge of novelty. Using AI to systematically explore and quantify tropes in existing scripts can help establish the novelty factor of a new project. Let’s note that this also works for politics, journalism, fiction and basically anything targeted at the mass market.

Auto-edition of video interviews

Interviews are a significant pain in the ass to edit. But using a LLM you could transcribe everything said, ask for it to pick the best part that will fill up X minutes and automatically slice and edit the video at the proper timestamps.

Conspiracy theory generator for social media

Whether we like it or not, social media is full of trolls trying to influence people’s choices and votes. A way of doing this is to attack specific pillars of a society (science, government, etc) by throwing an insane amount of conspiracy theories to destroy them. It doesn’t need to be consistent, it just needs to be massive. That is great, given that LLMs are excelling at making text that sounds good but that is utterly shallow. Pick your target, throw a LLM at Twitter and enjoy a massive ideological destruction.

Fan-fiction generation

Fans usually like their media so much that they want to keep exploring this world endlessly. Without making those stories canon, entertainment giants could easily generate literally endless stories by fine-tuning LLMs on the specific do’s and don’ts of a universe and let them generate content for their fans. As a bonus, the most successful stories could serve as a basis for major projects.

Data Analysis

Data visualization

Data visualization is a hard topic in the sense that managing all those graph libraries, SQL queries and the weirdest APIs like Pandas, it’s not very accessible to your average executive Joe. On the other hand LLMs are excellent at this, given a proper human intent. They are going to play a key role at making data more accessible.

Transform natural language signals into structured data

Scrape social media, listen to Slack messages or emails and turn this into structured data that you can quantify and analyze easily through graphs and statistics.

Loosely structured data cleanup

How many times data is provided in CSV form with completely inconsistent content? Poorly escaped lines, inconsistent IDs, etc. A usually tedious cleanup job could be entirely automated away with a properly trained LLM.

Reverse-engineer structures

Have you ever tried to understand what a company does from the outside? It’s usually very hard, given that the corporate website will tell you that they “deliver excellence” across a wide range of industries, present their “solutions” and “case studies” but will never go into the detail of what they actually did. The best way to understand the truth in my opinion is to look at job descriptions, both their quantity and their content. Gather them all together and you understand exactly which operational tasks, tools and hierarchy those companies have. Tedious by hand, but very suitable for LLMs to complete.

Natural Language Processing

Translate

All right this one is obvious in a conversational setup, but it of course also works if you are trying to internationalize a service. In an e-commerce or social media for example, the level of translations from a top LLM is good enough that you can trust it unsupervised in many languages for many non-critical use-cases.

Generate alt tags

Something that all CMSes should start doing: automatically generate alt tags for their image library. LLMs are now entirely capable of describing an image, and it’s so good for SEO and accessibility that this should become the norm very quickly.

Spellcheck

LLMs are also very good at spell-checking and can be used in a wide range of applications to help you improve your writing.

Find acronyms

The hardest thing when starting a project is to find a good name for it. Well not anymore, as now you can simply describe what your project does, ask Claude for a fitting acronym and there you go!

Parse free-form numbers

It’s not uncommon to end up with a data table where you need to parse prices or different kind of amounts but unfortunately they have been given in various forms, like “30 millions” or “45k”. While you can solve this with regular expressions, a cheap LLM can often be very efficient at parsing this.

Anything to Markdown

Given the ability of LLMs understand documents structure — textual or from images — they are excelling at producing markdown from anything. Just rasterize your PDF, throw it into a LLM and you’ll get your markdown version pretty easily.

Parse citations from academic papers

My understanding is that academic parpers follow a formal structure but in a semi-formal way technically speaking. Typically they are all linked to each other through citations, but their parsing is tedious. LLMs could empower this.

Smart replace in document

Imagine you write a long proposal for a client and refer repeatedly the name of their product or some important concept. But then your boss swoops in and asks you to remove or change all those references by another one. Sometimes search and replace can do the job, but sometimes it will affect the grammar or the structure of sentences. LLMs could do this job completely automatically.

Auto-adaptation of texts for different targets

Imagine writing a scientific revue. Maybe you want to address different levels of readers from the most advanced to kids. Or imagine a publisher that wants to make Shakespeare accessible to foreigners. LLMs are able to translate between languages but also between styles.

Re-phrasing of customer input

Customer support is a fantastic world where you get insulted for things you didn’t do. Instead, LLMs could act as a buffer between the customer and the support where aggressive, sarcastic sentences are turned into plain and clear ideas.

Content Generation and Management

Generate FAQ from website content

Gather all the content of your website, figure all the questions that it answers and generate the FAQ pages from this.

Generate decent usernames

It is quite hard to generate a decent username when subscribing to a platform. With a few smart questions and methodology, a pretty cheap LLM that could even run locally would produce many interesting name propositions in real time.

Create recipe/tutorial variations

When cooking, doing some work in your home, taking care of your garden or anything hobby-level in which you have no particular expertise, you will tend to follow tutorials to learn how to do things — and more importantly to achieve particular goals. The only issue with those tutorials is that they might have details incompatible with your particular situation. For example you want to cook a cake but you are alergic to one particular component. How do you replace it? That’s where the LLM can make educated guesses and alter the content dynamically to fit the user’s need.

Smart filling of templatized documents

Newsletter software allows you to place people’s name and a few other details within the text. But what if you could go much further than that? Create templates for documents like contracts, commercial outreach, etc. Then have a LLM fill up the blanks respecting grammar, gender or even making up whole sentences based on meta information: “Hi John, you expressed on our contact form that you need XXX, which can be filled by products in your YYY range. Let’s schedule a call?”.

Generate onboarding procedure and training path

Anyone running a company knows that transmitting the company’s knowledge is a tedious endavor. Pages and pages of process have been written over the many years of existence of the company, all at different levels of maturity. How do you introduce a newcomer to all this in a consistent order? You can feed all your documents to a long context LLM (like Gemini’s 2M tokens) and have it sort out documents in topological order and that are interesting in respect to a given job description.

Organize asset production based on company policy

Many companies have a process for rolling out a product or communication: social media assets, press releases, etc. While generating them directly will still be a human work, a LLM could allow high-level definition of guidelines in plain text, with more nuances possible than regular automation platforms, and automatically create outlines for the assets that need to be created.

SEO and keyword-centric upgrade of articles

CMS and other content management tools could receive specific directives regarding SEO and keywords that need to be present to perform not only live audits but also suggestions of modifications to the content in order to integrate the desired keywords.

Meeting prep, create meeting agenda

What worse than a poorly prepared meeting? Gathering information from previous meetings, ticket trackers and other digital platforms, an AI could outline the agenda of upcoming meetings, while at the same time helping each participant to gather their own content to present.

Auto-update documentation

All products and company processes need to be documented at various levels from the most technical to the most high-end. As the product grows and changes are made, it becomes hard to keep track of what needs to be updated in the documentation. Combining LLMs and embeddings, you could track the overall company activity and highlight parts of the documentation that become obsolete, list the missing parts and even automatically propose edits.

Dynamic course re-writing

Imagine a student learning online in front of their computer. Some topics will be easy but assuredly some others will prove more challenging. These courses are often evaluating the student’s skills all along the way. What if depending on those evaluations the content of the courses was adapted to the strengths and weaknesses of the student? Catch-up texts can be generated from the original course but focusing on the weaknesses and ellaborating on them further than in the initial content.

Infinite copy generator

How do you know which words are going to transform your audience best? What if you generated one version of your content for every single time that someone reads it? Then observe which versions worked the best and use this as reinforcement for your model, to produce more and more efficient versions of the copy.

Dynamic content

In the same vein, you can also observe the user’s behavior and browsing history in order to dynamically re-write or optimize pages when he lands there. Connect the dots with concepts freshly ingested, push forward detected interests, etc.

Image and Visual Processing

Transcribe handwritten notes

This might sound like a miracle but GPT-4o is able to read my handwriting. Not only this but it can transform it into a well-structured Markdown document. And then of course translate, summarize and all the perks. This can be helpful in number of scenarios from digitizing meeting notes to processing and translating on-the-fly antique manuscripts.

Pet control

Pets tend to behave differently when their owners aren’t home, like jumping on the bed or sofa. LLMs are definitely not the most efficient but they are for sure the easiest way to express to a machine “if a dog rolls in my sheets yell at them to stop”.

Generate color palettes

Just like LLMs are trained on word patterns, they are trained on visual patterns and including the understanding of colors. This means that you can generate smartly color palettes that actually work (as opposed to this color wheel madness you often see). This can help you generate your own UIs but even more than that what if the LLM were able to generate all the design tokens up to the user’s taste, ending up with a unique, custom and beautiful UI for every single user?

Art explanation

If like me you are art-illiterate but still end up in museum wondering what happened in a specific painting, only to find the name of the painter with a vague title next to it, with a lengthy audio guide telling you everything except what you want to know… you’ll understand this idea. Instead of audiobooks, musuems could provide interactive assistants fed with in-depth knoweldge on every work of the musuem but able to distil it in a way tailored to the visitor’s taste and to reply to their questions directly.

Picture-based food search

Google Maps is trying very hard to create ontologies of the real world, especially with its “questions” program asking you if a given service or food is available in various places that you have visited. However if you are not american you probably ended up confused when you got asked if your local high-end bakery was making smores. Food simply does not translate between cultures. That’s where a deep understanding of images could lead to a much more efficient search that would echo one’s way of expressing their wants.

Better narration for GPS

Did you ever take the Madrid highway with some US-optimized GPS voice? How long did it take you to take the wrong turn? With ample imagery available — street view, 3D maps, etc — you could absolutely have much more descriptive directions from the GPS. Referring landmarks, taking into account perspective, etc.

Drone or CCTV-based visual inspection of equipment or land

As you can describe what you want to see, you can have drones or cameras film something you want to inspect and ask the LLM to tell you if it matches your expectations or not. Look at satellite imagery and ask “tell me places where forrests have been depleted”. Look at a building and say “tell me if any tile is missing”. And so forth.

Auto-design simple, templated flyers, posters, etc

Some apps will help people organize events or do marketing. Especially for small businesses it’s going to be hard to create those assets on their own, as they will not have the means to work with bigger agencies and are most likely unaware of best practices. On the other hand the app could leverage LLMs to apply best practices, pick colors, use and customize proven layouts to generate all kinds of visuals.

Check that translations are meaningful in context by visually analyzing apps

A common example of translation error that infuriates me is around the word “check” in English that can be understood as two distinct French words: either as in “verify” or as “check this box”. And very often, the meaning is lost, leading to crazy translations like “Verify the terms and conditions to continue”. Since LLMs can read texts and context, they could be used to apply translation files on an UI and make sure that all buttons make sense.

Document Processing

Parse invoices

Invoice management is the bane of any small business. You receive hundreds of them, need to extract different items and taxes systematically, but on the other hand every single invoice has a different format. Fortunately LLMs are pretty good at extracting this information and putting it into a JSON — whether it comes from an email, a PDF, a picture of a ticket, etc.

Pick food at restaurant

Did you ever end up undecisive at a restaurant? Just snap the menu, feed it into a LLM and let it guide you into ordering something. It even works with hand-written texts that you can’t understand — Japan explorers will rejoice. If you are a restaurant you can even push this further and help users through a custom assistant.

Normalize recipes

As a nutrition app or related, you might want to make the link between the food listed in a recipe and the calories for example. But people writing recipes love to use the weirdest units or even have things implied — like some common ingredients not even being listed as ingredients. With the help of LLMs you can extract these ingredient lists, transform them into units that make sense and get the nutritional value of what you are cooking.

Convert mind-maps to structured linear document

Mind maps — or the Post-It method as well — produces a lot of ideas around one given topic but you might end up overwhelmed at the end of the process by the amount of information that needs to be processed. LLMs can transform those ideas into a linear structure, properly sorted and organized.

Paper forms digitalization

As an intern I have been copying lots of paper forms made on-the-spot for fidelity cards in a store. Or recently, we all have been filling up countless COVID forms whenever taking the plane. Using LLMs, you can understand and transcribe those forms completely automatically into a digital system.

Transcribe, tag and reference historical paper-only archives

Countless historical documents or books have been scanned but how many are properly referenced? You can guess that over thousands of years of history, we could set to map out all those documents, link them together, analyze references and ideas over time, to build a better understanding of our history and our currents of thoughts.

Customer Support and User Experience

Start workflows from emails

Customer supports will often be drowning in emails. You can parse them, detect intents and trigger the proper systems in your back-office to start procedures, without any human intervention.

Request routing inside the company

When the organigram starts becoming big, it might be hard to navigate the responsibilities and knowledge of the people in there. Especially for newcomers, it cn be a challenge to find the right person to talk to while also not bothering them unnecessarily. As CTO and founder I can answer most questions on most topics within my company, but should an intern come and ask me how to connect the printer? A reassuring AI could help people orient themselves in the hierarchy to know who they can confidently reach out for in order to receive help.

Prioritize incoming messages and notifications

You are probably like everyone else drowning in countless useless sollicitations, from services to which you subscribed 20 years ago to urgent business emails. Depending on the time of the day and your personal goals, you might want to be notified of one thing but not the other. Or you might want to receive notifications in bulk for some topics. For example I’d love to see Slack create a “what do you want to be notified about?” option and then burry irrelevant mentions and messages.

Configure complex features

When you use some apps, there will be features that are extremely complex to grasp. For example, try any product in Binance, it’s complex enough to throw you off if you are not eager enough to learn about it. Through the use of AI they could instead ease the user into setting the right parameters according to their own personal goals.

Voice message summarization

Some people love voice messages, some people like me are loathing them. Having an AI skip through the “sorry I’m sending you a voice message because I’m in the street and it’s easier to send a voice message […]” and instead deliver to you just the point of the message would be a great WhatsApp addition.

Conversation coach

We are constantly exposed to conflictual situations, especially in low-stake but annoying uses cases like negociating a refund overe an incorrect package received. Email and messaging apps could help you understand what you could obtain in that situation and redact emails for you, helping you every step of the way and reducing your mental load.

Automated test grading

The point of MCQs is that they are easy to grade, including if done by a computer. That’s why e-learning platforms use them so much. But given the advances of LLMs, it would be easy to imagine having them grade even textual responses, looking to see specific bits of information and telling you if it’s correctly explained or not.

Interpretation of complex diagnostics

Some diagnostics are not nice to hear, especially when they are particularly complex. From medical reports to SEO audits, if you are not an expert you might be confused by the terms and implications of those documents. A properly trained LLM could instead simplify them for you and even answer potential questions you might have.

Allow customers to do self-diagnostic on products

Vice-versa, some products are complex and have many failure modes. Companies internally have debugging procedure that can pin-point exactly what is faulty, but it’s hard for the regular customer to follow such procedures. Instead of paying a human being to tell you to turn it on and off again, such workflows could be assisted entirely by a LLM, specially driven by a tailor-made logic engine for your product.

IT support for end-users

The most feared and annoying department of a company is often the IT, that has the important task of securing the company’s data and intellectual property while also having to explain to users how to connect the Wifi. By the proper use of LLMs, with their general knowledge of how computer works, but with a specific training adapted to the policy, they could skim a lot of useless requests off the pile from IT departments.

Suggest A/B testing variants

A/B testing is great to test how the user is going to react to different UX or copies, but how might you do it? You need the ideas, after all. A well-trained model that knows the UX best practices for different industries could do this job of taking a human’s work and proposing potential optimizations.

Algorithm transparency

Many complex algorithms are ruling our lives. For example every electronic payment goes through a set of rules to determine if the action is legitimate or not. These departments are utterly closed and opaque to the rest of the company. Typically, if your card gets blocked then nobody in the bank can tell you why nor for how long it will stop working. Having a LLM being aware of the different rules of the algorithm, it could explain in simple words the reasons of this block to the bank advisor, and the available perspectives. This works for banking, but any sector with complex algorithm could leverage this.

Personal Assistance

Help user stay on plan

So many apps are helping us become a better version of ourselves, wether it’s for diet, exercise, jet lag, pet training, etc. But how often can you follow 100% of the plan? With a bit of intelligence you could let users report their deviations and help them stay on track without overreacting with counter-productive actions or simply getting demotivated.

Context-picking for events, emails, etc

Imagine an event in your calendar with few details. When the event comes up, the system could read your emails, meeting notes talking about this event and then infer useful information ranging from the latest tickets from the issue tracker, the weather if you need to go somewhere, remind you to take your IDs or advising you to dress in a certain way. Overall, for one item and a lot of context, the LLM could pick the top few elements that are relevant for you not to forget — wheter it’s an event, an email you are writing, a plane ticket, etc.

Long-term goal tracking

The human brain is very wired to small tasks and has a hard time taking a step back to see if you are achieving your long term goals. On the other hand an AI could be aware of your goals and rank each of your actions telling you if they seem to be helpful or not to achieve that specific goal.

Natural language passwords

Passwords are a notoriously hard problem. How do you make a password that is secure yet that you can remember? You could imagine generating complete phrases through LLMs but not only. When the user types back the phrase, you could use LLMs to normalize the text before hashing it so that spelling mistakes or punctuation or even word arrangements do not affect the outcome of the hashing.

Business and Legal

Categorization of items

The other day I wrote an article about GDPR, which parsed a big HTML page from the CNIL. The information there was semi-structured and I had to categorize things further in order to make sense of them. Same for the current article, the different ideas have been categorized by a LLM. This is a great tool to group similar concepts together.

Critique political programs

While LLMs are obviously biased — especially with the US culture wars — but are nonetheless able to project without ego into many persona. As such they are quite interesting to tools to use in order to review political programs and see how they are backed by facts and theory. Journalistic platforms could enhance their content with thorough review of every single politician, detect their parting from party ideas, and most important let people explore concepts on their own, for their personal situation or their vision of society.

Find out adequation between candidate and job description

Large companies will match candidates only based on keywords based on a first pass. However we know how biased this approach is, given specific technical knowledge is not shared by recruiters — and even less by those in charge of unpiling thousands of CVs, which must not be the highest ranking ones. On the other hand AI is great at matching a CV to a job description, describing quite well the fitting and lacking areas as well as the challenges to work with this person. Even further, based upon a transcript of the interview, you can ask questions and validate specific checkpoints by searching intelligently for the relevant parts, without having to listen to hours of recording.

Check ToC exhaustivity

When you reply to a RFP, a grant or any exercise of that style, there will be a list of requirements you need to meet and precise points that need to be addressed. Obviously it is more subtle than just filling up a form, you need to make sure that various aspects are answer consistently throughout the response. You can use a LLM to both extract from the RFP the list of points that need answering — or at least cross-check it — and also to check if your response provides adequate light on each of those elements.

Insurance claim/policy matching

Insurance policies are always a bit obscure. A well-trained LLM could allow customers to role-play use cases before subscribing, as well as gather all the necessary information in case of a real claim.

Social Media and Content Moderation

Automated content moderation

As OpenAI has proven to us with its extremely restrictive usage policy, LLMs can be used to detect offensive content — or any kind of content that you don’t want to see, offensive isn’t the same for everyone. In a day and age where social media operates at a great scale, being able to detect “forbidden” content would not only make the platforms safer but also more customizable. Indeed, what if instead of having one single policy, different communities had their own policy automatially applied? Free speech and safety for all!

Social media filters

In the same idea, what if instead having algorithms sorting your feed in the most opaque way, you could express what you want to see? The same as TweetDeck allows you to do by keywords for example, but with concepts instead. “Tell me all about space news” or “I’m sick and tired of meme X”. On top if filtering this could also mean groupping: different posts talking about the same topics could be groupped or even hidden after a threshold.

Specialized Applications

Ask questions about meeting transcript

Summarizing a meeting is good, but being able to look for specific information in it is the killer feature. “What did we conclude on the topic X?”. This is what I really want to see in those AI meeting platforms, especially to be used during ulterior meetings.

Step assistance in tutorials/recipes

When following a tutorial or a recipe, some steps that might seem obvious to the person writing it will probably be hard for you to follow if you are too new to the topic. Having a LLM to be able to write sub-steps to fill the gaps for you will be a great help.

Democratic platforms for citizen engagement

Politicians love to claim that they know what their people want, but how do they really know? With AI’s capability to categorize and summarize, you could turn griefs and ideas into structured input coming from the whole nation. A super-simplified procedure where you could complain about anything or ask for any change as it goes through your mind. Then processed and presented in any level of detail to your representative.

Real-estate property auto-description

So many platforms allow to post real-estate ads, but the quality of those is often mediocre to non-existent. What if using the proper context based on the pictures, the map information, the neighborhood meta-data and so much more, you could generate a proper text description highlighting the strengths and weaknesses of a given property?