7
- My Forums
- Tiger Rant
- LSU Recruiting
- SEC Rant
- Saints Talk
- Pelicans Talk
- More Sports Board
- Fantasy Sports
- Golf Board
- Soccer Board
- O-T Lounge
- Tech Board
- Home/Garden Board
- Outdoor Board
- Health/Fitness Board
- Movie/TV Board
- Book Board
- Music Board
- Political Talk
- Money Talk
- Fark Board
- Gaming Board
- Travel Board
- Food/Drink Board
- Ticket Exchange
- TD Help Board
Customize My Forums- View All Forums
- Show Left Links
- Topic Sort Options
- Trending Topics
- Recent Topics
- Active Topics
Started By
Message
How are you using AI in your professional life?
Posted on 9/13/26 at 6:15 am
Posted on 9/13/26 at 6:15 am
My company spends about $3,000 a month on AI, and we're getting a lot of value out of it. I can't give you a hard ROI number. What I can tell you is it's saving a lot of time and letting people get more done in a workday.
There are plenty of YouTube videos of people building games and running tests, but not many showing real examples of how people actually use this stuff at work. So here's what I do. We're a federal construction contractor, working in a lot of states. Small team, a lot of moving parts.
Daily reporting. I have AI connected to the SAM.gov API, and it sends me a daily report of new contract opportunities that fit our NAICS codes and set-asides. It's also connected to Slack, my email, and Procore. Every morning I get a report on each job: where we stand against the scope of work, what the contracting officers have been saying, and anything that changed. I know what's happening on every project without chasing project managers for updates.
Estimating and bidding. When a VA solicitation comes out, I have it read the whole package and help me build the cost and schedule estimate, including things like labor and logistics premiums for remote locations. It also drafts the boring but necessary stuff, like a request to the contracting officer for a bid extension when suppliers are slow getting us pricing.
Recruiting. This one surprised me. It's connected to a sourcing platform (SeekOut), so I describe the role, like a senior estimator with federal experience or a project manager who has done underground steam work at an occupied hospital, and it builds the candidate pool, screens it down to a shortlist, exports a tracker spreadsheet, and drafts the outreach emails in my voice. I'm running three searches right now that would have cost me a recruiter fee on each one.
Finance. It's connected to QuickBooks. When I wanted to know whether it made sense to buy equipment and bring pavement striping and sealcoat work in house instead of subbing it out, it pulled the P&L and AP data and helped me build the break-even model. It also produced a competitive analysis of who else is bidding that scope in our region. Not perfect, some reports it can't reach and I have to pull manually, but it's a real analyst on demand.
Contract admin. When we had a payment dispute with a subcontractor, it went through the email history and the subcontract, pulled out the clauses that supported our position, and organized the timeline. That's a few billable hours from a lawyer we didn't have to spend before we knew where we stood.
Documents. Proposals, memos, reports, spreadsheets, all formatted to our company standard.
The pattern: the value isn't one killer feature. It's that everything I used to have to ask someone for, or dig up myself, now shows up already summarized. The $3,000 covers a Team plan for the company plus API usage for the automations, and I run it as the account admin.
PS: I also use it for emails and posts like this one. I simply hit the record button and dump out my thoughts and AI cleans it up.
There are plenty of YouTube videos of people building games and running tests, but not many showing real examples of how people actually use this stuff at work. So here's what I do. We're a federal construction contractor, working in a lot of states. Small team, a lot of moving parts.
Daily reporting. I have AI connected to the SAM.gov API, and it sends me a daily report of new contract opportunities that fit our NAICS codes and set-asides. It's also connected to Slack, my email, and Procore. Every morning I get a report on each job: where we stand against the scope of work, what the contracting officers have been saying, and anything that changed. I know what's happening on every project without chasing project managers for updates.
Estimating and bidding. When a VA solicitation comes out, I have it read the whole package and help me build the cost and schedule estimate, including things like labor and logistics premiums for remote locations. It also drafts the boring but necessary stuff, like a request to the contracting officer for a bid extension when suppliers are slow getting us pricing.
Recruiting. This one surprised me. It's connected to a sourcing platform (SeekOut), so I describe the role, like a senior estimator with federal experience or a project manager who has done underground steam work at an occupied hospital, and it builds the candidate pool, screens it down to a shortlist, exports a tracker spreadsheet, and drafts the outreach emails in my voice. I'm running three searches right now that would have cost me a recruiter fee on each one.
Finance. It's connected to QuickBooks. When I wanted to know whether it made sense to buy equipment and bring pavement striping and sealcoat work in house instead of subbing it out, it pulled the P&L and AP data and helped me build the break-even model. It also produced a competitive analysis of who else is bidding that scope in our region. Not perfect, some reports it can't reach and I have to pull manually, but it's a real analyst on demand.
Contract admin. When we had a payment dispute with a subcontractor, it went through the email history and the subcontract, pulled out the clauses that supported our position, and organized the timeline. That's a few billable hours from a lawyer we didn't have to spend before we knew where we stood.
Documents. Proposals, memos, reports, spreadsheets, all formatted to our company standard.
The pattern: the value isn't one killer feature. It's that everything I used to have to ask someone for, or dig up myself, now shows up already summarized. The $3,000 covers a Team plan for the company plus API usage for the automations, and I run it as the account admin.
PS: I also use it for emails and posts like this one. I simply hit the record button and dump out my thoughts and AI cleans it up.
This post was edited on 9/13/26 at 6:16 am
Posted on 9/13/26 at 8:13 am to I Love Bama
I mostly use it to write random powershell scripts for me, or do some admin shite like creating spreadsheets and whatnot from interface metrics
Posted on 9/13/26 at 8:46 am to I Love Bama
I can't share what we use it for, but for the life of me I can't find one decent model that can read large PDFs well at all.
Most just make up shite as it goes, I know generally it's breaking the file down and reading the internal data so page numbers basically are meaningless. But none of the frontier models can read them without making up shite.
Would be a massive time saver, but in reality it's a time sink because we can't trust any of them reading PDFs we have found.
Tried Claude, chatgpt, Gemini, and grok. Chatgpt might be the best... Hard to say. At best it's right about 85-90% of the time of the information it ingests. Which is exceptionally bad. We are looking at converting PDFs to something more ingestable, but that costs a lot of time and money the second the models change
Most just make up shite as it goes, I know generally it's breaking the file down and reading the internal data so page numbers basically are meaningless. But none of the frontier models can read them without making up shite.
Would be a massive time saver, but in reality it's a time sink because we can't trust any of them reading PDFs we have found.
Tried Claude, chatgpt, Gemini, and grok. Chatgpt might be the best... Hard to say. At best it's right about 85-90% of the time of the information it ingests. Which is exceptionally bad. We are looking at converting PDFs to something more ingestable, but that costs a lot of time and money the second the models change
Posted on 9/13/26 at 8:56 am to I Love Bama
I’d be careful with all this shite if you are on an open model.
You are necessarily disclosing any of this to the public but the model is using your (and your clients proprietary info) to train and work for other people.
And you may be in breach of confidentiality agreements. You may not, but id check.
We were explicitly told over and over again we COULD NOT disclose ANY nonredacted client info in Claude until we had some sort of firewall add on added which took a couple of months.
You are necessarily disclosing any of this to the public but the model is using your (and your clients proprietary info) to train and work for other people.
And you may be in breach of confidentiality agreements. You may not, but id check.
We were explicitly told over and over again we COULD NOT disclose ANY nonredacted client info in Claude until we had some sort of firewall add on added which took a couple of months.
Posted on 9/13/26 at 11:54 am to Mingo Was His NameO
quote:
I’d be careful with all this shite if you are on an open model.
You are necessarily disclosing any of this to the public but the model is using your (and your clients proprietary info) to train and work for other people.
And you may be in breach of confidentiality agreements. You may not, but id check.
We were explicitly told over and over again we COULD NOT disclose ANY nonredacted client info in Claude until we had some sort of firewall add on added which took a couple of months.
Not sure why this is downvoted, because it is in fact a very legit concern, especially from a compliance standpoint. The problem is the horse left the barn so to speak before any decent monitoring/control tools were available. Those tools are now starting to catch up, but there's still not alot of very good ones out there just yet.
I've seen companies cheering the fact that they allow nearly any end user to deploy AI generated apps with little to no proper review or oversight. It's a disaster waiting to happen without proper controls in place. Some will even say it's standing in the way of progress, but it's usually not their arse on the chopping block if and when some of these user rolled apps get compromised.
This post was edited on 9/13/26 at 11:56 am
Posted on 9/13/26 at 12:44 pm to I Love Bama
quote:
$3,000 a month
I don’t think it’s uncommon to spend that much per tech employee right now. That’s why harnesses and open source models are all the rage right now. Luna too.
Posted on 9/13/26 at 2:31 pm to UltimaParadox
quote:
I can't find one decent model that can read large PDFs well at all.
Most just make up shite as it goes, I know generally it's breaking the file down and reading the internal data so page numbers basically are meaningless. But none of the frontier models can read them without making up shite
This.
I'm an SAP functional consultant. I use it mostly to breakdown complex ABAP programs into what it's doing line by line in functional terms.
We use Copilot which has great integrations with Teams, Onenote, Outlook, Word, etc outnof the box. I'll hold a requirements gathering call with whatever team I'm working a project for that usually is 3-4 hours, depending on how large the project is while I take personal notes in Onenote. I'll tell Copilot to review the recording and transcript of the call, consider my personal notes, and generate a requirements document. Typically takes me less than 30-45 minutes to review and edit/correct it compared to 8-10 hours for me to put it together myself.
I do the dame thing for creating functional specs. Across a major integration project it can save me another 10-12 hours of work.
Posted on 9/13/26 at 3:01 pm to UltimaParadox
quote:
I can't share what we use it for, but for the life of me I can't find one decent model that can read large PDFs well at all.
It depends on the type, contents, and legibility of the PDF. If you have text based PDFs, then the process theoretically should be trivial and nearly 100% accurate. It sounds like you may have the other type of PDFs though. Those PDFs appear contain text, but the contents of the PDF are actually images and not text. If that is the case, then the PDF must go through an OCR process that is CPU intensive and somewhat error prone even if the PDFs appear to be pretty legible.
Posted on 9/13/26 at 3:37 pm to UltimaParadox
quote:
I can't share what we use it for, but for the life of me I can't find one decent model that can read large PDFs well at all.
Most just make up shite as it goes, I know generally it's breaking the file down and reading the internal data so page numbers basically are meaningless. But none of the frontier models can read them without making up shite.
Would be a massive time saver, but in reality it's a time sink because we can't trust any of them reading PDFs we have found.
Tried Claude, chatgpt, Gemini, and grok. Chatgpt might be the best... Hard to say. At best it's right about 85-90% of the time of the information it ingests. Which is exceptionally bad. We are looking at converting PDFs to something more ingestable, but that costs a lot of time and money the second the models change
This is my issue. My main thing is needing it to pull PDFs, read PDFs, and output data from said PDFs.
However, like you said, I have found it making up stuff.
Has anyone found a good model? This is an offline closed system.
Posted on 9/13/26 at 4:07 pm to TigerFanatic99
I use to do this work many moons ago and I cannot imagine how much easier this would be now with LLM, meeting prep, etc. I could probably do 4-5 times the amount of work now than I used to do then.
Posted on 9/13/26 at 5:49 pm to jdd48
quote:
depends on the type, contents, and legibility of the PDF. If you have text based PDFs, then the process theoretically should be trivial and nearly 100% accurate. It sounds like you may have the other type of PDFs though. Those PDFs appear contain text, but the contents of the PDF are actually images and not text. If that is the case, then the PDF must go through an OCR process that is CPU intensive and somewhat error prone even if the PDFs appear to be pretty legible.
None of them can read text PDFs. Try it man, large PDFs like data sheets. Anything with text tables you might as well assume it is going to make everything up
We have had so many errors from so many models a lot of our guys are losing any confidence in it.
Literally none of them can read any decently formatted pdf that is more than a few pages.
I know for home use 85-90% might sound great.. For our work we are almost to the point to stop using it because it just inserts randomly wrong data constantly.
Posted on 9/13/26 at 6:05 pm to UltimaParadox
quote:
None of them can read text PDFs. Try it man, large PDFs like data sheets. Anything with text tables you might as well assume it is going to make everything up
We have had so many errors from so many models a lot of our guys are losing any confidence in it.
Literally none of them can read any decently formatted pdf that is more than a few pages.
If they are definitely text PDFs (and you mentioned they are quite large), then you're likely running into a context window limitation where the earlier parts of the PDF get "forgotten" the longer it processes. An alternative might be to tell it to process in page chunks instead of trying to process the entire PDF in one pass. That or maybe tell it to read the PDFs verbatim and not try to predict the next set of data or numbers.
This post was edited on 9/13/26 at 6:08 pm
Posted on 9/13/26 at 7:39 pm to jdd48
I'm trying chunks as well. Had a RAG knowledge base set that was throwing in stuff which was fake
Posted on 9/13/26 at 8:18 pm to I Love Bama
I work on a government defense contract. We were given some shitty enterprise version of copilot due to fears of classified information getting out. We're heavily restricted on what can be shared with it and it can't do shite anyway. It's basically just a chat bot that can rewrite your emails. It doesn't even have write access in excel. frickin useless yet our senior management is constantly hounding us about how we need to get all our teams to start using it. I do use AI at home but not at work
Posted on 9/13/26 at 9:29 pm to HailToTheChiz
I've tried chunks as well. However since every model has no concept of pages on PDFs it becomes a guessing game in what is contained in a chunk.
You end up searching the file yourself for the information, rather than getting false information.
The worst part though is even in small files they still hallucinate horribly.
Gemini Enterprise is the worst. Even giving it small documents and telling it to prove what it states with documentation. It confidently states wrong locations with fake information.
You can't point it to page numbers, since it has no concept of pages.
Codex at least say it can't process the file these days with the latest model, as it seems like they have given up on it lately. It will churn through almost a million tokens to give you no answer lol.
At this point we are trying different file formats. But every manufacturer on the planet puts their documentation in PDF. So unless it's something like a markdown file, I have little hope it can read anything remotely sophisticated.
Claude actually did better if you asked if for information instead of providing the source. But also once you try to tell it to point you to the reference to double check it generally has no way to consistently cite it's sources.
Some models definitely gotten worse over time.
LLMs are pattern matching to a fault, they just suck at folding human made documentation
You end up searching the file yourself for the information, rather than getting false information.
The worst part though is even in small files they still hallucinate horribly.
Gemini Enterprise is the worst. Even giving it small documents and telling it to prove what it states with documentation. It confidently states wrong locations with fake information.
You can't point it to page numbers, since it has no concept of pages.
Codex at least say it can't process the file these days with the latest model, as it seems like they have given up on it lately. It will churn through almost a million tokens to give you no answer lol.
At this point we are trying different file formats. But every manufacturer on the planet puts their documentation in PDF. So unless it's something like a markdown file, I have little hope it can read anything remotely sophisticated.
Claude actually did better if you asked if for information instead of providing the source. But also once you try to tell it to point you to the reference to double check it generally has no way to consistently cite it's sources.
Some models definitely gotten worse over time.
LLMs are pattern matching to a fault, they just suck at folding human made documentation
This post was edited on 9/13/26 at 9:32 pm
Posted on 9/14/26 at 12:30 am to UltimaParadox
Yeah I just experimented with about 10 Pdf pages. Could read parts but then seemed to make up the rest.
If the local models can't parse through pdfs, then what good are they?
What's the work around?
How else are you using ai in your office
If the local models can't parse through pdfs, then what good are they?
What's the work around?
How else are you using ai in your office
Posted on 9/14/26 at 3:00 am to HailToTheChiz
Was told a feature a customer requested was going to cost 7-10k, take a couple months to build + 1k per year to maintain a hosting solution long term.
Modeled tested and deployed it myself in a couple hours with zero hosting requirements/dependencies.
"Move, bitch get out the way."
Was told by our e-commerce team they didn't have the capacity or time to build a proper fulfillment model that needed just one lousy call to an API after I spoonfed the blueprint to them and said it would take months to develop after I discovered a flaw in the existing model already live.
Instead I built, tested and deployed the middleware to talk to the cloud API including creating a management application to support other products that use the same API without having to write an additional line of code in roughly 2 days including a migration component to map already completed transactions and handed them a *SINGLE* URL for them to swap out with the existing one that will take the order info and return a dynamic URL for each purchase keyed off a registration system we built.
"Move, bitch get out the way."
Creating a set of shared skills and posted to a new repo requested in Azure devops for my team associated with complex deployment environments that we've used for a decade so they can safely deploy powershell scripts from their AI accounts that prevent damage by making backups and creating a version/audit trail for each change. It's crazy we worked for a decade without any type of change management control until now. All they do is pull the skill into their desktop AI and as I refine and improve it, their AI clients get the updated skill rules instantly.
I could go on. Each project I get more comfortable with especially including the checks and balances should my boss ask to audit what I've done.
I've never felt more productive and free to execute my vision free of petty "sorry we can't do that" from other departments that has caused so many good ideas in the past to wither and die.
I wake up excited every day at how easy my job has become.
Modeled tested and deployed it myself in a couple hours with zero hosting requirements/dependencies.
"Move, bitch get out the way."
Was told by our e-commerce team they didn't have the capacity or time to build a proper fulfillment model that needed just one lousy call to an API after I spoonfed the blueprint to them and said it would take months to develop after I discovered a flaw in the existing model already live.
Instead I built, tested and deployed the middleware to talk to the cloud API including creating a management application to support other products that use the same API without having to write an additional line of code in roughly 2 days including a migration component to map already completed transactions and handed them a *SINGLE* URL for them to swap out with the existing one that will take the order info and return a dynamic URL for each purchase keyed off a registration system we built.
"Move, bitch get out the way."
Creating a set of shared skills and posted to a new repo requested in Azure devops for my team associated with complex deployment environments that we've used for a decade so they can safely deploy powershell scripts from their AI accounts that prevent damage by making backups and creating a version/audit trail for each change. It's crazy we worked for a decade without any type of change management control until now. All they do is pull the skill into their desktop AI and as I refine and improve it, their AI clients get the updated skill rules instantly.
I could go on. Each project I get more comfortable with especially including the checks and balances should my boss ask to audit what I've done.
I've never felt more productive and free to execute my vision free of petty "sorry we can't do that" from other departments that has caused so many good ideas in the past to wither and die.
I wake up excited every day at how easy my job has become.
This post was edited on 9/14/26 at 3:26 am
Posted on 9/14/26 at 7:00 am to HailToTheChiz
If it's just text, I would try having it first convert it to a markdown file. Then, in a new chat have it analyze the markdown file.
Posted on 9/14/26 at 8:14 am to I Love Bama
We fabricate frameless shower doors for glass companies. Almost all of my customers are low-tech and not math-savvy enough to use our online portal or calculate the prices themselves. Instead, they send in a hand sketch of the shower enclosure, and my customer service team has to then enter it into our system for quoting. It's obviously time-consuming.
I created a standard form my customers could fill out and sketch on. I then trained Claude to read these forms and hand-sketches and determine pricing. Took a couple afternoons, but I got it about 98% of the way there, which is inline with our hand pricing accuracy. Whereas a quote used to take 7-10 minutes, it now takes maybe 30 seconds. I made the standardized form into pads that I can hand out to my customers. I told them if they use that, we can quote the same day.
Now, the customers take the pads to their jobs, measure, sketch, fill out the options on the form, snap a photo, send it to an email, and AI automatically prices it. I still require the AI price to be approved by a human, but so far I have not seen any major screw-ups. Just a few dollars here and there that seem to even out over time.
What impressed me was the ability of the AI to interpret the drawings. Every customer draws them slightly different, so I had to run through quite a few training runs with AI, but it finally got it. I'd say all-in it was about 40 hours worth of work, and it used the $100/month Claude with no overages on usage.
I also use AI to help me in my production tracking/planning. We have an ERP that runs my factory. It tracks orders and talks to the machines, but the reporting is abysmal. It's very time consuming to get it into something legible for me to read, much less my guys on the floor to read.
I taught claude about my factory, the intricacies of glass fabrication, timing, constraints, and all that fun stuff. I then trained it to read the reporting from my ERP and turn it into a report that can be run at the click of a button. Now, I can see very quickly where there are hiccups in the process, what station is behind, where I have excess capacity, what orders are done, where they're at, etc. It also creates a report I can hand out to station leads telling them what's coming, what's behind, and time to complete.
I created a standard form my customers could fill out and sketch on. I then trained Claude to read these forms and hand-sketches and determine pricing. Took a couple afternoons, but I got it about 98% of the way there, which is inline with our hand pricing accuracy. Whereas a quote used to take 7-10 minutes, it now takes maybe 30 seconds. I made the standardized form into pads that I can hand out to my customers. I told them if they use that, we can quote the same day.
Now, the customers take the pads to their jobs, measure, sketch, fill out the options on the form, snap a photo, send it to an email, and AI automatically prices it. I still require the AI price to be approved by a human, but so far I have not seen any major screw-ups. Just a few dollars here and there that seem to even out over time.
What impressed me was the ability of the AI to interpret the drawings. Every customer draws them slightly different, so I had to run through quite a few training runs with AI, but it finally got it. I'd say all-in it was about 40 hours worth of work, and it used the $100/month Claude with no overages on usage.
I also use AI to help me in my production tracking/planning. We have an ERP that runs my factory. It tracks orders and talks to the machines, but the reporting is abysmal. It's very time consuming to get it into something legible for me to read, much less my guys on the floor to read.
I taught claude about my factory, the intricacies of glass fabrication, timing, constraints, and all that fun stuff. I then trained it to read the reporting from my ERP and turn it into a report that can be run at the click of a button. Now, I can see very quickly where there are hiccups in the process, what station is behind, where I have excess capacity, what orders are done, where they're at, etc. It also creates a report I can hand out to station leads telling them what's coming, what's behind, and time to complete.
Posted on 9/14/26 at 9:51 am to AaronDeTiger
I will try this. Thanks
Page 1 of 2
Popular
Back to top
Follow TigerDroppings for LSU Football News