Learning Library

← Back to Library

OpenAI Data Connectors Fall Short

Key Points

  • OpenAI launched new data connectors, adding integrations like GitHub, Linear, Zapier, Gmail, Outlook, SharePoint, and Google Calendar to compete with Claude’s similar tools.
  • The company warns that these connectors are not meant for deep research or extensive analysis of large personal datasets such as Google Drive spreadsheets.
  • In testing, the connectors struggled with complex queries, often providing only vague or approximate results rather than precise counts or detailed insights.
  • A technical limitation caps API results to 15 items per request, making comprehensive tasks—like analyzing hundreds of emails, calendar events, or documents—practically impossible.

Full Transcript

# OpenAI Data Connectors Fall Short **Source:** [https://www.youtube.com/watch?v=UvC33D6lwRI](https://www.youtube.com/watch?v=UvC33D6lwRI) **Duration:** 00:11:40 ## Summary - OpenAI launched new data connectors, adding integrations like GitHub, Linear, Zapier, Gmail, Outlook, SharePoint, and Google Calendar to compete with Claude’s similar tools. - The company warns that these connectors are not meant for deep research or extensive analysis of large personal datasets such as Google Drive spreadsheets. - In testing, the connectors struggled with complex queries, often providing only vague or approximate results rather than precise counts or detailed insights. - A technical limitation caps API results to 15 items per request, making comprehensive tasks—like analyzing hundreds of emails, calendar events, or documents—practically impossible. ## Sections - [00:00:00](https://www.youtube.com/watch?v=UvC33D6lwRI&t=0s) **OpenAI Launches Data Connectors with Limits** - OpenAI introduced new data connectors for services like GitHub, Gmail, and Google Calendar, enabling cross‑account search but warning that the feature struggles with deep, local‑data queries such as complex spreadsheet analysis. - [00:03:16](https://www.youtube.com/watch?v=UvC33D6lwRI&t=196s) **AI Shines on Narrow Topic Briefings** - The speaker explains that the system fails with broad email searches but delivers strong, comprehensive briefings when given a specific, time‑bound keyword that links limited data from Gmail, Calendar, Docs, and the web. - [00:06:53](https://www.youtube.com/watch?v=UvC33D6lwRI&t=413s) **OpenAI Enterprise OS Progress** - The speaker critiques OpenAI's current enterprise OS capabilities, notes early Operator's failures, then praises the 03‑model upgrade that makes Operator fast and accurate enough for real tasks like flight planning. - [00:10:25](https://www.youtube.com/watch?v=UvC33D6lwRI&t=625s) **Refine Query Precision for Data Insights** - The speaker reflects on a failed attempt to extract email and calendar patterns, emphasizing the need to craft more precise, well‑structured queries to effectively use low‑bandwidth data connectors. ## Full Transcript
0:00Open AAI released data connectors 0:02yesterday. Data connectors are basically 0:04OpenAI's answer to Claude connecting you 0:08to Gmail and Claude connecting you to 0:10calendar etc. being open AI, uh they 0:14added a bunch more that Claude had not 0:16previously added because of course this 0:17is a competitive arms race. And so they 0:20added GitHub, they added Linear, they 0:22added Zapier, they added a bunch of 0:24things. And then they also of course had 0:27Gmail, they had Outlook, they had 0:28Sharepoint, they had Google Calendar. 0:32Essentially they are saying that you can 0:35now search in your plus teams and pro 0:37account across a lot of the personal 0:40information that you create as you do 0:43work and they are careful to call out to 0:46their credit that this is not a perfect 0:48search mechanism. They specifically call 0:50out this shouldn't be used if you're 0:52searching for example in Google Drive 0:54for doing extended work across uh ma 0:56math on sheets. And so if you're doing a 0:59spreadsheet analysis, deep research is 1:01probably not best positioned for that. 1:03I'm sure they're right. But even when I 1:06gave it several queries, deep 1:09research, that were not designed to test 1:12it against what OpenAI 1:14suggested, it still didn't work. In 1:16other words, I tried to stay away from 1:18the warning spots. I tried to give it 1:19challenging queries similar to what I'd 1:22given deep research in the past when it 1:24only worked on the open web and it 1:27really fell down on local information 1:31and I was actually able because you can 1:33see the chain of thought for deep 1:34research to pull up that chain of 1:37thought during the queries I sent and 1:39pull screenshots of what deep research 1:43said it was doing. And I learned some 1:46fascinating stuff. It turns out that the 1:50API result that it is relying on to get 1:53results from calendar, to get results 1:55from Gmail, tops out at 15. In other 1:59words, if you want to say like you would 2:02to, I don't know, a poor hardworking 2:03executive assistant, please do a 2:05comprehensive analysis of last month, 2:08last month's email volume, cohort it 2:10out. Tell me who I need to be focused 2:12on. Tell me how I can use my time more 2:13efficiently. give me a sense of the 2:15types of emails that I need to respond 2:16to and the ones I don't. It can't do 2:19that. It has access, but because of the 2:22thinness of the data pipe it's working 2:24with, it is absolutely impossible. And 2:26trust me, I tried. I tried to do an 2:29email analysis of the last 100 emails. I 2:31tried to do a calendar analysis of the 2:33last 100 calendar docs. And this is me 2:35being kind. I would have said a thousand 2:37if I could, but I had a feeling. Uh I 2:40tried to do an analysis of the last 100 2:42uh Google Docs that I created. It does 2:46extremely limited searching. I could 2:48find evidence in in my query for a 100 2:50docs. I found evidence of it calling 2:52back three. In my query for a 100 2:54emails, it could not produce exact 2:57counts. It just kind of waved its hand 2:59in the air and gave me approximate 3:01numbers. And as someone who gets the 3:03email every day, I knew that it had 3:07correctly ga guessed categories, but it 3:10had wildly incorrectly guessed numbers. 3:13The numbers it was guessing for whole 3:16groups of my email were completely 3:19off and it just didn't do the groundwork 3:22of actually checking the email even 3:24though the data connector was 3:26there. So it just failed like we just 3:29don't have to call it anything else. 3:30Where did it succeed? You might wonder. 3:32Well, I will say I gave it a specific 3:35topic to look into and it did much 3:38better there. So, if you give it, let's 3:39say you give it a webinar you're 3:41planning or you give it an event that 3:42you want to do, something that has a 3:45defined time focus and you say, "Look 3:47across the web. Look across my email, my 3:50calendar, give me a comprehensive 3:52briefing for just this very tight topic 3:54that's very clearly delineated by 3:56keyword." Then it does pretty well. It 3:59can use that keyword as a guidepost 4:01across Gmail, across calendar, across 4:04the open web, across your Google Docs, 4:07and come back with something that is 4:09actually a decently good comprehensive 4:11briefing. How does it that it do that 4:13well? Because each individual data 4:16source is not going to be a ton of 4:18individual units of data. It's not going 4:20to be more than 15 in many cases. and it 4:24can then assemble that out into 4:25something that's really comprehensive by 4:27inferring and reasoning across all of 4:30them together which 03 does very 4:33well. It also helps if that event has a 4:36public presence because then it can do 4:38what uh deep research does best which is 4:40reason across the entire web at scale. 4:45So to me when I look at this if I step 4:47back I see this in the context of 4:50ongoing 4:52competition both between model makers 4:55between anthropic and open AI which I 4:57laid out at the beginning and also 5:00between model makers and specific 5:02verticals they want to go after. One of 5:05the questions that I got uh as this came 5:07out was is anything safe? Like they keep 5:10going after these verticals. Who's next? 5:11Granola. Did they get eaten? Because one 5:13of the data sources here is that like 5:15teams will now record your 5:17calls. Look, to be honest with you, I 5:21think the common rationale across these 5:24recent moves by both Anthropic and Open 5:27AAI is all about tokens and data in. 5:31It's all about training data. 5:32Everybody's hungry for it. And so 5:34they're building connections to Gmail 5:36for training data, for calendar for 5:37training data, anything they can get. 5:40They're building the meeting transcripts 5:41piece for training data. They're 5:42building um Anthropic is cutting off 5:45access to models in Windsurf, which was 5:48acquired by OpenAI in order to cut 5:50OpenAI off of training data. Now, it 5:54doesn't really matter because Windsurf 5:55can get thirdparty access instead of 5:56first-party access. But the point is 5:58like the intent the intent is to cut 6:00them off on training data to keep your 6:01your rival from getting training data. 6:04And so if you're building in the space, 6:05the question you should be asking 6:07yourself is how easy is it for a model 6:10maker to get access to training data 6:13that they would find high value that 6:15would have real rewards if they got it 6:17right. And so my proposal to you is to 6:22look for places where that data would be 6:23hard to get where they couldn't just add 6:25an MCP server and get the data because 6:28that is basically what OpenAI did. And I 6:30think the question is not would they add 6:33an MCP server to get the data if they 6:35could. It is is collecting the data to 6:39do this in line with their larger vision 6:41that they have stated for the company 6:43because you know throwing elbows between 6:45open AI and anthropic aside this is 6:48right in line with what we would expect. 6:50OpenAI has been really clear about their 6:53plan to be the default OS for the 6:55enterprise. If you're going to be the 6:57default operating system for work, well, 7:00you got to do meetings. Shouldn't have 7:02surprised anybody. You got to do Gmail. 7:04You got to do Calendar, you got to do 7:05Outlook, you got to do SharePoint. This 7:08isn't that surprising. Now, are they 7:10doing it well enough that I would trust 7:12them with all of that now at the 7:13enterprise level? No, they're not. In 7:16fact, this is reminding me a lot of the 7:19initial release of Operator back in 7:21January when Operator came out and it 7:24was frankly pretty terrible. I used it 7:25about three or four times and uh it was 7:28inaccurate. It was slow. It took 7:30forever. It froze up on easy things like 7:33add to cart. They just didn't use it 7:35anymore. Well, I had a feeling they 7:37would make it better because after all, 7:40it's in beta. And last week they did. 7:43They added their 03 model as the driving 7:45model behind operator. Operator got 7:47about 10 times faster and 10 times more 7:50accurate. It's actually a useful tool 7:52now. They didn't make as big a fanfare 7:53out of it, but I now find that I 7:55actually go to it and imagine specific 7:57use cases because it is fast enough. I 7:59used it for a real task. I did not 8:02pre-plan just this week. Uh, and if 8:05you're wondering, it was flight 8:06planning. And yes, I know they trained 8:07for flight planning and all of that, but 8:09it wasn't even usable for flight 8:10planning in January. So, it's not just 8:12that they trained for it, it's that they 8:13got the right model. They took the time. 8:16They probably collected data anonymously 8:18from some of us users as we were using 8:20it. um and they were able to eventually 8:22make the model better at browsing the 8:25internet. I expect the same thing to 8:28happen with this connectors play. This 8:30is a long-term play. I don't consider it 8:32particularly usable today. I consider 8:35the move toward data for the enterprise 8:37and workplace something that is not 8:40surprising and something where open AI 8:43is going to get much better at this over 8:45the next six months. They need more data 8:47to reason across. Now, this doesn't mean 8:50that they will immediately become 10 8:53times more intelligent at handling all 8:55of that data. I do think there are real 8:59questions about the ability of these 9:01models to reason well across very large 9:05data streams that are incredibly messy 9:08that are created by humans. Have you 9:09looked in your notion or your wiki at 9:11work? It's probably pretty dirty. 9:14Everyone kind of rolls their eyes and 9:15throws stuff in there. When you have a 9:17dirty repository of unstructured text 9:20data like that, it is inherently not a 9:22great place to ask the AI to start to 9:25make meaning. And yet that is often the 9:27case for a lot of our text 9:30repositories. And so I think my 9:32challenge to you is how good is your 9:35prompting and how precisely are you 9:37asking for work to be done? One of the 9:39characteristics that we are learning for 9:41AI in 2025 is that when you prompt well, 9:45when you prompt cleanly and clearly with 9:47a very specific task in mind, you often 9:49get surprisingly good results. But when 9:52you ask for something fuzzier, harder, 9:54more like you would ask a very senior 9:56researcher where you don't even fully 9:58know the question, you often get quite 10:00poor results. And that really explains 10:03how my experience went with connectors 10:06today. I was able to get my specific 10:08query around, you know, events and 10:11webinars and understanding how they all 10:13uh assembled into a briefing and kind of 10:15give me the sense of the emails and the 10:16calendar and the agenda and the public 10:18profile for this specific event. That 10:20went well. It was a very specific query. 10:22My more generalized discovery question 10:25around patterns in my email and patterns 10:27in my calendar that went terribly. like 10:30it just did not go well at all and it 10:33could not pull the volume of data needed 10:35to make sense of it. And so I am taking 10:38away from that that part of the 10:40challenge for work here in 2026 is how 10:43can we more effectively structure our 10:46queries so we precisely ask what we 10:48intend. That takes a lot of human work. 10:51It's not easy but I think that's a big 10:54work skill that we all can stand to get 10:56better at. Clearly, I could stand to get 10:57better at it because I batted like one 10:59for three, one for four on my test 11:01queries today. I need to figure out ways 11:04to use this tool that are more more like 11:06a scalpel and less like a chainsaw. This 11:09is not a chainsaw tool. It's a very 11:10precise tool right now. It may get 11:12higher bandwidth as they increase the 11:14scope of those connectors. It's not that 11:17high bandwidth now. So, that is the 11:20skinny on what happened with data 11:23connectors. I hope you enjoyed this 11:24quick review and uh we'll get into a bit 11:27more of like AI and work I think in the 11:29next piece because there's a lot more 11:31that is behind the scenes of this move 11:35that I want to dig deeper on. All right. 11:38Cheers.