Latest / Logosofia no Dia a Dia / Why AI Can’t Read Your Gmail (Yet)
Transcript
- [Prompt] custom: Gemini is in the process of rolling out an integration with its wide range of integration services p | Hosts: corn, herman [Response] Corn: Daniel sent us this one, and it's basically the question I've been muttering at my screen for six months. He's been trying to use the Gmail integration in ChatGPT and Gemini, the thing that's supposed to let your AI assistant actually read your inbox and send replies. And the experience was, quote, brittle and buggy. The read path missed actual replies sitting right there in his inbox. The write path just failed outright despite the connector being set up correctly. Herman: Which is wild when you think about it, because we've had the Model Context Protocol for almost a year now. Anthropic introduced MCP back in November twenty twenty-four. The plumbing exists. Individual developers can wire up a Gmail MCP server and get decent results in a prototype. But the moment you try to use it through ChatGPT or Gemini at production scale, it falls apart. Corn: That's the tension Daniel's pointing at. He's built prototypes himself that handle this more reliably than the paid products. But he doesn't want to build his own daily-use chatbot. That's why he pays for ChatGPT. So his question is, what's actually breaking here, and how far away are we from seamless read-write access across email, calendar, contacts, and Drive? Herman: The short version is, it's not an AI reasoning problem. It's an integration surface problem. The pipes are leaky, and the translation layer between natural language and API search syntax is where most of the failures live. Corn: You ask your AI, did I get a quote back from Acme Shelving? It says no results. But the reply is sitting three threads deep in your inbox with a subject line that says Re: Industrial shelving quote, updated pricing. The AI didn't fail to reason about the email. It failed to find it, because the search query it constructed didn't match the actual data. Herman: That's the thing Daniel's really getting at. The technical foundations are in place. MCP defines how tools get discovered and called. The Gmail API exists and works. But the gap between a prototype that a developer tunes with hardcoded search operators and a production system that has to translate generic natural language into precise API calls, that gap is enormous. Corn: Let's unpack what's actually going on under the hood, because the answer isn't that the AI is dumb. It's that the pipes are leaky, the safety layers add friction, and the business incentives push vendors toward caution over capability. Herman: The way I see it, there are really three layers of failure here, and they stack on top of each other. The first is just the sheer complexity of the API surface, plus the rate limits that throttle what an AI can actually do in real time. Corn: The rate limit one is sneaky, because it's invisible to the user. You just see no results and assume the AI is incompetent. Meanwhile, under the hood, the thing burned through a thousand quota units trying to paginate across two hundred threads and hit a wall. Herman: The Gmail API gives you a million quota units per day per project. That sounds like a lot until you realize every list call costs five units, and a single natural language query like find the quote for industrial shelving might need to scan hundreds of threads to be confident it hasn't missed anything. For one user in a prototype, that's fine. For ChatGPT managing millions of users sharing a single API project's quota, it's a hard constraint. Corn: The AI has to be stingy with how deep it searches, which means it stops looking before it finds the thing. Not a reasoning failure. A budget failure. Herman: And that's just layer one. Layer two is the thread reconstruction problem. Gmail stores messages as a flat list with thread IDs. The web client does all this work behind the scenes to present a nice threaded conversation. But the MCP read tool just hands back raw message data. The AI has to reconstruct the conversation itself, ordering by timestamp, handling drafts, dealing with split threads where someone changed the subject line. Corn: If it gets the threading wrong, the write path breaks too. Sending a reply means specifying a thread ID. If the AI constructs the message with slightly wrong headers, Gmail creates a brand new thread instead of appending to the existing one. That's almost certainly what happened with Daniel's failed write call. Herman: Which brings us to layer three, the one nobody wants to talk about. Google and OpenAI are terrified of a write action going wrong. Sending an email to the wrong person, replying to a thread with something sensitive, deleting a calendar event. So they layer on safety mechanisms, content filters, confirmation dialogs, extra rate limiting. Each of those adds a failure point that a prototype doesn't have. Corn: The prototype has no safety layer and one user. The production system has a dozen safety layers and millions of users. Of course the prototype works better. The surprising thing isn't that the production integrations are brittle. It's that anyone expected them not to be. Herman: Let me make this concrete, because the thread reconstruction problem is where most of the read failures actually happen. Picture Daniel's inbox. He's looking for a quote from Acme Shelving. The reply is sitting there, subject line Re: Industrial shelving quote dash updated pricing. Daniel asks the AI, did I get a quote back from Acme Shelving? Corn: The AI translates that into something like search Gmail for Acme Shelving quote. Which doesn't match the actual subject line at all. The word Acme isn't in the subject. The word Shelving is. The word quote is, but it's buried after a dash. The AI's search query and the real email are speaking slightly different languages. Herman: A developer building a prototype would hardcode something like from colon acme at example dot com subject colon quote after colon June first twenty twenty-six. That search is precise. It finds the email instantly. But ChatGPT and Gemini can't hardcode anything. They have to take Daniel's natural language, guess what search operators map to it, and hope the guess is right. Corn: The guess is often wrong in ways that feel absurd to the user, because the email is right there. The AI just looked in the wrong place with the wrong keywords. It's like asking someone to find a book in a library by describing the cover, and they only check the fiction section because your description sounded like a novel. Herman: And it gets worse when you add pagination. The Gmail API returns results in pages. If the AI's first page of results doesn't contain the right email, it has to decide whether to paginate deeper. Each additional page costs five quota units. Across millions of users, the system has to be conservative. So it often stops after one or two pages and reports no results. Corn: Which makes the user think the AI is broken. When really it's just being frugal with API calls because the quota math is brutal at scale. A million units a day sounds generous until you divide it across a user base the size of ChatGPT's. Herman: Here's the part that really gets me. The write path has its own version of this fragility, and it's arguably worse because the failure is visible. When Daniel tried to send a reply through ChatGPT's Gmail connector, the call failed. My bet is it was a threading failure. The AI tried to append to a thread ID, but it constructed the message headers slightly wrong, maybe the References header didn't match, maybe the In-Reply-To was off, and Gmail said nope, new thread. Corn: Which is actually Gmail behaving correctly by the spec. The thread ID parameter is strict. If the headers don't line up, Gmail assumes it's a new conversation. The AI has to reconstruct the exact header chain from the raw message data, and as we just established, the raw message data is a flat list that the AI is already struggling to parse correctly. Herman: This is where the prototype versus production gap becomes almost comical. In a prototype, you're the only user. You can afford to pull fifty messages, reconstruct the thread carefully, verify the headers, and compose a reply with the correct threading. It takes a few seconds. Nobody else is competing for quota. The safety layer is nonexistent. In production, you've got millions of users, a quota budget, safety filters scanning every outbound message for sensitive content, and a latency budget that says the user expects a response in under three seconds. Corn: The production system cuts corners the prototype doesn't have to cut. It paginates less. It spends less time reconstructing threads. It guesses at search operators instead of using hardcoded ones. Every shortcut increases the failure rate, and the failure rate is already high enough that users like Daniel just assume the feature doesn't work. Herman: The technical mechanisms are only half the story. The other half is about what this brittleness does to people's heads. Daniel said something in his prompt that I think is the most damning line in this whole situation. He said, I wasn't surprised to see that it didn't work. Corn: That's the trust deficit in one sentence. He paid for the feature, set it up correctly, tried it, and his baseline expectation was failure. And the system met his expectations. Herman: Which is worse than not having the feature at all. If the integration didn't exist, Daniel would just search his inbox manually and move on. But because it exists and fails, it trains him that AI can't be relied upon for anything that matters. That's learned helplessness, and the vendors are baking it in. Corn: He tries the read path, it misses the email. He tries the write path, it fails. That's two strikes in one session. How many users come back for strike three? Herman: Most don't. And that's the business tragedy here. Google and OpenAI have invested enormous resources into these integrations, but the first impression they're delivering is broken. They're teaching their own users not to trust the product. Corn: The safety impulse is understandable though. Imagine the headline: AI assistant sends confidential email to wrong recipient. That's a front-page disaster that sets adoption back years. So they wrap the write path in content filters, confirmation dialogs, moderation checks, rate limiters. Herman: Each of those safety layers introduces a new failure point. The content filter flags something innocuous as sensitive and blocks the send. The confirmation dialog times out. The rate limiter kicks in because the user made three requests in a row. A prototype has none of this. It just sends the email. Corn: There was a case in May, a ChatGPT user reported that the Gmail write action sent a draft to entirely the wrong thread because the AI misidentified the thread ID. The user had to manually delete the draft. OpenAI's fix was to add a confirmation dialog before sending, which users now blindly click through because they've been trained to ignore confirmation dialogs since nineteen ninety-five. Herman: The dialog becomes theater. It protects the vendor legally but does nothing for actual safety, and it adds friction that makes the feature feel clunky. Meanwhile the underlying thread ID bug is still there. Corn: This gets to something deeper about MCP itself. The protocol defines how tools get discovered and called, which is genuinely useful. But it says nothing about retry logic. Nothing about idempotency, so the AI doesn't know whether resending a failed write will duplicate the email. Nothing about conflict resolution when two actions touch the same calendar event. Nothing about state management. Herman: A prototype can assume a clean state. One user, no concurrent edits, no network partitions. The production system has to handle someone editing a calendar event on their phone while the AI is trying to reschedule it on their laptop, and the network blips halfway through the write. MCP has no standard for any of this. Corn: Every vendor is building their own operational layer on top of MCP, which means the reliability story is different for ChatGPT versus Gemini versus Claude. There's no shared infrastructure for the hard parts. Herman: The calendar integration makes this painfully concrete. The freebusy query is straightforward, you ask the Google Calendar API whether a time slot is open, it says yes or no. But writing an event means checking for double-booking across multiple calendars, handling recurring events where the AI needs to decide whether to edit one instance or the whole series, managing attendee notifications so you don't spam twelve people with update emails. Corn: The AI has to simulate human judgment about what's a reasonable time to book. Is nine PM too late? Is six AM too early? Does this person have a thirty-minute gap between meetings or do they need a buffer? The API doesn't answer those questions. The AI guesses, and sometimes it guesses wrong. Herman: Drive adds yet another layer. File search across Drive means understanding permissions. Is the file in My Drive or a shared drive? Does the user have edit access or just view? Is it the latest version or an old revision? The Drive API's search syntax is powerful but it's its own query language. The AI has to translate find the shelving quote PDF into something like title contains shelving and mimeType equals application slash pdf. Corn: If it gets the MIME type wrong, or searches the wrong drive, or misses a shared folder, the result is zero files found. The user sees failure. The AI didn't fail to understand the request. It failed to translate the request into a query syntax that matches how the data is actually organized. Herman: Microsoft sidesteps a lot of this with Copilot for three sixty-five. It runs inside the Microsoft Graph, which has native access to the same data store as Outlook. No API translation layer, no search syntax guessing, no quota management across an external API boundary. It just queries the graph directly. Corn: Which is faster and more reliable, but it's locked to the Microsoft ecosystem. If your company runs on Google Workspace, Copilot can't help you. Google and OpenAI are trying to do this cross-platform, which is inherently harder. They have to go through public APIs that weren't designed for AI assistants to begin with. Herman: Given all these constraints, what can you actually do today to get more reliable results? Because Daniel's not wrong to want this to work. The pieces are there, they're just not assembled in a way you can trust yet. Corn: First thing, and this is the one that'll save you the most frustration: stop using natural language for search. I know it feels wrong. You're talking to an AI, you want to talk to it like a person. But the translation layer is where most of the failures live, so bypass it. Herman: Instead of "find the quote from Acme," give it the structured query directly. "Search Gmail for emails from acme at example dot com with subject containing quote since June first." That maps almost one-to-one onto what the Gmail API actually expects, so the AI isn't guessing at search operators. It's just passing through what you gave it. Corn: If you don't know the exact sender address, at least give it the domain and a date range. The date range is especially important because it constrains how many threads the AI has to paginate through. Less pagination means fewer quota units burned and a higher chance the thing actually completes before the system gives up. Herman: Second actionable thing, and this one's going to sound like I'm telling you not to use the feature you paid for, but hear me out. For write actions, treat the AI as a draft generator, not an autonomous sender. Have it compose the email, read it over, then manually paste it into the correct thread yourself. Corn: The write path is where the brittleness goes from annoying to dangerous. A missed search result wastes thirty seconds. An email sent to the wrong thread, or worse, the wrong person, that's a problem that takes real time to clean up. The thread ID reconstruction is just not reliable enough yet to trust blindly. Herman: The confirmation dialog doesn't save you, because as we said, everyone clicks through those on autopilot. The only real safeguard is you verifying the thread context before it goes out. It's an extra step, but it's the difference between the AI being a useful assistant and a liability. Corn: The third thing is more of a watching brief. Pay attention to which integrations are native versus which are cross-platform. Microsoft Copilot inside three sixty-five, or Google's own Gemini inside Workspace, those have direct data access. They don't go through the public API translation layer. They query the same data store the application itself uses. Herman: That eliminates whole categories of failure. No search syntax guessing, no quota management across an external boundary, no thread reconstruction from flat message lists because the native integration can just ask the same threading engine that powers the Gmail or Outlook web client. Corn: The tradeoff is lock-in, obviously. If your company is on Google Workspace, Microsoft Copilot can't help you, and vice versa. But for reliability right now, the AI that lives inside your ecosystem is going to outperform the one knocking on the front door through a public API. Herman: Cross-platform MCP-based integrations will improve, but I'd put them at twelve to eighteen months from being reliable enough for critical workflows. The operational layer, the retry logic, the idempotency guarantees, the conflict resolution, none of that is standardized yet, and standardization moves slowly. Corn: Last practical thing. Test the integration yourself with a low-stakes task before you rely on it for anything that matters. Send yourself a test email from a different account. Ask the AI to find it. Then ask it to reply. See where it breaks. Herman: Document what you see. Which search phrasings worked and which didn't. Whether the write path threaded correctly or spawned a new conversation. That kind of failure-mode documentation is exactly what the product teams need to prioritize fixes. Right now they're flying somewhat blind on what's actually breaking in the wild. Corn: Daniel's experience, the read path missing replies, the write path failing outright, that's not an edge case. It's the typical experience. But the more users who can say "here's exactly how it failed and here's the query I used," the harder it is for vendors to treat this as a someday problem. Corn: Which brings us to the bigger question. Is MCP going to evolve into something that handles all this operational stuff, retry logic, idempotency, conflict resolution, or are we watching each vendor quietly build their own proprietary layer that bypasses MCP entirely? Herman: My bet is fragmentation first, consolidation later. Short term, Google's going to optimize Gemini for Workspace, Microsoft's going to keep deepening Copilot inside the Graph, and OpenAI will build whatever custom integration layer they need to make ChatGPT work reliably with third-party services. None of them will wait for a standards committee. Corn: Because the competitive advantage of being first to solve this is enormous. The vendor that delivers reliable read-write across email, calendar, and Drive, where you actually trust it to send the reply to the right thread, that vendor wins a huge chunk of the productivity market. Herman: Right now nobody has it. The pieces are all there, as Daniel pointed out, but assembling them into something that works at scale, across ecosystems, without terrifying the legal department, that's a two to three year project. Maybe longer if the industry splinters into walled gardens. Corn: The next time your AI says it couldn't find that email, ask yourself a different question. Not "why is the AI so dumb," but "which pipe broke?" Was it the search translation? The pagination budget? The thread reconstruction? The answer tells you where the real work needs to happen. Herman: That's the thing I hope sticks with people. The reasoning is usually fine. The integration surface is what's failing. Fixing that isn't an AI problem. It's an infrastructure problem. And infrastructure takes time. Corn: Now: Hilbert's daily fun fact. Hilbert: In the early fifteen hundreds, the horseshoe crab population around the Seychelles was so dense that sailors reported navigating by the bioluminescent glow of their blood in shallow waters at night. Corn:...bioluminescent navigation by crab blood. Corn: That's our show. Thanks to Hilbert Flumingtop for producing. This has been My Weird Prompts. If you want to send us your own question, email the show at show at my weird prompts dot com. Herman: We'll be back next week.