Back to Blog
    The Answers Were Already in Our Inbox

    The Answers Were Already in Our Inbox

    11 min read

    Every business I know has a filing cabinet nobody can open. Ours had the locksmith's phone number locked inside it, and it took an AI research assistant to pick the lock.

    For years we stored customer knowledge the way a raccoon stores dinner: everywhere, enthusiastically, and with no system anyone could explain afterward. Email. Teams chats. Meeting recordings. Signed agreements we filed with great ceremony and never opened again. Somewhere in that pile sat the direct line, the job title and the mobile number for nearly every customer contact we've ever worked with. Getting to any of it meant somebody squinting at email signatures for a week like a monk illuminating a manuscript, only with worse lighting and a Teams notification every nine seconds.

    So in September we built a research assistant inside Microsoft Copilot, named it Brain Picker, handed it to the team, and asked them one question. Who do you know?

    What happened next is the most fun I've had with a contact list. Which, I grant you, is a bar so low I could run for mayor of this town.

    The gap we didn't want to admit

    We were moving our service desk onto a new system, and a new system is a fussy houseguest. It wants a clean contact list. Who works where. What they do. A phone number that rings an actual human being.

    Our records had holes you could drive a Zamboni through. 704 contacts were missing a phone number or a title. Some names were misspelled. One "mobile" number turned out to be the front desk of an entire school. We are a technology company, and our contact list had more empty bubbles than the Scantron of a kid who spent study hall playing Oregon Trail.

    The maddening part is that the information existed. The tech who fixed the Wi-Fi at a property last spring had the facilities manager's cell phone sitting in an email signature. The person who signed the contract had the CFO's direct line. Everybody held one piece. Nobody held the picture.

    Asking one person to go collect all of it would be like hunting for a cassette in a 1987 Camaro. You know it's in there. You just have to check under both seats, behind the visor, in the T-top bag, and inside the case for a completely different album, and when you finally find it, it's been eaten by the tape deck.

    What our AI research assistant actually does

    Brain Picker lives inside Microsoft 365 Copilot, the AI assistant a lot of businesses already pay for and mostly use to summarize meetings they should have skipped. You type a person's name and their company.

    The user manual ends there. It's shorter than the instructions on a bag of microwave popcorn, and harder to get wrong.

    From there it does the digging, like a private eye with a photographic memory and no interest in lunch. It reads that person's emails to you and copies the signature block at the bottom, because the email signature is the most accurate document in American business. People will embellish a résumé until it qualifies as historical fiction, but nobody fudges their own direct line. It searches your Teams chats and meeting recaps for the moment someone said, "This is Jane, she runs facilities." It checks your agreements and spreadsheets. It looks the person up in your contacts. Then it writes a short report in plain English, with every fact tagged to the email, the sender and the date it came from.

    When it's done, it saves that report as a Word document in a shared library on our SharePoint site. Nobody copies and pastes anything. Nobody retypes a phone number with two digits swapped.

    From there, the reports get compiled into a running history of each customer, a narrative of who they are and what we've done together, and that becomes part of our permanent record. Anyone at Summit can search it through Jeeves, our internal agent. Jeeves deserves his own post, and he'll get one.

    One detail matters more than all the others. Brain Picker can only see what the person running it can already see. Nothing new gets opened up, so the security auditor can put the fainting couch back in storage. That's a real limit, and it's also why the whole thing works as a team sport. My inbox knows the executives who sign contracts. Our technicians' inboxes know the people who call when the printer develops a personality disorder. The answers live in different places, so different people have to ask.

    How the team uses it

    We made the job as small as humanly possible. The team opens a list on SharePoint called the Research Queue, finds a name they recognize, copies a ready-made prompt, pastes it into Brain Picker and clicks Confirm. If you can order a burrito on an app, you are wildly overqualified.

    Behind the scenes, a small program checks for new reports every 15 minutes. When it finds one, it fills in the blanks in our master contact list and pulls the name off the queue once that contact is complete. It only fills empty fields. It never overwrites something we already had. And it's a snob about job titles, in the best way. Every title it finds gets graded on where it came from. A signed agreement outranks the person's own email signature, which outranks a passing mention in a Teams chat, which outranks the company directory, which most organizations update with the urgency of a Blockbuster membership card. Newer sources beat older ones. A blank title only fills in when the source is solid or when two separate sources agree. A title we already had only changes when a top-shelf source says otherwise, and even then the old one stays on file in case the new one turns out to be a typo with ambition.

    The first morning the team used it, they ran 24 lookups before 9:20 a.m.. The real yield was about one useful new fact per lookup. That won't make the evening news. But every one of those facts was a hole we no longer had to walk around, and every one came with a source attached, so anybody could check it.

    The answer was in the contract

    The moment I knew this was worth rolling out came before the team ever touched it.

    I ran Brain Picker on a long-time customer whose contact records were thinner than a gas station napkin. Titles blank. Direct lines blank. The report came back with titles and phone numbers for both contacts, and the source made me laugh out loud. It was our own signed service agreement, which lists the customer's contacts in an appendix. We had filed it with the solemnity of a time capsule and opened it about as often.

    Your contracts know more than your CRM does. They're just terrible at volunteering it.

    What went wrong

    I won't sugarcoat this part.

    It started quoting itself. Every report Brain Picker saved became something it could find the next time someone asked about the same customer. So it would read its own earlier guess, treat it as a second source, and grow more confident, like a cable news panel booking its own previous segment as the expert guest. We fixed it by hiding the report library from search. People can still browse it. The AI can't cite it as evidence.

    It promoted a customer to "Approver." On day one it read our own Research Queue, spotted a label we use internally, and reported it as a real job title. For about an hour, a perfectly nice person carried a title that sounds like a minor Marvel villain. We caught it because we read every change the system made to the contact list, put the old values back, and added rules that refuse labels like "Contact" or "Employee" as titles. Read what the AI writes before it becomes the record. Every time.

    Some people are simply unfindable. One contact went through eight separate lookups without a phone number turning up anywhere. Somewhere out there is a person living like it's 1994 and Caller ID is a passing fad. That one needs an actual phone call.

    And it isn't free. The parts that read email and Teams require a paid Microsoft 365 Copilot license for each person using it. We built it for people who already had one.

    The bigger gap: what walks out the door

    Contact details turned out to be the warm-up act.

    The gap that should keep every owner up at night is the knowledge that lives in about four people's heads. How a customer's network was wired five years ago. Why you stopped buying from a certain vendor. What somebody promised in a meeting nobody wrote down. None of it is in a system. When one of those people leaves, that knowledge walks out the door with their badge, their parking spot and the good stapler.

    We had plenty of that at Summit, and we decided to go get it before it left.

    So we built a second version called Context Research. I pick a topic and the people who know it. Each of them gets one task in Microsoft Planner. To hand over years of what they know, they do this:

    Click one: open the link in the task. Click two: press Start my research.

    Two mouse clicks. Declining a calendar invite takes more effort. (The very first time, Copilot asks each helper for permission to save, so the debut costs a few extra clicks. After that, two clicks forever.)

    Then three helpers go to work on that person's own email, chats and files, and each one files what it finds, word for word, with sources, into one shared folder. Nobody writes a memo. Nobody sits through a two-hour knowledge-transfer interview that everyone treats like jury duty. I pull the pieces together into a single document the whole company can use.

    Two clicks from the person who knows, and the company keeps what they know for good. I can't think of a better trade in this business.

    Why this matters for a business like yours

    None of this needed a new platform, a data warehouse or a consultant camped in your conference room for six months billing you for the privilege of learning your business. It runs on Microsoft 365, which most of our customers already own. It's hard to beat what you're already paying for.

    It's also built to leave. One script removes every piece we created, and the reports already saved stay put. When the job is done we retire it cleanly, instead of keeping it around forever like a treadmill that became a coat rack.

    The value shows up in three places. Time, because the digging that used to eat someone's week now happens in the background while they do their real job. Trust, because every fact carries its source, so you can check anything in seconds. And safety, because the AI never sees more than the person asking could already see.

    What I'd tell another CEO

    Brad Hutchison, our vice president, was bragging about Brain Picker to an industry partner over dinner tonight, and the questions didn't stop for fifteen minutes. When the person across the table lets their steak go cold to ask how something works, you write it down.

    Most companies are sitting on far more answers than they realize. They're buried in signatures, attachments and meeting recaps, spread across people who each hold one piece. AI is very good at the patient, unglamorous work of finding those pieces, as long as you tell it exactly where to look, make it show its sources, and read its work before you trust it.

    If you're curious what AI could do with what your business already knows, let's talk. The answers are probably already in your inbox. Getting them out isn't about buying more software. It's about asking the right people the right question, and letting the machine do the digging.