The founder of a small B2B company told us in a first meeting that they had already given Claude their ideal customer profile, their qualification rules and a connection to the CRM, and let it decide which accounts to work. Halfway through the demo they asked one question: how far away is your MCP? They wouldn't commit to a single seat before they had tested us through their assistant. And they were honest about why. They wanted everything inside the chat.
That founder is not alone. In the same few weeks:
- A buyer whose current vendor had just launched an MCP server said they started playing with it the same day, and that every tool of this kind would have to offer one "or go under."
- A CEO of a consultancy records every meeting, pastes the transcript into a trained Claude project and gets summaries good enough to make them ask what a vendor adds.
- A company had just bought a company-wide AI account and listed its second priority for the autumn as benchmarking Claude against Perplexity. Sales tools came after that.
- The manager of a meeting-booking agency said:
"The MCP connection sounds really cool. Honestly, that was what drew me in most from the very first call."
The order of the purchase has flipped. A few years ago a team bought a CRM, a data tool and a dialer, and then wondered how to fit AI in. Now the team picks one AI assistant for everyone first, and every other tool is judged on whether it can plug into it.
What changed: the assistant became the front door
When a company standardizes on one assistant, that assistant becomes where people start their work. They ask it what happened with a customer, which companies to call next week, what to write in a follow-up. A sales tool that only lives behind its own login becomes one more tab to remember, and the people who have learned to ask the assistant first don't open it.
The thing that lets an assistant reach into another system is called MCP (Model Context Protocol), an open standard that Claude, ChatGPT, Copilot and a growing list of other assistants support. A tool that offers an MCP server lets the assistant search it, read from it and, if you allow it, change things in it, signed in as you. That's why "do you have an MCP?" now comes up in first meetings, often before pricing.
For the buyer, this changes what to evaluate. The question isn't "which tool has the smartest AI." Your team already picked the AI. The question is which tools give your assistant something it can't get on its own, and let it work there safely.
What the do-it-yourself version quietly misses
If you can connect Claude to your CRM and ask it to build a list, why buy anything else? We wrote about the week this feels like a superpower and where it wears off. Seen from the buyer who picked the assistant first, five things go missing.
The data isn't on the open web. An assistant that searches the internet finds what companies publish about themselves: the website, the press release, the CEO. It doesn't find the head of finance at a 40-person company, the direct number, or the fact that the company was added to the register last spring. A sales leader at a training company put it simply:
"We use AI assistants a lot, but you're right. You can't rely on a general AI alone to find the right people."
The export-and-paste loop eats the time it saved. A sales manager at an IT consultancy told us:
"I have people right now trying to hack that together in ChatGPT: exporting CRM lists and trying to find lookalikes by hand."
Export, paste, ask, copy the answer into a spreadsheet, check against the CRM, import. Every step works. Together they turn a five-minute question into an afternoon, and the result is a file that is out of date the day someone else touches the CRM.
The chat isn't a record. A great answer in a chat thread lives in that thread. Your colleague can't see it, your CRM doesn't know it happened, and in six months nobody can find out why the account was dropped. Selling needs one place where what was said, decided and promised is kept, and a chat history isn't that place.
The assistant sees what the connection sees. If you wire an assistant straight to a database with an admin key, it reads everything, including the private conversations and the deals other reps own. That's a decision nobody would make on purpose for a new hire, and it gets made by accident for an assistant.
Some things it can't produce at all. A consultancy CEO's homemade meeting summaries were excellent. Their phone calls, though, weren't recorded, and as they put it, "that's a difference." An assistant can summarize what it's given. It can't give itself a call recording, a direct number or a verified register match.
None of this argues against the assistant. It argues for being picky about what you connect to it.
Seven questions for any sales tool when your assistant comes first
Ask these in the first meeting. A vendor who has thought about it will answer in a sentence each. A vendor who hasn't will start talking about their own chatbot.
1. Can my assistant connect to it, and which assistants? Ask specifically about the one your company uses. "We have an API" isn't the same answer: an API needs a developer, an MCP server needs someone to paste a URL and sign in.
2. Does the assistant see exactly what I see, and nothing more? The connection should sign in as the user and follow the same permissions as the app. If a rep can't open a colleague's private meeting in the tool, their assistant shouldn't be able to either. Ask what happens when someone leaves.
3. Can it read the selling, not just the data? Company and contact data is the obvious part. The valuable part is the record: what the buyer said in the last meeting, which deals have no next step, which replies are owed. If the assistant can only reach the database, you're back to pasting transcripts.
4. What can it change, and what does it ask first? Reading should be generous. Writing should be narrow and visible. Find out which actions an assistant can take, whether anything is sent to a buyer without a human, and whether anything that costs money happens without a confirmation.
5. Does spending need my yes? Many data tools charge per revealed contact. An assistant that is asked for "the phone numbers for these 200 people" will try to get them. The tool should tell the assistant the price first and need an explicit confirmation before credits are spent.
6. Where do the results end up? A list built through the assistant should land as a list in the tool, where your team can call it, not as a table in a chat thread. The same goes for a saved search or a target profile. If the answer lives only in the chat, you've built another spreadsheet.
7. Can I see and cut off the connection myself? You should be able to see which assistants are connected, which account each one reads from, and revoke one in a click without calling support.
And one more, the one that decides whether the tool is worth paying for at all: what does it give my assistant that it can't get by itself? If the honest answer is "a nicer interface," the assistant has already replaced it.
Where the assistant should stop
Picking the assistant first is a sound decision. Letting it do everything isn't. Three lines are worth holding, whatever tools you buy:
- A human approves before anything reaches a buyer. Drafts are fine. Sending is a person's call.
- One system keeps the record. The assistant asks and acts, but what was said and decided is written into the CRM or sales platform, not left in a chat.
- Changes to what the team treats as fact get a second look. A deal's stage, a qualification answer or a contact marked as gone affects everyone who reads it next.
These are the same three rules we suggest in our guide to starting with AI in sales, and they matter more, not less, when the assistant is the front door.
The afternoon test
Before you sign, run this with a trial account and the assistant your team already uses. Use your own questions, not the vendor's demo script.
- Connect it. Time it. If it takes more than ten minutes or needs a developer, note that.
- Ask for a list you actually need. "Find 30 companies like our three best customers in the Gothenburg region, with a head of finance." Check five of the answers by hand.
- Ask about a real customer. "What did we last discuss with them, and what did we promise?" Is the answer sourced, and is it right?
- Ask for something it shouldn't do. Ask for a colleague's private meeting, or for contact details on 500 people. Does it stop, or ask for a confirmation and show the price?
- Look in the tool afterward. Did the list land there? Can a colleague open it? Is the connection listed where you can revoke it?
If all five go well, the tool works with your assistant. If step 2 fails, the tool doesn't add data. If step 4 fails, don't connect it to anything that matters.
What we built for teams that start in the assistant
We're an AI-native sales platform, and we built our MCP server because customers like the founder at the top asked for it before they asked about price. This is what it does today.
Works with the assistant you already use. Claude, ChatGPT and other assistants that support MCP connectors connect with a URL and a Funnelfeedr sign-in; no client ID or secret. Microsoft Copilot connects through Copilot Studio, set up by an administrator. On a Claude Team or Enterprise plan, an owner adds it once and everyone else just clicks Connect.
The assistant gets what the person gets. The connection signs in as the user, to the account they pick, and follows the same permissions and seat as in the app. A private conversation is neither listed nor counted. Under Settings › MCP connections you see every connected assistant, which account it reads, when it was last used, and you can move it to another account or revoke it.
Data an assistant can't find on its own. Company search across the Nordic markets with the main filters from the app, ranked by your ideal customer profile or by similarity to companies you name, plus lookups by name, domain or organization number, and the contacts we know at each company.
The selling, not just the data. The assistant can search your meetings, calls, emails and CRM notes ("who mentioned pricing", "what did they say about the integration"), open one with its summary, action items and transcript, look up deals with stage history, contacts, tasks, replies owed and qualification, list what you owe people, see how often each objection came up, and search your knowledge base.
Narrow writes, priced spending. It can create lists, add or remove companies and contacts on them, save a search and create an ideal customer profile, so the results end up in Funnelfeedr where the team can call them. Contact emails and phone numbers stay masked until revealed, and a reveal is two steps: the assistant first gets the price in credits, then needs a confirmation before anything is spent.
What it doesn't do. It doesn't send emails, change deals, write to your CRM or start calls. Those stay in the app, where a person approves them.
Already running your week from Claude or ChatGPT? Book a demo and we'll connect your assistant during the call →AI assistant · MCP · Claude · ChatGPT · Copilot · sales stack · sales tools · evaluation · B2B
