Your AI is giving your team answers they can't trust. And they know it.

Since May I've run a series of roundtables with GTM leaders, consultants and builders. I asked how AI implementations were going, and the same things kept coming up.

Sound familiar?

Here's what they told me.

Of everything people brought up, this came up the most.

A consultant has a client who pulls sales data out of the CRM every Monday and wants the same sort and filter in Excel each week, that way every rep gets their own pipeline in Power BI. The client has been at it for months. AI still won't repeat the same filter twice. "That seems like a fairly simple action," the consultant told me. It isn't.

One attendee wrote a decisions file to teach Claude how he likes to work. A couple of weeks later he asked why it kept ignoring him.

Claude: "You actually told me five different times, and you told me to do it this way twice, and you told me to do it this way three times, so I pretty much just pick randomly."

His reaction: "So I'm the problem is what you're telling me."

Another consultant watched a CFO load a whole file into AI and present the analysis to the room. "Look at all this analysis, this is true."

The math doesn’t math.

"It just makes math up."

In June, one company let reps connect Claude to the CRM through the app marketplace. A rep asked it a question. It went and changed a deal stage in the CRM.

That hit comp.

Session after session, someone said it out loud: people trust what the AI hands them and stop thinking.

An example.

Host an in-person conference.

The old way -> look at the attendee list, ran it through their ICP and routed accounts to reps.

Now -> AI builds a dynamic ICP from the event content and hands the reps their lists. The reps follow up, and nobody asks anymore why an account made the list.

"It's kind of taken away that aspect of thinking and review."

Here's a typical scenario, and more than one attendee described it.

A company buys a Claude team plan and gives it to literally everyone. The quality of what got built started to drift. One consultant saw Claude Max accounts go to every Salesforce user, reps and CS included. Half of them held admin licenses, because nobody ever set up the permissions. Connect Claude to Salesforce through MCP, tell it to delete what you don't need, and it will. Now the consultant takes calls from people who've had some "boo boos." A consultant's words: "legit monkeys with guns." Someone spotted a Gong API key that carries full delete power.

That risk is sitting there.

Another consultant found 1,000 contacts imported with no phone number or email.

Access and governance came up on their own again and again.

The Wild West is real. One attendee described two modes living side by side: the official tools everyone is supposed to use, and "the Wild West, where anyone can build anything."

Another watched "anyone could create a skill, anyone can pull data from somewhere" turn into a drop in the quality of what people built.

A startup gave everyone access and the CMO called it "an absolute sh*t show."

Another attendee caught employees running company data through personal paid ChatGPT accounts, PII included.

Yet another consultant described a pharma company building a governance layer, because one team seeing another team's data costs millions in penalties.

Data quality came up over and over.

One consultant had a prospect ask for a sales forecast. Their answer: you've got to be kidding me. No clean pipeline, CRM not fixed, 25 things that decide the number, and politics on top.

A win/loss analysis came out skewed because the AI never saw the email threads. An account hierarchy came back hallucinated.

One attendee's team has been "switching tools constantly." Claude is their latest bet.

Another built a bulk tool in Replit, then the platform shipped an MCP and the whole job became a click for an agent.

Their take on the LinkedIn firehose: "a little exhausting sometimes."

A third makes ChatGPT and Claude fight over an idea and picks a winner, and sometimes both lose.

The question underneath all of it: which one is the most reliable and takes the least maintenance.

You knew this was on the list right?

I've been surprised by how often people think that the rise of these tools is going to completely obliterate their spend.

Token spend came up on its own, more than once.

As an example, someone runs local models for the heavy lifting because "Claude is very expensive," saves Claude for the hard problems, and lets the local model grind overnight.

A consultant watched a new model release come out "very token hungry" and wants to catch spikes like that and adjust, since the bill runs at runtime on their clients' systems.

Another applies the old RevOps rule, use the least complicated tool, to model choice. Moving from Opus to Haiku can mean a 20x difference, so their workflows only call the model the job needs.

One attendee switched from Claude to Codex because it takes fewer tokens.

Outbound is turning into the same email from everybody.

One attendee said it plainly: "outbound messaging is becoming more homogenized, where everybody's getting the same messaging from all these different companies."

Then asked a question that matters: “maybe we put our effort in the wrong place trying to hyper-personalize.”

Another sees the slop where people feed AI junk and let it write the copy.

Their rule: use it for the outline and the data, then add the human touch.

That leaves me with the question… how do you do that when you don’t own the tech?

I’d love your thoughts on that.

And now the run downs…

Here's what went sideways.

Here's what attendees built on their own:

Most are one-person builds. Few of them could show me a measured result.

Some of it pays off sometimes… 20-30% of the time according to the research.

Here's what attendees told me is working for them.

A consultant said it out loud: SaaS is in real trouble.

Salesforce announced Headless 360, which tells you that you won't need their UI because you'll build your own and keep their data architecture.

Partners are building context layers to stay sticky.

And as the consultant put it, "if I could build everything myself and I didn't have to pay for all of these tools... why the hell would I not?"

Another attendee has met post-seed companies that want to run AI native: a spreadsheet, a place for data, Claude Code, no other tools. A third is building for a future where the main interface is a Codex or Claude app, not a website.

My take: If you aren't adding real value as a piece of software to whatever's going on in the database, you're a dead product in the next three years.

One attendee, a few months into a new role, uses AI to build account plans, territories and ramp models for a team of six reps, so they can defend real numbers to the board instead of fluffy forecasts.

CROs commit to another 30% because last year was 30%.

The plan shows where the pipeline comes from.

Their quick win with the executives was a DocuSign integration that puts the contract-sent date in a CRM field.

Another attendee pointed out that RevOps usually plays the realist when a PE board assumes AI means even more growth.

My take: the jury is really out here. I mean really. We commit to forecasts that look like last year's growth but what is that based on? History… and we know how that goes. And still there Are workflows that are creating real impact. Maybe the point is that we don't actually understand how to measure them properly. And quick wins are available. Absolutely all over the place.

When you've been running an AI company as long as I have, you have seen literally everybody talk about how your stuff was going to magic them a gajillion dollars. And we're starting to see that expand, and revenue teams are starting to need to play the bubble bursters.

The same question kept coming back: whose job is AI now? Who decides what gets built, and who gets access? Some of the people asking are between jobs after a layoff. And more than once, someone told me they wanted real conversation with peers.

Education sits on the front line.

An attendee says clients arrive one of two ways:

Another watched a CRO and a VP expect Gemini to hand over the answer. The reply: "You didn't hire me for generic."

And people who never wrote code are now talking about GitHub and BigQuery tables.

My take: No one seems to be cautious about this brave new world. And it's literally turning everything on its head. And we're watching people who don't know how to do tech do tech. And hope for the best.

In May the worry was everyone running loose: homegrown tools of uneven quality, who gets access, how to know if any of it works. By summer the problem changed shape. People re-taught their agent the same pipeline every session, fought memory drift, and watched saved instructions pile up. Wrong answers and tool churn stayed on the table the whole way.

My take: We saw some shifts there maybe. Because the reality is, the sample size is not huge. And just because the people who are showing up in my roundtables are reporting that they're working on context and prompting properly and those sorts of things, doesn't mean that there aren't still a lot of other people struggling with the same issues of homegrown tools of uneven quality, access problems, and trying to learn how all of it works. People who show up in MySpace tend to have some AI experience.

Attendees want three things.

These are today's problems.

Here's the thing. AI works with you, and your team already knows it. They're the ones re-teaching it, checking it and patching it. Hit me up if you'd like to see what getting my help looks like.

Here's how often each topic came up on its own across the sessions.