If you run a mid-market company right now, there are probably a few AI pitches sitting in your inbox this week. A consultant with a slide deck, a software vendor with a slick demo, maybe your MSP offering a new AI package, or someone on your own team who is excited and wants to lead the charge. They all sound confident, and from the outside it is hard to tell who actually knows what they are doing.
We talked about this on our last live stream with Sagar and Alex, and the gauge I keep coming back to is pretty simple. Watch whether the person is willing to have the difficult conversations about the messy work that has to happen before AI touches your data. In my experience, that willingness is probably the number one sign they know what they are talking about. If they jump straight to tools and integrations, they are skipping steps, and skipped steps tend to come back and bite you, sometimes slowly and sometimes very quickly.
Tools and integrations are the easy pitch
From the service side, it is honestly easier to lead with the shiny stuff: the best new tool, the integrations they can turn on, the agent that will handle your inbox. Those conversations are fun, they demo well, and everyone walks out of the first meeting feeling like progress happened. I get why people go that route.
The trouble is that a great demo runs on clean sample data with nobody's permissions in the way. Your company has years of files in odd places, a CRM with three spellings of the same customer, and folders nobody has opened since the person who built them left. Someone who knows this work will bring that up early, even if it makes the first meeting less exciting. Someone who doesn't will promise results and let you find the mess later.
Messy data costs you every time AI reads it
One thing people miss is how much the state of your data affects cost. In theory you can point AI at everything and let it sort things out, and it can help with some of that. You still pay for it. Think about how much intelligence, and how many tokens, goes into the AI just figuring out what a piece of data even is, every single time it reads it. That cost keeps going, because your data keeps changing and somebody has to maintain it.
When a little groundwork is done first, the AI spends its effort on the actual task. A good partner will walk you through that tradeoff in plain terms and help you decide which data is worth cleaning up now and which can wait. If nobody brings up data readiness at all, you'll probably find out about it later through a higher bill or answers your team can't trust.
"Give AI access to everything" is a red flag
A lot of companies love the idea of a company brain, where everyone can pull from everything the business knows. I love the power of that too. You still don't want a random employee pulling up the CEO's email, and that example is less exaggerated than it sounds. The mindset of throwing all the data in, giving AI access to everything, and figuring it out as you go is where a lot of these projects go wrong.
Sagar made a point on the stream that stuck with me. AI will find out that last summer's interns still have access to the executive finance folder, even when no human on staff knows it. It doesn't care how deep a file is buried in SharePoint. Alex added that access decisions belong to the people who own the data. HR decides who sees HR data, finance decides for finance, and IT handles the admin side underneath. So ask anyone you're evaluating how they would handle permissions before connecting anything. If the answer is vague, or it sounds like "we'll set up some admin roles," keep asking.
Governance starts as a conversation between people
When people hear data security, unstructured data, and governance all in one breath, it can feel overwhelming, and the natural reaction is to go shopping for software that promises to handle it. Software can help, but the core questions still take real human conversations: which data lives where, who owns it, who should see it, and what happens when someone asks for access. Even if you get the technical side of the data right, those questions need answers first, or at least alongside the build.
So when you're talking to a consultant, vendor, or MSP, pay attention to whether they want your stakeholders in the room. The good ones will ask to talk with the people who own the data and run the workflow. They'll also push back if leadership wants to wave a policy through just to close a deal faster.
Use the same test on your own team
This applies whether the person leading AI is external or internal. Plenty of companies hand AI to an excited employee or to IT and hope it works out. That person might be great. Ask them the same thing you'd ask a vendor: are they skipping steps? Can they explain how they'll handle data cleanup, access, and ownership, and are they comfortable telling leadership that some ugly work has to come first?
As a business leader, making that call is your job. You can make it without being the technical expert in the room, as long as you notice when someone jumps to the end result without explaining how they'll get your data and permissions ready.
Questions to bring to your next AI meeting
Before you sign anything or buy another subscription, try these in your next conversation with a vendor, consultant, MSP, or internal lead:
- What do you need to know about our data before you connect anything?
- Once AI is reading our files, how will you decide who can see what?
- Who on our side needs to be part of those decisions?
- What cleanup do you expect up front, and what does it save us later?
- What would make you tell us to slow down?
Pay attention to whether they lean into these questions or steer back to the demo. In my experience, the people who lean in are usually the ones who have done this before.
If you want help that starts with the messy part and still gets you to AI that works in your business, that is how we approach it at Spej. Talk with the Spej team.
