An ops manager at a 40-person distributor we talked to last spring had just turned on AI automation to auto-approve reorders under $500. Three days in, it approved a $4,200 reorder for a part that had already been discontinued. The AI hadn't hallucinated anything. It read the inventory count in the ERP, which said 0, checked the "active SKUs" tab in a shared spreadsheet, which still listed the part as live, and made the call the spreadsheet told it to make. Nobody had updated that tab in six weeks.
This is the pattern behind most failed AI automation projects at small and mid-size businesses right now. The AI isn't broken. The data it's reading is.
Why AI automation quietly falls apart on scattered data
Every SMB we've worked with has some version of the same setup: a CRM for contacts, QuickBooks for money, a shared drive full of spreadsheets for whatever the CRM and QuickBooks don't cover, and Slack threads holding the context that never made it into any of them. None of these systems agree with each other, and until recently that was fine, because a human was always in the loop to notice when the numbers didn't match and go figure out which one was right.
AI automation removes that human. That's the whole pitch: no more manual reconciliation, no more re-typing the same order into three places. But an AI agent doesn't know which of your three "truths" is actually true. It picks one, usually whichever system its integration happened to be pointed at, and acts with full confidence. A support bot pulls a shipping address from the CRM instead of the more recent one in the order system. A pricing agent quotes last month's rate because the spreadsheet with the new pricing hasn't synced to the tool it's reading from. None of this shows up as an error. It shows up as a wrong answer delivered fast and with total confidence, which is worse than a slow answer that's right.
The common fixes, and why they don't hold
Most teams try one of three things once they notice the problem, and all three treat the symptom instead of the cause.
The first is more automation, adding a Zapier or Make workflow to sync data between the two systems that disagreed. This helps for exactly as long as it takes someone to add a fourth tool the sync doesn't cover. We've seen companies running 15+ of these point-to-point syncs, each one a separate thing that can silently break, with nobody watching all of them at once.
The second is a bigger, smarter AI model. Teams assume the model is the weak link, so they upgrade from GPT-4o to a frontier reasoning model, or add a "verification agent" to double-check the first agent's work. A smarter model reading bad data just produces a more articulate wrong answer. It doesn't matter how good the reasoning is if the premise it's reasoning from is stale.
The third is a data cleanup sprint. Someone spends two weeks manually reconciling the CRM against QuickBooks against the spreadsheets, gets everything matching, and declares victory. This one actually helps, briefly. Then a new order comes in through the old process, gets entered in one system and not the others, and you're back to three versions of the truth within a month. Cleanup without a structural fix is just cleaning the same spill on a schedule.
What "connected" actually has to mean before automation is safe
The tradeoff nobody wants to hear is that connecting your systems takes longer than turning on an AI tool, and it doesn't feel like progress the way a new chatbot does. But it's the difference between automation that saves hours and automation that creates a new category of mistake at machine speed.
Here's the bar we hold internal automation work to before letting an AI agent write or act on anything, not just read:
- One system of record per entity. Pick which tool owns "customer," which owns "order," which owns "inventory count." Not two. Everything else reads from it or writes back to it. It doesn't maintain its own separate copy.
- A defined sync direction and frequency. If QuickBooks is the source of truth for invoice status, the CRM should pull from QuickBooks, not the other way around, and you should know whether that sync runs every five minutes or every night.
- Reconciliation rules for conflicts. When two systems disagree, and they will during the sync window, there needs to be a rule for which one wins, not a person eyeballing it after the fact.
- An audit trail on writes. If an AI agent updates a record, you need to see what it changed, when, and based on what input. Without this, debugging a bad automated decision means guessing.
- Permission boundaries that match the automation's scope. An agent that can read pricing shouldn't automatically also be able to write pricing changes to the system customers see, unless you've decided that's acceptable and built guardrails around it.
If you can't answer all five of those for a given workflow today, that workflow isn't ready for an AI agent to act on it unattended. It might be ready for AI to draft something a human approves — that's a much lower bar, and often the right first step.
A decision path that actually works
When a client asks us where to start, we walk through this in order, and we don't skip steps to get to the AI part faster, because skipping steps is how you get the $4,200 reorder problem.
First, map where each type of data currently lives and how many copies of it exist. Most teams are surprised to find customer contact info alone living in four places: the CRM, the invoicing tool, a support inbox, and somebody's personal spreadsheet of "VIP accounts."
Second, pick the system of record for each entity type based on which tool already does the best job of it, not which one is newest or most impressive. QuickBooks is usually the system of record for money. Your CRM is usually the system of record for the relationship. Resist the urge to consolidate everything into one giant platform. That's a different, much bigger project, and it's not a prerequisite here.
Third, build the connections between systems before you build anything AI does. Get order data flowing automatically from your e-commerce platform into QuickBooks. Get support tickets flowing into the CRM as activity. This step alone usually removes 5 to 10 hours a week of manual re-entry, and you'll feel that before AI even enters the picture.
Fourth, only once data is flowing reliably and you trust it, layer in automation. Start with read-only or draft-and-approve actions, and only move to fully autonomous actions on workflows where a wrong call is cheap to reverse.
What this looks like in practice
A property management company we worked with had leasing data in AppFolio, prospective tenant info in a shared Google Sheet their leasing agents kept because AppFolio's lead form felt clunky, and maintenance requests coming in over text and email. They wanted an AI assistant to triage maintenance requests and auto-schedule vendors. Sensible idea. Not viable yet, because "which unit needs which repair" existed in three disagreeing places.
We spent the first three weeks doing the unglamorous part: making AppFolio the system of record for units and tenants, building a real intake form that fed maintenance requests directly into it instead of text messages, and setting up a nightly reconciliation check that flagged mismatches instead of letting them sit silently. Only after that did we turn on the triage assistant. It now correctly routes about 90% of requests without a human touching them, and the ones it can't handle get flagged instead of guessed at.
That ordering — connect first, automate second — is the whole difference between an AI project that pays for itself and one that generates a new kind of mess.
If you're staring at a similar tangle of tools that don't talk to each other and wondering whether AI can help before you sort it out, it's worth getting a second set of eyes on the actual data flow first. We do a free 30-minute Process Teardown where we map one of your painful workflows end to end and show you, in hours and dollars, what it's actually costing — no pitch, no obligation. You can also look through some of the systems we've connected for other companies before adding AI on top, to get a sense of what "ready for automation" actually looks like.
0 Comment