Agentic AI Needs Organization Before Technology
Before buying an AI agent, a company needs to decide who decides what, how rules are written down, and how know-how gets passed on.
The agent arrives before the rules do
A mid-sized distribution company wants an AI agent to handle reorder decisions: watch stock levels, flag shortages, place purchase orders automatically. The pitch is simple and the demo works. Then someone asks: who approves an order over 5,000 euros? What happens when two suppliers quote different lead times? Who do you call when the agent orders the wrong quantity because a product code changed last month?
Nobody has a clean answer, because nobody had to answer it before. The operations manager made those calls in her head, mixing a few written rules with a lot of judgment built up over years. That judgment was never written down because it didn't need to be — she was always in the room.
An AI agent removes her from the room. It still has to make the same calls, but now it needs the logic spelled out: thresholds, exceptions, escalation paths, who gets notified, what counts as a reasonable deviation. If that logic doesn't exist on paper, the agent either makes up its own version of it — badly — or it freezes at every edge case and gets routed back to a human, which defeats the point of automating anything.
This is the part that gets skipped. Teams evaluate agent platforms on accuracy, integration options, and price per call, and barely touch the harder question: do we actually know, in writing, how this decision gets made today? Most of the time the answer is no. The decision lives in someone's head, in a Slack thread from eight months ago, or in a spreadsheet nobody has opened since the person who built it left.
Why the technical layer isn't the hard part anymore
Connecting an agent to a CRM, an ERP, or an inventory system is largely solved. APIs exist, integration tools exist, and the cost of wiring systems together keeps falling. What doesn't get solved by better software is a company that has never formalized how a decision gets made, who owns it, and what "good enough" looks like for an edge case.
An agent that reorders stock, qualifies a lead, or drafts a contract clause is not doing something new — it's doing something a person already does, at a pace and volume no person can match. If the underlying rule was always informal, scaling it with AI just scales the inconsistency. The company doesn't get a faster version of a good process; it gets a faster version of whatever was actually happening, flaws included.
This is why pilots stall after the demo. The demo runs on clean, pre-selected cases. Production runs on the messy ten percent — the supplier who changes terms mid-contract, the customer who asks something outside the script, the invoice with a typo in the reference number. Handling that ten percent was never a model problem. It's a governance problem: nobody defined who the agent escalates to, on what grounds, and how fast.
What to do before you buy an agent
Treat the agent purchase as the last step, not the first. The groundwork is organizational and it has to happen before a vendor conversation goes anywhere.
- Pick one decision the agent would make and write down the current rule exactly as it's applied today — not the rule as it should be, the rule as it actually is, including the exceptions people make without saying so out loud.
- Name a single owner for that decision who signs off on the written rule and who the agent escalates to when a case falls outside it. If two people can override the rule differently, fix that first.
- Set a threshold for what the agent handles alone versus what goes to a human, in concrete units — an amount, a time window, a confidence score — not a vague "use judgment."
- Check where the data the rule depends on actually lives, and whether it's current. A rule based on a supplier list from two years ago will produce confident, wrong decisions.
- Run the written rule manually for a short period before any agent touches it, to confirm it produces the outcome you expect across a handful of real edge cases, not just the obvious ones.
This work is slower than installing software, and it exposes gaps that are uncomfortable to look at — a rule that contradicts another team's rule, a decision nobody actually owns. That discomfort is the signal the groundwork is doing its job.
What tells you it's working
Once the agent is live, the metric to watch isn't how many tasks it completes — it's how often it escalates, and whether those escalations make sense. A rising escalation rate on genuinely new situations means the rule is holding up. A flat rate while errors pile up downstream means the written rule doesn't match reality and needs revising, not the model.
Talk to us about the groundwork, not just the tool
ArkonLabs builds custom business software and measured AI deployments for companies that need the decision logic sorted out before any agent touches it — not a generic platform bolted onto an undocumented process. If that groundwork is where you're stuck, reach out through www.arkon-labs.com.