The sifting is the hard part

After working at Amazon, it became clear:

Amazon is a universe of systems that follow standard operating procedures.

It basically works like this: an employee has to solve a task. The task falls under a specific team. That team has an SOP. To solve the problem, you need to work with X number of other teams — and each of those teams has its own SOP for how outside teams are supposed to partner with them.

It's high-process work. The thinking isn't hard — the sifting is. Finding, identifying, and gathering the correct process takes forever.

Take the chargebacks process as an example. It can drag on for years: pulling evidence and data, finding the right finance team, writing up the case, getting it approved across several layers of leadership, getting the case opened, having a reviewer review it, following up, getting the findings, and explaining those findings to the vendor.

In that same stretch, a vendor may have cycled through several CSMs.

With the right system access, permissions, and human review, much of this work can potentially collapse from months to days.

That's one example. Now multiply it by every process at Amazon that follows this same pattern.

And it's not just Amazon, and it's not just retail. Take enterprise SaaS: before a big company signs a contract, they run the vendor through a security review. That means a security questionnaire (sometimes 200+ questions), a SOC 2 report, pen test results, a data processing agreement, sign-off from the buyer's security team, legal redlines on the MSA, procurement approval, and often a live call with sales engineering to walk through the architecture. Every team involved has its own SOP for how it wants to receive and review that information.

In many organizations, it can take four to eight weeks to close an enterprise deal on paperwork alone — after the buyer already said yes. Multiply that by every deal in the pipeline, and it's the same pattern as chargebacks: not hard, just slow, because the process is scattered across teams that don't talk to each other by default.

The real roadblock: understanding, not tooling

Here's the catch: before you can point an AI agent at a process, you have to understand that process manually yourself. I think this part is essential. If a team does not understand the underlying process, distrust of automation is reasonable: people cannot validate a system they cannot explain.

The antidote to distrust is understanding.

And understanding isn't hard. It's just the paragraph I wrote above — it takes a minute to explain to someone. Which is the next point: clear communication and storytelling to your team matters enormously for adoption. If people can understand the process simply, the team can evaluate where automation is appropriate, where review is required, and what evidence the system must provide. That's how you get an adopted customer.

Level two: AI as a diagnostic collaborator, not only an execution layer

The chargebacks example above is tool-oriented — AI executing a known process fast. But AI is closer to a colleague than a tool. Treat it like one.

Once you've mastered fixing chargebacks after they happen, you can have AI analyze your entire business — operations, product, packaging, everything — to stop chargebacks from happening in the first place. At that point you don't even need the recovery tool anymore; there's nothing left to recover.

This analysis would need to happen inside an approved environment with appropriate access controls, data-retention policies, and contractual permissions. The model should receive only the information necessary for the task.

Here's how you'd do it:

• Download all your labels, product specs, designs, carton designs, and packaging specs — everything official.
• Pull your last few POs and run them through.
• Describe the system and process your team uses to accept POs, pack, and ship product.
• Note which carrier you use to transport product.
• Download your efficiency metrics from Vendor Central.
• Feed in product specs: weight, cases per package, liquid or solid, plastic or not — the full picture of process, product, and packaging.

Give the AI complete knowledge of all three, and more often than not it can identify why you're receiving chargebacks — and propose how to fix it.

Once the data is available and appropriately governed, an initial diagnostic can take days rather than weeks. The real blocker isn't information anymore — it's obtaining the documents.

So why enterprise over SMB?

The distinction is a spectrum rather than a clean divide. Smaller companies often have less formal process sprawl, but more tribal knowledge and fewer dedicated owners. AI can help them document and standardize work that has never been written down — feed it how you currently (informally) do something, and it can turn that tribal knowledge into a real process: the SOP the SMB never had time to write.

Large enterprises face the inverse problem at greater scale: the procedures exist, but they are fragmented across teams, systems, approval layers, and permissions. Their opportunity is not simply creating process; it is retrieving, reconciling, and executing the correct process repeatedly. The chargebacks example isn't one problem to solve once; it's a pattern that repeats endlessly, at scale, with real dollars attached every time it repeats.

That's the opportunity. SMBs need AI to create structure. Enterprises need AI to cut through structure that already exists. The opportunity scales with the coordination burden: every repeated delay compounds across teams, cases, and dollars — which is why process navigation is such a valuable target for carefully governed AI.