Why the Best Enterprise AI Strategy Ends With Less AI
The customer quotations below come from recorded aviation-industry conversations. They’ve been lightly cleaned for readability, and names and company identities are omitted.
TL;DR
The smartest AI strategy may leave you with less AI in production. Use agents to get a dirty process moving, observe their decisions, and turn repeated behavior into deterministic workflow. Keep AI where interpretation and judgment matter. If a model must repeatedly rebuild context across disconnected systems, you’re paying a fragmentation token tax instead of buying intelligence.
Why are companies adding AI without changing the work?
We’ve reached the point where adding AI to a workflow can feel more responsible than asking whether the workflow should exist.
Tobi Lutke recently helped popularize the phrase “slop grenades” for AI-generated work tossed to somebody else without enough human judgment. The sender looks productive. The next person absorbs the cost.
Enterprises can do the same thing with a much larger budget. They bolt AI onto email, spreadsheets, portals, legacy ERP screens and disconnected databases, then celebrate because information moves faster. The model looks intelligent because it can cross the gaps the company never fixed.
The numbers should make us uncomfortable. In a GFT survey of 945 CIOs and CTOs, 84% said legacy-system limitations had caused their organization to cancel an AI pilot or project. A Wharton Human-AI Research and GBK Collective study found that 67% of departments use AI agents in some form, yet only 28% have AI fully embedded in core processes. Only one in five met all three workflow-transformation measures. Among those that did, 67% reported strongly positive ROI.
That gap matters. AI adoption and operational transformation aren’t the same achievement.
One aviation distributor described the human version of the problem:
“We’re spending 90% of our time finding the part, entering all the information and going through that sequence, all the way to shipping.”
The team didn’t need another intelligent surface. It needed the work to stop consuming the people hired to sell.
Why is faster movement between fragments a false win?
This was the point that became clearer as I thought through the argument: faster movement between fragments is a false positive. It can look like a win while the transaction underneath remains broken.
An AI assistant finds the email faster. It copies the PDF into the ERP. It reads a marketplace response, updates a spreadsheet and sends a summary to somebody in Purchasing. Every local step gets quicker, but the transaction still has no durable home. Customer intent lives in one place, vendor evidence in another, approval in a message and the latest status in the head of the employee everyone calls.
You improved interaction speed. You didn’t necessarily improve transaction continuity.
That distinction matters in aviation because the quote isn’t an isolated document. It can become a purchase, receipt, inspection, stock record, allocation, shipment, invoice, payment and years of aftermarket history. The part number, condition, trace, certificates, decisions and exceptions have to survive the trip.
One quality leader gave us the cleanest picture of fragmentation I’ve heard:
“I have a Google Sheet nightmare going on because I have 27 tabs.”
AI might search those tabs in seconds. The 27-tab operating model would still be there tomorrow.
How many tokens are you spending just to find the transaction?
Here’s the question I’d put in front of every CIO, operator and AI vendor:
How many tokens must you spend to discover where the transaction went, or to rebuild context the system should never have lost?
Before an agent can make a useful recommendation, it may need to identify the customer, match the part, find the current RFQ, reconcile vendor responses, locate the certificate, check inventory, recover an approval, inspect prior pricing and determine whether the last action actually happened.
You pay for that reconstruction every time. The bill arrives as tokens, latency, integration maintenance, security exposure, duplicate actions and human review. I call it the fragmentation token tax.
That calculation has to include more than model pricing. Count the connectors that must stay alive, the records copied into new stores, the permissions duplicated across tools, the failures somebody has to reconcile and the audit trail rebuilt after the fact. A cheap token can still carry an expensive piece of context if the business has to rediscover that context on every pass.
The market can mistake this work for intelligence because the agent looks busy. But finding five broken pieces and putting them back together isn’t the same as improving the decision. One competitor has said openly, and will remain anonymous here, that a meaningful portion of its AI intelligence wouldn’t need to exist if the underlying workflow were better. I respect the honesty. It proves the point.
An customer explained the cost in one sentence when they were initially evaluating us:
“We missed that one part. Sorry. And that’s the important part.”
Moving forty emails faster doesn’t help if the critical line still disappears.
Where does AI genuinely belong?
AI earns its place in three broad areas.
First, it can interpret unstructured inputs: language, email, PDFs, images and noisy records. Second, it can work at the edges where rules change too quickly or interact too dynamically to code economically. Third, it can support decisions made with incomplete information. Narrow models for classification, forecasting, retrieval and anomaly detection belong in the picture too.
Most other work should become deterministic when practical. Validate the field. Apply the permission. Update the order. Reserve the inventory. Generate the invoice. Record the event. Verify the outcome.
Agents give us a better way to reach that state. We no longer have to perfect the process before digitizing it. Start with the dirty workflow. Let an agent reason across it, under control, and get it moving end to end. Log every decision. Watch the exceptions.
Then pay attention to repetition. When an agent handles the same condition the same way again and again, you may be looking at a business rule that hasn’t been implemented yet. Extract it. Test it. Put it into deterministic workflow. Keep repeating the cycle until the AI is doing the work that still needs a brain.
The agent is scaffolding. The process is the building.
An aviation operator once showed us what happens when the scaffolding becomes permanent:
“Some items get left behind even though we have to deliver by a certain date. Right now, that’s mostly tracked on a Google Sheet.”
The objective isn’t a smarter way to monitor the workaround. The objective is to make the workaround unnecessary.
What changes when the transaction stays connected?
This is where ERP.Aero belongs in the conversation, and only here, after the operating principle is clear.
ERP.Aero’s role isn’t to become another intelligence layer floating above a fragmented aviation stack. The stronger architecture keeps the transaction connected from demand and pre-RFQ activity through quote, sourcing, purchasing, receiving, inspection, inventory, repair when required, quality, shipment, invoice, settlement and aftermarket history.
That doesn’t mean every aviation company eliminates every outside system. Marketplaces, carriers, payment services, customer channels and specialist tools will remain. The difference is that the transaction has an authoritative operational home. External signals can enter it without forcing the agent to recreate the business every time somebody asks a question.
iRFQ makes the point at intake. Unstructured demand can be captured and normalized into the workflow instead of becoming another copy-and-paste job. ELIA can then work from connected customer, part, inventory, sourcing, pricing, quality and transaction context. The purpose isn’t to give users another place to chat. It’s to put interpretation and decision support close to the transaction that already knows the work.
The responsibilities become clearer:
- ERP.Aero maintains objects, relationships, permissions, known rules, transaction state and deterministic execution.
- ELIA interprets intent, evaluates uncertainty, supports planning, identifies exceptions and recommends permitted next actions.
- People retain policy, approval, technical judgment and material commitments beyond configured authority.
That is how agents become builders instead of glue.
Another customer captured both the opportunity and the honest fear of changing the foundation:
“You’ve got to go from four or five systems to one, and that’s kind of scary. But besides that, you can integrate everything.”
Connectedness has implementation work. Fragmentation has permanent work. We should compare both honestly.
What should you ask before approving the next AI project?
I’d start with five questions.
What part of this workflow truly requires interpretation or judgment? What work repeats predictably enough to become a rule? Where does the authoritative transaction live? How much AI effort is being spent reconstructing context? If the agent disappeared tomorrow, would the process still make sense?
Those questions change the design. They also change the economics.
Less general-purpose AI means lower inference cost, fewer unpredictable paths and a smaller surface to secure, test and audit. It lets teams use stronger models where the value justifies them and plain software where plain software does the job better.
This applies to product development too. An agent can help explore the problem, draft the implementation, test alternatives and expose patterns. That doesn’t mean every production step should remain agentic. Mature systems convert learned behavior into reliable capability.
I’ve spent close to two decades around aviation operations, and the best operators already think this way. They use judgment where judgment matters. They standardize what can be standardized. They don’t ask a technician to rediscover the same repair process every morning or a salesperson to rebuild the same customer history before every quote.
Our AI systems should meet the same standard.
Use AI to reveal the process. Use the process to reduce unnecessary AI. Then let intelligence work on the parts that still deserve intelligence.
About me and ERP.Aero
I didn't start here but I landed in A&D and aviation software. After nearly two decades working around aviation parts, sales, sourcing, distribution, repairs and the systems behind them, he focuses on removing the work that keeps good people from using their judgment.
ERP.Aero is an aviation native operational platform connecting quoting, sourcing, inventory, certification, repair, fulfillment, finance and embedded intelligence. iRFQ helps bring inbound demand into the workflow, while ELIA is designed to apply intelligence within the transaction rather than act as permanent glue between fragments.
