The Highest-Odds AI Pilot Is a Translation Job

Most AI pilots fail. We’ve written about the MIT number before — 95% of enterprise pilots show zero P&L impact — and the pattern behind the failures is almost always design, not technology.

But there’s a flip side to that research nobody talks about: the pilots that do work cluster around one job. Taking messy human input — speech, emails, call notes, free-text tickets — and converting it into clean, structured data your systems can act on.

Unstructured in. Structured out. That’s the translation job, and it’s the highest-odds first AI project a mid-market company can run.

Why this one category keeps working

Gartner estimates 80-90% of new enterprise data is unstructured, and it’s growing roughly three times faster than the structured kind. Your CRM, your ERP, your ticketing system — they all run on structured fields. The gap between what your business generates and what your systems can use is the single biggest pile of unrealized value in most companies. Humans have been bridging that gap manually: re-typing call notes, reading emails to figure out what category they belong in, listening to voicemails and creating tickets.

That bridging work is exactly what large language models are unreasonably good at. Not hype — shape. The problem itself is built for this technology:

The output is constrained. You’re not asking the AI to be creative or strategic. You’re asking it to map open-ended input against a known list — your service categories, your product catalog, your priority levels, your CRM fields. Closed vocabulary, finite schema. The model can’t wander far, and when it does, it’s obvious.

The baseline already exists. Somebody is doing this conversion today, by hand, and you can measure them: minutes per ticket, cost per intake, error rate per hundred records. That’s your “before” number — the thing the failed 95% never capture — handed to you for free.

The review loop is natural. Structured output is checkable at a glance. A dispatcher can verify a categorized service call in five seconds. Compare that to reviewing an AI-written proposal, where checking the output takes nearly as long as doing the work. The economics of human-in-the-loop actually work here.

Failure is graceful. When the model isn’t confident, it routes to a human. That’s not a workaround — it’s the design. The pilot doesn’t have to be perfect to be valuable; it has to handle the easy 70-80% and know when to hand off.

McDonald’s is running this exact play — with a known list of cheeseburgers

In early June 2026, at its Worldwide Convention in Las Vegas, McDonald’s announced ArchIQ — an AI operating platform built on its multi-year Google Cloud partnership — including a generative AI drive-thru voice assistant nicknamed “Archy.” It’s piloting at five US locations, US-first by design, with broader deployment planned through 2026 and nationwide reach projected for 2027.

Strip away the headlines and look at what the drive-thru actually is: a torrent of unstructured audio — accents, idle chatter, kids yelling in the back seat, “uhh, gimme the thing with the...” — that has to be resolved against a completely known menu. Every utterance maps to a finite list of items, sizes, and modifiers, and the output is a structured order ticket. It is the purest unstructured-to-structured conversion problem in commercial existence, run at colossal volume.

The reported early numbers (company figures, not yet independently verified): over a million transactions processed, roughly 90% completed without human intervention, in both English and Spanish. The 10% that fall outside the model’s confidence get handed to a crew member — graceful failure, by design.

Two details matter more than the tech. First, this is McDonald’s second attempt — it wound down an IBM-partnered version in 2024 after a 100-plus-location test produced enough order errors to go viral. They didn’t abandon the category; they fixed the design and re-piloted. Second, the rollout is deliberately staged: five stores, one country, expand on evidence. A company with the resources to deploy everywhere on day one is choosing pilot discipline instead. If the scoped, measured, staged approach is good enough for a brand serving tens of millions of customers a day, it’s good enough for a 100-person company in Alberta.

What this looks like at mid-market scale

You don’t have a drive-thru. But your team still deals with unstructured data:

Service call intake. A customer calls, describes a problem in their own words, and someone transcribes it into a work order — equipment type, issue category, urgency, location. An LLM does the first pass against your known category list; a dispatcher confirms. We’ve seen this cut intake time from minutes to seconds per call, and the structured data it produces makes your scheduling and parts forecasting better as a side effect.

Ticket triage. Every new support ticket arrives as free text and has to be categorized, prioritized, and routed. This is the most documented win in the category: one production IT support deployment reported saving 300-plus staff hours and roughly $15K per month across 2,000+ monthly tickets, and contact-center implementations have reported double-digit reductions in handle time. The model reads the ticket, assigns category and priority from your taxonomy, and routes it — with low-confidence tickets flagged for a human.

Email and document intake. Invoices, POs, claims, vendor correspondence — extracted into the fields your ERP or accounting system needs, with a human approving exceptions. The known list here is your chart of accounts, your vendor master, your claim types.

Field reports and meeting notes. Technicians and reps capture what happened in fragments and voice memos. The model converts them into structured CRM updates and job records, so the data actually lands in the system instead of dying in a notebook.

Notice what every one of these has in common with the drive-thru: a known list on the output side, a measurable manual process on the input side, and a human who can verify the result faster than they could produce it.

How to pick yours

Run the same diagnostic we use with clients. Find a workflow where someone on your team currently reads, listens to, or re-types unstructured input into a system — then ask four questions:

• Is the output a known, finite list (categories, fields, codes)?
• Can you measure the manual version today, in minutes and dollars?
• Can a human verify the AI’s output faster than doing the job themselves?
• Is there a clean handoff path when the model isn’t confident?

Four yeses and you have a pilot with real odds — the kind that produces a P&L number at day ninety instead of a shrug. Then run it like the 5% do: named workflow, written baseline, designed review loop, one accountable operator.

McDonald’s needed two attempts and a staged five-store pilot to get this right. You get to skip their first mistake and start with the discipline. The translation job is sitting in your intake queue right now — it’s been there all along, waiting for the technology to catch up.

It finally has.