Most of the AI deployment conversation is about who gets paid to do it. Aaron Levie’s post on September 30 was more useful than that, because he wrote down what “it” actually is.

His list, lightly trimmed. Legacy systems moved to the cloud, data organization and access updated, software connected to agents in new ways, workflows reengineered for agents, human-in-the-loop figured out per process, evals generated and maintained, and the whole system updated as new models ship. Then the honest part. “And the full list may even be longer.”

I read it twice and did one thing with it. I sorted the seven items by who has to own them. Four you can buy. Three you can’t, and those three decide whether the other four are worth paying for.

Why the old implementation model breaks

Levie’s distinction is the one I’d underline. Software implementation used to mean installing a known category of tool, stepping back, and letting the customer run it. With agents, “you’re deploying work output in a process.”

An ERP never decided anything on its own. An agent that reconciles invoices does, every time it runs. Somebody has to own what happens when it’s wrong, and that somebody is rarely the firm that set it up and moved on to the next client.

The money is already moving

Petar Petrov tallied it in his September 30 newsletter. Between April and early July, OpenAI, Anthropic, Microsoft, Amazon and Google committed around $9.75 billion to teams that go inside a company and stay until the AI is used. I covered what that meant for Accenture back in July.

The number in his piece that matters more for a mid-size company comes from McKinsey, as he cites it. Among companies using AI, about 20% have redesigned workflows around it. Among the ones getting real financial results, it’s 55%.

Workflow redesign is item four on Levie’s list. It’s also the line between companies that see returns and companies that see invoices.

Seven items, two owners

Workstream	Can you hire it out?	What breaks if you skip it
Legacy systems to cloud	Yes	The agent has nothing to plug into
Data organization and access	Only the plumbing	The agent sees too much, or too little
Software connected to agents	Yes	Brittle integrations that fail on every update
Workflows reengineered	Partly	AI bolted onto the old process
Human in the loop	No	Nobody knows who signs off before the agent acts
Evals	No	You can’t tell if it works, or if it got worse
Updates with each model	Yes, by contract	The system ages in months

The three “no” rows share something. They’re decisions, and a vendor can only implement decisions somebody already made.

A partner can configure permissions on your payroll data. They can’t tell you who should have them. They can wire an approval step before the agent emails a client. They can’t tell you which emails need one. And evals, the set of real cases with known correct answers, can only be written by whoever knows the answers. That’s your operations lead who has done the work by hand for ten years, not an engineer who met your company last Tuesday.

Levie said it in seven words five days earlier, “You can’t automate what you can’t measure.” I’ve argued before that the eval is the part you keep when prompts and models change. Partners rotate off. Your evals stay.

What we do at IQ Source before anyone deploys

If you hire a deployment team before settling those three decisions, you’re paying them to make the decisions for you, using whatever they learned in a week of interviews.

That’s the gap AI Maestro is built for. It sits before deployment rather than replacing it, and it typically runs two months with deliverables every two weeks. Four come out of it. A Process Reality Map of how the company actually works (it rarely matches the org chart), a definitions audit that finds where the same word means two things in two departments, an AI Opportunity Score, and a ranked list of opportunities that ends in a Go/No-Go gate. It runs on whatever AI tools the company has already approved, so nothing new gets bought to start.

What comes out is the half of Levie’s list you can’t contract. Which data, who approves, how it’s measured. With that settled, the other half gets quoted for less, because the implementer stops guessing.

Levie closed by saying it’s a great time to be a forward deployed engineer. Agreed. It’s also a good time to walk into that first meeting with your list already sorted.

Sort your deployment list with AI Maestro first