Five questions IT leaders are already asking about agents

TL;DR. On 19 March 2026, Aaron Levie (Box) published what he heard from 20+ enterprise AI and IT leaders, including people in regulated businesses. The useful public part is five questions: where the data lives, what the agent may do, whether it inherits a person's whole view, how tokens are budgeted by team, and whether one platform is supposed to cover the stack. This page is his note, retold as a buyer checklist. A score of our build stays off it.

A dinner note, written down

He described meetings and a dinner, then a short list of what kept coming up. Agents were the topic. Chatbots were the old topic. Coding was still the agent use those leaders actually saw in production; other knowledge work was starting to show up, and some organisations already had live uses. Cost cutting was barely in the conversation. Better work for customers was.

That scene is his. We did not host the dinner, and we are not turning his coding line into a pitch. The checklist below is the part a buyer can carry into a meeting this month.

1. Is the data somewhere an agent can use?

People have spent years with documents in more than one system. A colleague can still walk down the hall, remember which folder matters, and make a judgement. An agent cannot do that walk.

Levie's note puts the hard task in one sentence: getting content into a place an agent can work from, securely and without a treasure hunt. Years of "we will tidy that later" become the blocker the day something other than a person has to operate on the files.

The question for your own estate is plain. Is the material this job needs already in a place the agent can use, with a source a person can open? Or is it still scattered across drives that only the old hands can navigate?

2. What is the agent allowed to do?

The same note treats a second problem as still open: governing what an agent may do with that data inside a workflow. A chat that answers a question is one thing. A system that files, sends, or changes a record is another.

Ask it before the pilot grows up. Which actions are in bounds for this job? Which actions need a person to say yes? What gets written down when it acts?

A model that replies is not yet that system. We have written the difference separately, in a model alone is not an agent. When the agent can use tools, the fence around those tools is its own topic: why AI agents need sandboxes. This page does not re-argue either one. It only marks the question the leaders were already asking.

3. Does it see everything the person sees?

Identity showed up next to governance. His wording is the uncomfortable version: can the agent have access to everything you have? In a workplace with many agents working "on behalf of" someone, that is a lot of exposure for a task that needed a slice.

The picture is a badge. A person's pass opens the rooms their job needs. You do not photocopy that pass onto a helper that runs overnight and then wonder which doors it tried. Leaders in his note were asking for a smaller grant: access for the task, kept apart from the person's whole view.

How a product cuts that grant is a design. It stays with engineering. This page keeps the question, and leaves the mechanism off the site.

4. Who budgets the tokens, by team?

Tokens are the meter on the bill. His note says companies want room to try uses, and they still need a ceiling. The questions forming in those rooms were about budgets across teams and uses, not a single IT line that hopes for the best.

He expects that spend to sit with operating cost over time, beside the other costs of getting work done, rather than hiding inside a technology budget forever.

A trail of recent calls can tell you who hit which route, and some usage metadata, without opening the prompt. That trail is not a budget. A budget is a decision: which team may spend how much, on which use, before the quarter surprises finance. His note says that decision is arriving. Bring it to the meeting as a question. Do not treat a log viewer as the answer.

5. Are you betting the stack on one platform?

The last point in the note is interoperability. The enterprises he sat with were already deploying more than one AI system. He does not expect one platform to cover them all. Buyers, he wrote, are getting sharper about how agents work across that mix, and he expects that mix to shape the stack.

Take his sentence as his. The buyer question underneath it is enough: which jobs is this system for, and which systems will still sit beside it when the pilot ends? A purchase that assumes one logo will absorb the rest is a different bet from a purchase that names the neighbours.

What this page is for

Five questions, in the order a busy IT lead can actually ask them:

1. Is the data somewhere an agent can use?
2. What is the agent allowed to do?
3. Does it see everything the person sees?
4. Who budgets the tokens, by team?
5. Are you betting the stack on one platform?

They come from one public note, dated 19 March 2026. The "20+" is his count of the room. If a vendor answers all five with a single adjective — enterprise, governed, production — ask for the sentence under the adjective.

We are not publishing a pass or fail against the five. The comparison of this note to our build stays with engineering.

One related answer, already written

One neighbouring question is already on the site: who asked the model. That piece is the visitor book: who called which route, and whether it succeeded, without a dump of the prompt. It is recent operator history. It is not a forever archive, and it is not a token budget.

It sits beside question 4 as a trail, and beside question 3 as a reminder that "who used the front door" is a different sentence from "what the agent was allowed to see." Read it for that trail. Leave the five questions above as the checklist.

If you are the person who will sit in the room when someone says "agents in production," bring these five. Talk to us or try the demo with the questions, not with a request for a scorecard.

Frequently asked questions

Is this AiHPC's own survey of IT leaders? It is Aaron Levie's public note of 19 March 2026. The room size is his. We were not there.

Does OrchAI pass these five questions? This page does not score that. The checklist is for the buyer. An internal comparison stays off the site.

Is a request log the same as a token budget? A request trail shows recent calls and usage metadata, without the prompt. A token budget is a spending decision by team and use. The log is the trail. The budget is the decision his note says is coming.

Should we wait until one platform does everything? His note says those leaders are already running several systems. Ask which jobs this one is for, and which systems remain next to it.