Quick disclosure: I work at Box as a Scaled Customer Success Manager, and a big part of my job is helping customers get value out of Box AI. Obvious bias. I'm going to argue against it in places.
Here's the pattern I keep seeing.
Someone opens AI Home for the first time and types in a filename. Or a client name. Or "Q3 pricing deck." They get back a conversational response instead of the file they wanted, quietly go back to the search bar, and never open it again.
That's a framing problem, and I think we created it — Box people, admins, champions — by talking about AI Home as an upgrade to search.
It isn't. It's a different tool that happens to search as one of its steps.
The question that resolves it: do I need a file, or do I need an answer?
Search is retrieval. It returns a set of things you can sort, filter, and act on. AI Home is reasoning. It reads across your content and produces a response.
Sort the task into one of those buckets and the choice stops being much of a debate.
First, the AI Units thing
This goes early because it's the objection I hear most, and it's usually a misunderstanding.
Box Agent in the web app doesn't consume AI Units. Not in Standard Mode, not in Pro Mode. The same applies to custom agents in Standard Mode and to conversational search.
AI Units are consumed by things like API calls, Box Automate workflows, Expanded Mode agents, and metadata extraction — automated, at-scale work. Not a person asking a question in AI Home.
So if your team has been rationing AI Home to protect the unit pool, you're rationing against the wrong meter.
There is a real constraint: a daily per-user Box Agent quota that resets at midnight local time. But that's your own headroom for the day, not your company's AI Units budget.
When search is the better tool
"Use AI for everything" is bad advice, and the people in this community know it.
- You know what you're looking for. Filename, exact phrase, something you touched last week. One keystroke, instant. Asking an agent to go find it is slower and less certain.
- You need an exact string. Invoice numbers, contract IDs, case numbers, employee IDs. You want a literal match, not an interpretation of what you probably meant.
- You need a defined result set. Search hands you the things that match your query. AI Home hands you an answer built from the content it determined was relevant.
"Every file that mentions this vendor" is a search job.
It's also why traditional search stays important when completeness and repeatability matter — audits, investigations, legal workflows, or anything where you need to review and act on the actual result set rather than an answer derived from it.
- You need real filtering. File type, size, date range, owner, folder scope, active vs. archive, title-only, metadata values, and Boolean operators. One thing worth knowing: AND, OR, and NOT have to be uppercase, or Box treats them as ordinary keywords. If your content is well governed, a precise metadata query can beat semantic guessing when precision is the goal.
None of that is a weakness of AI. It's just a different job.
When AI Home is the better tool
Notice that almost none of these are about finding things.
- You can't describe it, only recall it. "The pricing exception we made for a healthcare customer last year." You don't know the filename. You may not even know which document contains it.
- The answer lives across documents. You're picking up an account mid-cycle and need to know what you actually owe this customer. The commitments are spread across an MSA, two amendments, and an order form, and they don't fully agree.
Search gets you four files. AI Home can tell you where they conflict.
- You need patterns, not documents. Themes across a quarter of customer feedback. What's coming up more often in support tickets than it did last month.
There's no single file that contains that answer. It only exists once something reads across all of them.
- You want an output, not an input. A security questionnaire arrives with sixty questions. The answers exist across Trust docs, old RFP responses, and a policy nobody's opened in a year.
Search finds the sources. AI Home can draft the responses.
- You're new to the content. New hires, cross-department requests, anyone inheriting someone else's folder structure. Guessing at another team's naming conventions is a genuinely bad experience, and this is where AI Home becomes especially useful.
The workflow I actually recommend
Use search to define the scope. Use AI to reason over it.
Filter down to exactly the documents that matter — file type, date range, folder, metadata — then select those results and hand them to Box AI. You get search's precision on retrieval and AI's reasoning on analysis, with less guessing about scope in between.
That sequence resolves most of the tension in this post. The two tools aren't competing for the same job. One defines the set. The other thinks about it.
The biggest reason people think AI Home is "wrong"
A lot of the "it gave me a bad answer" stories I hear have something in common: the question was asked with no sources attached.
Unscoped is the default. The agent goes looking across the content you have permission to see, and sometimes it lands on a plausible source that's the wrong one — an old draft, another team's version, a template nobody deleted — and produces an answer from it.
Often the problem isn't that the tool is unreliable. It's that nobody told it where to look.
Scoping takes seconds. @ mention the two or three files that matter. Drag them in. Browse to the folder. The heavier version is the search-first workflow above, for when being wrong is expensive.
This is also the answer to the completeness limitation I raised earlier. Defining the source set yourself doesn't eliminate every limitation of AI, but it removes a major source of ambiguity: what content the answer should be based on.
Try these three prompts today
Reading an argument doesn't change how anyone works. Each of these only really makes sense in AI Home. Attach the relevant sources first.
- On a folder of agreements: "Compare the renewal and termination terms across these and flag anywhere they conflict or the dates don't line up."
- On customer feedback, tickets, or call notes: "What are the top themes, and which are getting more frequent over time?"
- On a long document you're about to be asked about: "What would someone challenge me on here? Give me the three weakest points and where they are."
That last one is my favorite, and I almost never see it used. It isn't retrieval. It isn't really summarization either. It's asking AI for a read on content you already have.
Hubs as AI context
Here's the shift I think is coming, and I don't hear many people talking about it yet.
We've mostly treated Hubs as a destination — a curated place for people to go and browse. But a Hub can also be a durable way to define context for AI.
Think about an HR Hub containing only current policies. A security Hub containing approved Trust documentation. A sales Hub containing the collateral reps should actually be using.
Instead of asking the agent to work out which content is authoritative every time, you've already curated the boundary. Scoping by @ mention solves the problem for one question. A Hub can solve it for a category of questions.
That's what makes this interesting to me. If a Hub is going to be used as AI context, you're no longer curating content only for humans to browse. You're curating what the agent should know.
And that may change how we think about building Hubs in the first place.
So, file or answer?
That's the framework I've started using with customers.
If you need a known thing, search for it. If you need an answer derived from your content, use AI Home. And if the answer really matters, define the source set first, then let AI reason over it.
Search retrieves. AI reasons. The best workflow often uses both.
I'm curious where this framework breaks for other people. When have you found standard search better than AI Home — or AI Home better than you expected?
I'd especially like to hear from admins who've invested in metadata structures. My instinct is that the better your governance, the more often standard search stays the right tool, but I don't know whether that holds up in practice.
And if people find this useful, I'll keep writing about what I'm seeing. There's more to unpack around Hubs as AI context, custom agents, and what good content structure looks like once AI becomes another consumer of your information.