Data Agent - Smart properties

[ISmolin]

Hello there. I spent about 5 whole days just playing with prompts and doing tests for 47 companies - sample, before scaling. Everything worked fine, was checked by the team’s experts, and the data was available. Now, when I start using it in a commercial, either the data is unknown or missing altogether. Credits are charged. Great business

The prompt worked fine is below:

Research the typical pricing positioning of this company within its industry based on publicly available information (such as product catalogues, pricing pages, customer segments served, and brand positioning) and determine whether it primarily targets a higher-priced, average-priced, or budget-focused market segment.

Research this company’s website and product categories and analyze price point and range. Categorize the brand overall as: Low-end, Mid-end, or High-end/Luxury Pick thresholds based on the flagship category:

If the flagship category is any of (Sofas, recliners, coffee tables, TV stands, sofas, sectionals, Beds, dressers, nightstands, wardrobes. Dining sets, bar stools, buffets, Desks, office chairs, shelving, Patio sofas, dining sets, loungers, Desks, cubicles, task chairs, conference tables, Guest room beds, lobby seating, banquet chairs, Waiting room chairs, hospital beds, institutional tables, Cabinets, bookcases, shelving units, Cabinets, bookcases, shelving units, Console tables, benches, room dividers, Expandable tables, modular sofas, storage beds, Wood, metal, plastic, upholstery):

- Low-end: average below $800
- Mid-end: average $800–$2,500
- High-end/Luxury: average above $2,500

Return ONLY a comma-separated list using exactly these labels:

“High End / Luxury”, “Mid-end”, “Low-end”.

If you can’t confidently decide, return exactly “Unknown”.

Do not include any additional text in your answer.

Now it returns no data.

Please advise how to acieve this goal as it’s frustrating toi waste again 5 hours playing with prompt and getting inconssitent data outcome.

[STierney]

Hey @ISmolin - thanks for posting in the Community!
I understand that the prompt you were using for Companies is not working the same outside of Companies. Is that correct? I ask, because your prompt seems to be designed specifically for the Companies object, which would make it difficult to translate this into other objects. My expectation is that the language would need to be updated in an effort to align it with whatever object(s) you’re looking to find this data for.
Shane, Senior Community Moderator

[ISmolin]

Hello, you got it wrong. It was created for companies and it’s used for companies. When I tested the same prompt creating it - you can choose just 3 companies to see the results - on this stage all was perfect and data was correct. Once I implemented it and started to enrich segments with smart properties - the same prompt return results not for all companies or its unknown. I mean when you scale - smart property fails. I as a customer should click 2-3 times to get results for companies - why??? And of course I’m getting charged for each click - fill smart properties. I see the product is raw guys and it’s just onvious

[ISmolin]

When here - all is good:

And when you use in a commercial for a larger number of companies, it fails (the same smart property and LLM), come on guys:

It is about inconsistent data outcomes as well. Credits are cheap - yes, but not when you apply them to 1-2K companies for several smart properties -the idea was and is to automate things and use HS as the single source of truth, but not hoping that maybe the data is correct let’s pray

[STierney]

Hey @ISmolin!
Thanks so much for providing this context for my own understanding. I’m wondering if @karstenkoehler or @danmoyle might have any helpful insight here.
Shane, Senior Community Moderator

[danmoyle]

You seem very frustrated @ISmolin. As a fellow HubSpot user, I’ll first say I’m sorry you’re so frustrated by the new Data Agent. I don’t know that I can fully solve this, but I want to try and help you and others who might see this. Thanks for the tag and kindness @STierney.

I wonder if the scaling is caused by scale limits? As I understnad it, HubSpot’s AI may impose rate limits or timeouts for bulk operations on 1-2K records, leading to incomplete web research. Or maybe data variability as you scale? Some companies lack accessible pricing data (e.g., no public catalogs), which is expected but amplified in bulk. I wodner if going from 47 companies to maybe 75 would help surface limits or data variability? If it were me, I’d try to test on increasing segments rather than all at once.

Another thought, reading your prompt, is to ask an AI to help imrpove the prompt. I’m thinking the long, specific prompt with category lists and theresholds might be improved so as not to overhwlem the model as you scale. I’ve seen some folks out there I admire talk about doing this with prompt engineering. I asked a couple of AI engines, and here’s one I’d try:

Analyze this company’s website and public pricing info for furniture products (sofas, beds, desks, etc.). Categorize as:
- High End/Luxury: average >$2500
- Mid-end: $800-$2500
- Low-end: <$800
- Unknown: if unclear
Return only: “High End / Luxury”, “Mid-end”, “Low-end”, or “Unknown”.

So I’d try batch processing as one workaround (Enrich via lists or views in smaller groups, growing as you go, and use workflows for automation.)
​

Then I’d contact support and reference the thread. You might be able to request beta access to improved scaling (a guess as they keep improving these brand new tools, I don’t know anything behind the scenes about betas).

[Henrik-02]

The Data Agent is handy for quick insights but it still struggles with anything beyond basic calculations on smart properties.
If your smart property uses multiple conditions or rollups from associations, the agent often misinterprets or ignores parts of the logic. Best workaround right now is to create a separate calculated property that mirrors the smart one (using the same rules) and point the agent there instead. It reads those much more reliably.
Tested it on a few deals last week and the difference was night and day. Have you run into the same limitation or is it something else?

[MarkM1G]

HubSpot’s Data Agent absolutely chokes past ~75 records in our experience. We moved the classification task into an external pipeline that chunks and validates before writing back, this brought consistent results regardless of volume. Had to fill in a custom property for 15k+ accounts - took forever, but without a single issue

[Berangere]

Hey @MarkM1G and thanks so much for sharing your valuable feedback and workaround, we really appreciate it!

[WebGuruz_Tech]

Hey,

What you’re experiencing is actually quite common when moving from controlled prompt testing to real-world usage at scale. Let me break down what’s actually happening.

There are a few reasons this breaks at scale, and they’re worth knowing separately.

When you tested on your 47 companies, you were likely running the preview inside the Smart Property setup (a controlled, small-scale test). When you go commercial and bulk fill across a full segment, the environment changes, and that’s where things fall apart.

Rate limits. HubSpot’s AI has limits on how many records it can process in bulk at once. At 47 companies, there’s no pressure on the system. However, at a commercial scale, records start timing out and returning empty, even when the data exists publicly.

Dirty domain data. Smart Properties with “Web research” use the “Company Domain Name” field to know where to look. If that field is blank or incorrect on any records, the agent has nothing to search and returns nothing, but still charges credits.

The workflow “Custom Prompt” has no internet access. If you’re running this via a workflow, that version of the agent can’t browse the web at all. Web research only works through Smart Properties with the “Web research” data source selected.

Now, here are a few things you can do

Simplify your prompt first. Your current prompt is thorough, but a long prompt with large category lists and multiple conditions can overwhelm the model at scale, even if it worked perfectly in testing. Try something cleaner, like:

Analyze this company’s website and public pricing info for furniture products (sofas, beds, desks, etc.). Categorize as: High End/Luxury: average above $2,500; Mid-end: $800–$2,500; Low-end: below $800; Unknown: if unclear. Return only one of these: “High End / Luxury”, “Mid-end”, “Low-end”, or “Unknown”.

Shorter prompts tend to perform more consistently at scale.

Clean your domain data before running anything. Filter your view to only records where the company domain name is populated and correct. This alone will eliminate a significant chunk of empty results.

Batch your enrichment. Rather than bulk filling your entire segment at once, work through it in batches of 50-100 records. It’s slower, but it avoids the timeout issues, and your success rate will be noticeably higher.

Move to a workflow for ongoing enrichment. Use the “Data Agent: Fill Smart Property” action inside a workflow instead of manual bulk fills. You control the pace of enrollment, which makes the whole process far more reliable.

Hope this helps!