> ## Documentation Index
> Fetch the complete documentation index at: https://docs.surfaice.pro/llms.txt
> Use this file to discover all available pages before exploring further.

# Agent limitations

> What Hugo — the Surfaice agent — remembers, what it cannot do, and how to work with limits honestly.

**Hugo** is the AI agent inside Surfaice. This page describes its limits so your team can set realistic expectations before relying on outputs in systems of record.

Surfaice is built for high-stakes retail real estate work. That means being clear about what Hugo **can** do in your workspace today — and what still requires a human or your system of record.

## Memory and context

| What Hugo remembers                                          | What it does not guarantee                                         |
| ------------------------------------------------------------ | ------------------------------------------------------------------ |
| The current chat thread and workspace instructions           | Unlimited perfect recall across every past conversation            |
| Artifacts you deliberately save in the workspace             | A full mirror of Lucernex, SharePoint, or your mailbox             |
| Connected-system data at query time (under your permissions) | Data from systems you have not connected or are not allowed to use |

Long threads may be summarized so Hugo can keep working. If a detail mattered three weeks ago and was never saved to a workspace artifact, ask Hugo to look it up again in the source system.

## Skills vs freeform chat

| Use a **skill** when…                                               | Use **freeform chat** when…                    |
| ------------------------------------------------------------------- | ---------------------------------------------- |
| You need a repeatable workflow (abstraction, CAM triage, key dates) | You are exploring or asking a one-off question |
| Inputs are structured (store #, PDF, date range)                    | You need a quick summary across systems        |
| Output should match a known template or App                         | You are drafting email or narrative text       |

Skills route through tested procedures. Freeform chat is more flexible but easier to mis-scope — always specify store number, landlord, project, or folder when you can.

## Portfolio and corpus questions

* **"All leases" or "every store"** requests can take a long time and may return partial coverage when the corpus is large
* Hugo should report **how many** stores or files it inspected. If you do not see a count, ask: *"How many leases did you include, and which folders?"*
* Maps, dashboards, and **Preview** artifacts depend on the skill finishing successfully. If Preview is empty, retry with a smaller scope — see [Working with Hugo](/guides/working-with-hugo)

## Scanned documents and OCR

Many retail tenants work from scanned lease PDFs. Hugo can transcribe and abstract scans, but quality varies:

* Low-resolution scans, handwritten notes, or stamp overlays may produce missing or wrong fields
* OCR transcription and full lease abstraction are different workflows — say which you need
* Always compare abstracted values to the signed PDF before updating Lucernex

## Punch walk and field workflows

Construction punch walks (where enabled) have product-specific behavior:

* **Compliance checklists** work best as on-screen forms — not three questions in one paragraph
* If you chose **inspect sections in order**, Hugo should advance automatically; if it keeps asking permission to open the next section, say *"continue in order"* or restart with that mode selected
* **Verify the section** before confirming a deficiency — voice or freeform descriptions can land in the wrong trade bucket (for example GC vs installer)
* **Contact tables** on the start card should match your system of record (for example Sitefolio, a construction project tracker). If they disagree, treat Hugo's table as a draft and correct before the walk

## Response time

Simple chat answers are usually fast. Heavy workflows — full-folder inventory, multi-store extraction, voice punch walk, or large portfolio maps — can take **30 seconds to several minutes**. That is expected when Hugo is reading many files or running structured skills.

For complex jobs, see [Working with Hugo](/guides/working-with-hugo) or contact [support](/support).

## What Hugo will not do

* Expand your permissions to data you cannot already access
* Replace sign-off in Lucernex, your CMMS, or email send without your explicit action
* Provide legal advice — Hugo can summarize lease language; your counsel decides strategy
* Guarantee 100% accuracy on first output — **treat first outputs as drafts**

## What you should always verify

1. **Citations** — does the cited page or system record support the claim?
2. **Store and landlord identity** — ambiguous store numbers cause wrong answers
3. **Dates and dollars** — escalation tables and CAM pools are high-risk fields
4. **Coverage** — for portfolio questions, confirm how much of the corpus was searched
5. **Saved artifacts** — if you need to find work later, save or export the accepted result

## Related

* [Chat with Hugo](/guides/chat-with-hugo)
* [Working with Hugo](/guides/working-with-hugo)
* [Trust FAQ](/concepts/trust-faq)
* [Permissions](/concepts/permissions)