Back to website Submit a ticket My Tickets
Welcome
Login  Sign up
  1. Solution home
  2. Usage guides
  3. Ask Hortis : AI Connector

Ask Hortis FAQ

Havard Ostgaard
Modified on: Thu, 3 Sep, 2026 at 11:54 AM

Content


Getting started

Do I need to pay for an AI subscription to use Ask Hortis?

Ask Hortis works through an AI connector and there is a free plan available from Claude.ai that offers this capability. However, the free plan will have lower usage limits and only one connector slot, but it is a good starting point if you want to try it out. As far as we know, other AI platforms only allow for AI Connectors on paid plans. 

We don't use AI much yet. Where is a safe place to start?

Start with questions rather than changes. Asking "how many accessions do we hold in the Asian collection?", "which materials have no location recorded?" or "which accessions have notes mentioning frost damage?" costs you nothing if the answer is wrong, and it teaches you quickly how precise your prompts need to be.

When you move on to changes, three habits keep the quality-control burden small:

  1. Work in small batches you can check. One bed, one genus, or twenty records, not the whole collection.

  2. Ask it to show you before it writes. "List the records you would change and what you'd change them to" gives you a reviewable list first.

  3. Turn off the tools you don't want it to use. In Claude, the "Search and tools" menu lets you disable individual connector tools for a conversation, so you can run Ask Hortis effectively read-only while you build confidence.

Good low-risk first jobs: finding duplicates, finding blank fields, listing accessions using names no longer accepted by the shared taxonomy, and summarising notes.

Can I try it without being a Hortis subscriber?

Yes. You can sign up to a free trial of Hortis. You will start with an empty dataset, so you can export a subset from your current system, import it into the trial, and experiment on that. Keep in mind that importing large amounts of data via Ask Hortis will take some time and consume usage tokens. 

How much time will this actually save?

This is too early to say, but it is clear that repetitive tasks, deep research and complex cross referencing of collection data against external data can be done much faster with Ask Hortis. 

Control, permissions and governance

Can the AI features be turned off?

Not from inside Hortis today. Ask Hortis is opt-in in the sense that someone has to deliberately set up the connector in the AI tool and sign in with their Hortis credentials, which is a multi-step process that won't happen by accident. But there is currently no Hortis setting that prevents staff from doing it.

Two things worth knowing in the meantime:

  • Anyone using Ask Hortis is working with their own Hortis account and permissions. It cannot see or change anything they couldn't see or change by logging in normally.

  • If your organisation is on a Team or Enterprise plan, only an Owner can add a custom connector for the organisation, which gives you a control point on the Claude side. That doesn't stop someone from connecting to Hortis through a personal AI account.

An organisation-level switch is technically straightforward for us to add. If it matters to you, tell us, because that's how it gets prioritised.

Does the AI have more access than I do?

No. It acts with your credentials and your role. An administrator or editor can do the same things through Ask Hortis that they can do in the interface, and a user with read-only access stays read-only. Everything is written to the activity log in the usual way.

Will the activity log show that a change was made through AI?

Not yet. Because Ask Hortis acts on your behalf, changes currently appear in the activity log under your name, exactly as if you had made them yourself. We have received requests to show a visible distinction between "Richard changed this" and "Richard asked AI to change this", and we agree that this can be useful and we are looking into this. 

Can Ask Hortis change our public-facing website?

It can change what you publish, not how the public site works. For example, you can ask it to make all plants with a particular tag public. You can't ask it to add a feature to the public site, such as letting visitors search by tag. Site features are something we build, guided by what the community asks for, so send those requests to us.

Mistakes, undo and recovery

If AI makes a bulk change that's wrong, can I undo it?

Usually yes, if you catch it in the same conversation. We've tested this: you can say "that was a mistake, please revert those changes" and it will work back through what it did. Two things make this safer than it sounds:

  • Hortis keeps a full audit history. Changes are added on top of each other rather than overwriting what was there.

  • Deletions go to the archive, not to nowhere. If someone asked it to remove a large number of records, those records can be restored.

We haven't tested every scenario, so treat a large bulk change as something to check rather than something to trust.

What if someone else made the bad bulk change?

That's harder today. If it was hundreds of records and you weren't part of the conversation, correcting it currently means working through the records manually or asking AI to reverse the change from a fresh conversation, which needs you to describe accurately what was done. The activity log will show what changed and when.

The underlying history is all retained, so a general "revert these changes" function is technically possible for us to build, but it isn't in Hortis today. 

Accuracy and trust

Will Ask Hortis admit when it doesn't know, or will it make something up?

We can't promise it will never be wrong. In practice, we've found it's reliable about its own limits: if you ask for something Ask Hortis can't do, such as bulk-adding photos, it tells you it can't rather than pretending it has. Where more care is needed is analytical answers, such as counts and summaries, where a plausible-looking number is easy to produce and easy to miss.

Practical advice: ask for the working ("which records did you count?"), spot-check anything you'd act on, and tell us when you catch something wrong. We monitor how the tools are being used and can fix problems quickly, but only if we know about them.

AI relies on what's on the internet, and the internet is uneven. How do we control the quality of its sources?

This was a fair challenge and there are two things you can do about it. First, tell it which sources to use, for example POWO, IPNI or a named regional flora, rather than leaving the choice open. Second, tell it to show which sources it used for each statement, so you can verify rather than trust. Both are just instructions in your prompt, and using them together turns an unverifiable answer into a checkable one.

What limits or instructions should we set up from the start?

The most useful ones we've seen so far:

  • Name the authorities it should use for nomenclature and geography.

  • Ask it to cite sources for anything it fetches from outside Hortis.

  • Tell it to ask before writing, and to report what it changed afterwards.

  • Tell it to say when it doesn't know instead of estimating.

Beyond that, the honest answer is that curators and horticulturists using this daily will work out better conventions than we can guess at. We'll expand this section once we get more feedback from the community. 

Data privacy, ownership and ethics

This section is about AI tools generally rather than any one product. Ask Hortis has been tested on Claude and ChatGPT and both appear below as worked examples. Every provider sets its own terms and those terms change, so treat the examples as an illustration of the questions to ask rather than a substitute for reading your provider's current policy.

Will our collection data be used to train AI models?

It depends on three things, and most of the confusion around this question comes from mixing them up.

  1. How the data reaches the model. Records read live through a connector, meaning an integration such as Ask Hortis, are not necessarily treated the same as text you type, paste or upload into the chat. Some providers exclude connector content from training and some don't.

  2. Which plan you are on. Business and enterprise tiers generally exclude customer content from training by default. Consumer tiers usually depend on a setting, and that includes paid personal plans, so "we pay for it" is not an answer on its own.

  3. Your settings and habits. A model-improvement toggle, a private or incognito mode, and the thumbs up/down feedback button all change the answer. Submitting feedback can put an entire conversation back in scope even if you have opted out of training.

How Claude handles it. Anthropic draws the line by content type. Data used to improve their models excludes raw content from connectors, including remote MCP servers, which is what Ask Hortis is, though content is included if you copy it directly into the conversation. Beyond that, commercial plans (Claude for Work, meaning Team and Enterprise, and the API) are not used for training by default, while consumer plans (Free, Pro and Max) depend on a model-improvement setting you control. Incognito chats are never used for model improvement. The thumbs up/down button stores the related conversation for up to five years, again excluding raw connector content.

How ChatGPT handles it. OpenAI draws the line by plan instead. For Business, Enterprise and Edu customers, information accessed from connectors is not used to train models by default. For Free, Plus, Go and Pro users, information accessed from apps may be used for training if the "Improve the model for everyone" setting is on. There is no connector exemption on the consumer tiers, so the setting and the plan are what protect you.

The practical difference: on a consumer plan, the same connector query is outside training data on Claude but potentially inside it on ChatGPT with the default setting left alone.

Other platforms will differ again. Google, Microsoft, Mistral and the rest each set their own defaults, and some tools are thin wrappers that pass your data to a model provider under terms of their own. If you run an open-weight model on your own hardware, nothing leaves your infrastructure, which is the approach institutions with strict security requirements tend to take.

Assume the terms will change. Anthropic changed its consumer terms in 2025 to introduce the model-improvement choice described above. OpenAI's help pages are revised regularly, sometimes within days. Whatever we write here can go out of date, so check your provider's current documentation before relying on it, and re-check when you change plans or when a provider emails you about updated terms.

What should we check before connecting any AI tool to our collection?

Six questions that will get you a clear picture of any provider:

  • Does content read through connectors count as training data, or only what is typed and pasted into the chat?

  • Does our plan exclude training by default, or is there a setting someone has to find and switch off?

  • Is there a private or incognito mode, and is it exempt from everything or only from training?

  • How long is conversation data retained, what happens when we delete a chat, and does submitting feedback change that?

  • Who at our institution is able to connect the tool, and with which permissions?

  • Do any of these answers change for our region?

Who owns our collection data?

You do. Connecting an AI tool to Hortis does not transfer ownership of anything, and the major providers' terms do not claim your content as theirs. Two things are worth separating from ownership, though, because they are where the real questions sit.

Access and retention are not the same as ownership. Your data can remain entirely yours while still being stored on someone else's servers for a defined period. Ask how long, and what deletion actually does.

Ownership does not remove the need for trust. When you connect any cloud service you are relying on the supplier to do what its terms say. That is the same judgement you already make about your email provider and your finance system, and it is worth making deliberately rather than by default. For organisations that would rather not extend that trust, the alternative is running models in their own environment so data never leaves. It is also one reason we are looking at what AI capability should eventually sit inside Hortis itself.

We hold sensitive records, such as localities of rare wild-collected material. Should we expose those?

The risk is not abstract: precise collection localities for rare species, orchids being the obvious case, have real value to poachers. Sensitivity usually attaches to particular fields rather than to whole records, which is helpful, because it means you can be careful in a targeted way.

This applies whichever AI tool you use:

  • Don't paste sensitive localities or coordinates into a chat. Where a provider does exempt connector content from training, that exemption typically does not cover text you type or paste.

  • Keep restricted fields restricted, and unpublished. An AI assistant acts with your permissions, so what you restrict in Hortis stays restricted.

  • For sensitive work, use the strictest mode available to you. That means a business or enterprise tier, a private or incognito conversation, or a model running in your own environment.

  • Compliance obligations don't change because the tool is AI. If a record is sensitive under your own collections policy, it is sensitive in a chat window too.

Is sharing data for training actually a bad thing? Couldn't it help the community?

A fair question, and one we did not want to leave one-sided. The instinctive reaction is "it's our data, so no", but the counter-argument is real: models that have seen good botanical data give better botanical answers, and every garden using them benefits. It resembles the argument people had about Wikipedia in its early years, where the more productive response turned out to be improving it rather than dismissing it.

The distinction worth drawing is between the bulk of collection data, where the case against sharing is mostly one of principle, and the narrow set of genuinely sensitive fields, where the case against is concrete and specific. Deciding those two separately gets you further than one blanket answer either way.

How much energy does using AI like this consume?

Less than people often assume for this kind of task, though the honest framing matters. Generating images and video is very energy-intensive. Short text exchanges, which is what asking questions about your collection amounts to, are a tiny fraction of that per query. Training a large model is a very large one-off cost, and across the industry the total energy used serving queries is now substantial as well, because of sheer volume rather than any single question being expensive.

So the environmental question is real at industry scale, but asking Hortis to find your duplicate records is not a meaningful part of it. 

Where to check the current terms

These change, so go to the source rather than to us:

What it can and can't do today

Can it help clean up messy plant data, such as duplicate records?

Yes, and this is one of the strongest current uses. You can ask it to analyse the collection for duplicates and inconsistencies. Because these models handle language well, they're also good at things a database query struggles with, such as reading free-text notes and finding the records where the note contradicts the structured fields.

Can it fill gaps in records we imported from incomplete spreadsheets?

Yes, and this is a good fit. A worked example: if you have wild-collected material with coordinates but no locality description, you can ask it to derive the region from the coordinates and fill that in. Do it in batches, review a sample before accepting the rest, and keep in mind that derived data should be recorded as derived.

Can it fix incorrect or outdated botanical names?

It can help you find them, which is the hard part. We maintain the shared taxonomy in Hortis and update it regularly, so accepted names stay current on our side. What AI adds is the ability to work in the other direction: ask it to find every accession using a name the shared taxonomy no longer accepts, then review and update. It won't watch for changes and notify you unprompted.

Can it import plant lists or run bulk actions from a CSV?

You can use it to work through lists and apply bulk changes, and this is where the time savings are largest and the care needed is greatest. The download-edit-reupload round trip that some gardens use for bulk notes work can often be shortened. Treat a bulk write as a change you will check, and read the undo answer above first.

Can I photograph a garden bed, or use drone imagery, and have AI identify the plants and upload them to Hortis?

Not today. Ask Hortis does no image processing at present, so this can't be done end to end.

On the underlying idea: a photo of a whole bed or drone footage is a hard case, because the resolution and the overlapping foliage make reliable species-level identification unlikely. Large individual specimens are more plausible. A photo of a single plant taken while walking the garden is a much better prospect.

There's a workflow that works now and sets you up for later. As you map, record the coordinates and a photograph and accession the plant as unknown. A point plus an image is a lot of information, so identification can be done retrospectively, in bulk, once tooling supports it. Image handling is a natural next step for the tool, and mapping speed is exactly the kind of bottleneck we want to attack. 

Can AI search the photos on my phone to identify a plant?

Not through Hortis. General-purpose plant identification apps do this, and some phone galleries have their own search. What we're interested in is the step after identification, which is getting the result into your collection records reliably.

Can it help with native flora data or tracking microclimate adjustments across a large landscape?

Partly, and it depends entirely on whether the data exists somewhere it can reach. If the native flora dataset or the microclimate records are accessible, cross-referencing them with your collection is the kind of work these tools do well. If the data isn't available in a usable form, AI can't conjure it. The blocker is usually data availability rather than capability.

Can it do historical analysis of our collection?

Hortis is good at showing you the collection as it is today, and the historical record is all there, but longitudinal analysis has not been easy to get at. AI makes that considerably more accessible, so questions about trends over decades, acquisition patterns, or survival by provenance become reasonable to ask. We also want to surface more of this in the software itself.

Could we connect our own archives, publications or garden history to this?

Yes, with a caveat about how. If that material is brought into Hortis as a dataset, then Ask Hortis can query it alongside your collection records, which is a genuinely interesting research prospect for material that isn't findable through a normal web search. The route is getting the content into Hortis rather than training a separate private model. 

Feedback and support

How do we report a problem or ask for a feature?

Tell us. We also monitor how these tools are being used and can see when things go wrong, which lets us fix issues quickly, but a message from you is faster and gives us the context.

Several answers above depend on community feedback, particularly an organisation-level off switch, AI attribution in the activity log, and a revert function. Requests are how those get prioritised.


H
Havard is the author of this solution article.

Did you find it helpful? Yes No

Send feedback
Sorry we couldn't be helpful. Help us improve this article with your feedback.