AI for scientific instrument troubleshooting: a buyer's guide

AI for scientific instrument troubleshooting means software that answers "why is my instrument doing this, and how do I fix it?" from the instrument's own documentation. A good tool finds the right manual, finds the right page, and shows you that page so you can check the answer. A weak one writes a confident paragraph and leaves you to hope it's right.

This guide is for scientific instrument operators, the people who run shared instrument facilities, and field service engineers and leaders who are looking at these tools. We build one of them, so we have a point of view. We have tried to be fair about where other options are the better choice, and we say so plainly below.

Why instruments are a hard problem for AI

Instrument knowledge is spread across many documents. Take Thermo Fisher's Vanquish UHPLC, the front end of many LC-MS setups. The pump's manual points readers to separate operating manuals for the other modules, a system manual and the Chromeleon software help Reference [1]. The pump manual alone is 292 pages. Its troubleshooting chapter is built around a table of instrument messages, each with causes and fixes, and the manual has dozens of numbered figures Reference [1].

A shared facility can mix brands. Carnegie Mellon's Center for Molecular Analysis lists 11 instruments from six manufacturers Reference [2].

The manuals themselves say mistakes carry a cost. The Vanquish pump manual says its safety notices exist to prevent "personal injury, damage to the device, or loss of data" Reference [1]. So test accuracy first.

What the research says about general chatbots

General chatbots are fluent, but they make up sources. A study in Scientific Reports found that 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated Reference [3]. A 2025 study found 19.9% of GPT-4o citations were fabricated, and the rate was higher for less-studied topics Reference [4]. No one has measured this for instrument questions.

Grounding the AI in a curated set of documents helps Reference [5], but it does not make errors go away. Commercial legal research tools built this way still gave incorrect or poorly supported answers 17% to 33% of the time Reference [6]. The lesson for buyers: ask to see the source for every answer, and treat any vendor who claims zero errors with suspicion. That includes us.

We know of no peer-reviewed study of AI troubleshooting for analytical instruments like HPLC systems or mass spectrometers. The closest, a 2026 study of a chatbot for radiotherapy machine faults, found it did well at finding past repair events, but the authors said "important safety and workflow risks remain" Reference [7].

Your options today

1. Manufacturer tools and support portals

The large instrument makers offer software and online portals, not just a phone line. Four examples: Agilent's Lab Advisor runs diagnostic tests, calibration and maintenance routines on Agilent LC and CE systems Reference [8]. SCIEX Now lets customers search knowledge articles and submit and track support cases Reference [9]. For customers with a service agreement, Shimadzu's LabTotal Smart Service Net collects system check reports so Shimadzu can recommend fixes Reference [10]. Waters' help center lets you browse guides, videos and manuals by product or serial number Reference [11].

Where they are strongest: depth on their own instruments. Their tools can read the instrument's own records. Lab Advisor gathers error codes, firmware revisions and maintenance log entries Reference [12]. Shimadzu uses system check reports to recommend fixes and "prompt on-site repair" Reference [10]. SCIEX offers remote diagnostics and field service engineers Reference [9].

Where they fall short: their software stops at the brand boundary. Lab Advisor works on Agilent LC and CE systems Reference [8]. If your organization runs five brands, expect five portals and five logins. Some of it needs a service contract. Shimadzu's Smart Service Net, for example, requires a Shimadzu service agreement Reference [10].

Best for: organizations that run a single instrument brand, warranty work and any repair that needs a certified engineer.

2. General AI chatbots

ChatGPT, Claude and Gemini all have free versions Reference [13] Reference [14] Reference [15]. They can be a quick way to read up on a concept, like what a check valve does, as long as you check what they say.

Where they fall short: they only see your manual if you upload it, and they only know your model and firmware if you tell them. OpenAI's own researchers say language models "guess when uncertain" instead of admitting it Reference [16]. Check any error-code answer against the manual. As the studies above show, they also invent sources. Universities such as Harvard, UT Austin and UC Berkeley also restrict what research data can go into AI tools the university hasn't approved Reference [17] Reference [18] Reference [19].

Best for: background learning and general chemistry questions, with the answers checked elsewhere.

3. Your own shared drive, plus search

If your organization keeps manuals and SOPs on a shared drive or a SharePoint site, you can search them yourself or add a search tool on top. Enterprise search tools like Glean connect to more than 275 work apps and data sources, including Google Drive, SharePoint, Confluence and Slack Reference [20].

Where they are strongest: you control the content, and it includes your own SOPs and notes.

Where they fall short: a basic search box finds files, not answers. It will tell you which 400-page PDF mentions "pressure test," and someone still has to read it. Glean goes further: its assistant writes answers from your documents with citations Reference [21], and can point to the exact snippet that supports each statement Reference [22]. It is a general workplace tool, so test it on your own manuals before you buy. Put one scanned manual in the trial and check whether search can find text inside it.

Best for: teams that mostly need to find their own internal documents.

4. Instrument-specific AI tools

Some AI tools are built specifically for scientific instruments. Rayni, our product, is one of them. It answers questions across instrument types and manufacturers in one place, using the manuals, SOPs, notes and videos loaded for each instrument. Every answer cites the exact manual page it came from, so the person at the instrument can check a step before acting on it.

Where Rayni is strongest:

• One place for every vendor, so your team doesn't need a separate portal for each manufacturer.
• Answers draw on your own SOPs, notes and videos alongside the manufacturer's manuals, so what your team learns from one breakdown is there for the next.
• Each answer links to the exact page, and the system removes any citation it can't back up before you see it.
• It is in daily use. Innovative Lab Services, an independent service company, grew its Rayni coverage from 35 instruments to more than 250 and reports that 44% of its engineers' questions were resolved in the first reply Reference [23].

Best for: organizations that run instruments from several manufacturers, shared instrument facilities and independent service teams.

5. Forums, colleagues and training

Chromatography Forum has run since 1999 Reference [24] Reference [25]. Reddit communities and ResearchGate Q&A threads add more. CHROMacademy offers HPLC troubleshooting training Reference [26]. University students and staff get free access Reference [27]. Others can use a free tier with 5% of the content or pay $469 a year for full access Reference [28].

Where they are strongest: strange problems that need an experienced person's judgment, and building your own skills.

Where they fall short: you wait for someone to reply, and the person replying may be working from a different model. Check any fix against your own manual.

Best for: unusual problems and long-term learning.

How a grounded troubleshooting system works

Here is how Rayni does it. Other tools may work differently, so ask each vendor to walk you through its steps.

1. Read manuals the way a person would. We split documents at their real section boundaries, so a procedure like "Replacing the rotor seal" stays whole, and every figure gets a written description so diagrams can be found.
2. Find the right manual, then the right sections, and write the answer from those pages only.
3. Check every citation before you see it. If the answer cites a page the system didn't retrieve, we remove that citation.

Across the thousands of answers Rayni gave to real users from January to September 2026, 90% cited at least one specific manual page.

Security and data questions to ask

At many organizations, a cloud tool has to pass a security review before anyone buys it or even pilots it. Montclair State University, for example, requires a HECVAT review first Reference [29]. Ask these questions early, so the review has fewer surprises.

• Do you train AI models on our data? The answer should be no, in the contract.
  How Rayni handles it: For Rayni's premium customers, we do not train any AI model on customer data, and that commitment is in the contract.
• Do your AI providers keep our data? Retention and training are separate controls, so ask about both.
  How Rayni handles it: Rayni's premium customers get zero data retention, so the AI vendors Rayni uses keep no copy of your data.
• How is our data kept separate from other customers?
  How Rayni handles it: Isolation is enforced at six independent layers, and an automated test that tries to read across customers runs on every release.
• Where is data stored and processed?
  How Rayni handles it: In U.S. data centers, encrypted in transit and at rest.
• Who can see what inside our organization?
  How Rayni handles it: Admins manage users, Experts add and edit instrument documents, and Basic users can only read and ask questions. Admins can require two-factor sign-in.
• Can you show which sources produced a given answer, months later?
  How Rayni handles it: Yes. Every answer records the exact passages it was built from.
• Have you completed a HECVAT? Universities use this questionnaire to vet vendors, and version 4 added AI-specific questions Reference [30].
  How Rayni handles it: Rayni has completed these questionnaires.

How to run a fair trial

A guide in Lab Manager magazine on evaluating AI tools gives advice we agree with: test the tool on your organization's own data, and write down your acceptance criteria before the trial starts Reference [32].

A simple version:

1. Pick five to ten instruments your organization actually uses, including at least one older model.
2. Collect 20 real questions from your team's recent problems, including a few you already know the answer to.
3. For each answer, check the cited page. Mark it right, partly right or wrong.
4. Note how often the tool says "I don't know" when it should.
5. Ask the people who will use it daily which tool they would open next time.

When you don't need a tool like this

• Your organization runs instruments from one manufacturer and has the best tier service contract.
• Problems are rare and simple, and your team already knows the fixes.

---

Want to see how Rayni handles your instruments? Pick two of your trickiest recent problems and try them at www.rayni.ai.

---

References

All sources accessed October 1, 2026.

1. Thermo Fisher Scientific. Vanquish Pumps C, Pumps F (VC-Pxx, VF-Pxx) Operating Manual, doc. 4820.4405 rev. 5.1. Copy hosted by Carnegie Mellon University, Center for Molecular Analysis. https://www.cmu.edu/chemistry/research/facilities/cma/instruments/manuals/4820.4405-en_-rev_5.1-_vanquish_pumps_c,_f_operating_manual.pdf (official source: www.thermofisher.com/HPLCmanuals)
2. Carnegie Mellon University, Department of Chemistry. Center for Molecular Analysis: Instrumentation. https://www.cmu.edu/chemistry/facilities/cma/instruments/index.html
3. Walters WH, Wilder EI. Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports. 2023;13:14045. https://doi.org/10.1038/s41598-023-41032-5
4. Linardon J, Jarman HK, McClure Z, Anderson C, Liu C, Messer M. Influence of topic familiarity and prompt specificity on citation fabrication in mental health research using large language models: experimental study. JMIR Mental Health. 2025;12:e80371. https://doi.org/10.2196/80371
5. Zakka C, Shad R, Chaurasia A, et al. Almanac: retrieval-augmented language models for clinical medicine. NEJM AI. 2024;1(2):AIoa2300068. https://doi.org/10.1056/AIoa2300068 (open access: https://pmc.ncbi.nlm.nih.gov/articles/PMC10857783/)
6. Magesh V, Surani F, Dahl M, Suzgun M, Manning CD, Ho DE. Hallucination-free? Assessing the reliability of leading AI legal research tools. Journal of Empirical Legal Studies. 2025;22(2):216 to 242. https://doi.org/10.1111/jels.12413 (preprint: https://arxiv.org/abs/2405.20362)
7. Knill C, Devan S, Matrosic C, Moreau J, Zhang Z. AI-driven troubleshooting for TrueBeam systems: development and testing of a GPT-4.1 chatbot. Technical Innovations & Patient Support in Radiation Oncology. 2026;39:100420. https://doi.org/10.1016/j.tipsro.2026.100420
8. Agilent Technologies. Lab Advisor software. https://www.agilent.com/en/product/liquid-chromatography/hplc-ce-software/instrument-control-software/lab-advisor-software
9. SCIEX. Support. https://sciex.com/support
10. Shimadzu. Global Service Agreement LabTotal (Smart Service Net). https://www.shimadzu.com/an/service-support/service/global-service-agreement-labtotal/index.html
11. Waters Corporation. Support. https://www.waters.com/nextgen/us/en/support.html
12. Agilent Community. Gather and review log and results from Lab Advisor software (LC Portal article KP714). https://community.agilent.com/knowledge/lc-portal/kmp/lc-articles/kp714.gather-and-review-log-and-results-from-lab-advisor-software
13. OpenAI Help Center. ChatGPT Free Tier FAQ. https://help.openai.com/en/articles/9275245-chatgpt-free-tier-faq
14. Anthropic. Claude pricing. https://claude.com/pricing
15. Google. Gemini subscriptions. https://gemini.google/subscriptions/
16. Kalai AT, Nachum O, Vempala SS, Zhang E. Why language models hallucinate. arXiv:2509.04664. 2025. https://arxiv.org/abs/2509.04664
17. Harvard University Information Technology. Generative AI guidelines. https://www.huit.harvard.edu/ai/guidelines
18. The University of Texas at Austin, Information Security Office. AI tools. https://security.utexas.edu/ai-tools
19. University of California, Berkeley, Office of Ethics, Risk and Compliance Services. Appropriate use of generative AI tools. https://oercs.berkeley.edu/appropriate-use-generative-ai-tools
20. Glean. https://www.glean.com/
21. Glean Docs. AI Answers. https://docs.glean.com/user-guide/assistant/ai-answers
22. Glean Developers. Deep linked citations. https://developers.glean.com/guides/chat/deep-linked-citations
23. Rayni. Case study: Innovative Lab Services. https://www.rayni.ai/resources/case-studies/innovative-lab-services
24. Chromatography Forum. https://www.chromforum.org/
25. Separation Science. Chromatography Forum: a hub for the analytical chemistry community. https://www.sepscience.com/chromatography-forum-a-hub-for-the-analytical-chemistry-community-6518
26. CHROMacademy. HPLC troubleshooting. https://www.chromacademy.com/hplc/troubleshooting/
27. CHROMacademy. Academic membership. https://www.chromacademy.com/academic-membership/
28. CHROMacademy. Pricing. https://www.chromacademy.com/pricing/
29. Montclair State University, Information Technology Division. Vendor assessment (HECVAT). https://inside.montclair.edu/administration/information-technology-division/information-security/vendor-assessment-hecvat
30. EDUCAUSE Review. HECVAT 4: Better than ever. February 2025. https://er.educause.edu/articles/2025/2/hecvat-4-better-than-ever
31. Glean. Pricing. https://www.glean.com/pricing
32. Russell E. Evaluating AI tools for laboratory use: what lab managers should actually test. Lab Manager. https://www.labmanager.com/evaluating-ai-tools-for-laboratory-use-what-lab-managers-should-actually-test-35619