AI in research support: From individual prompting to shared AI workflows

← Back to resources

Research support services work with funding requirements, draft proposals, unpublished research ideas, internal assessments and tight deadlines. AI can help structure this information, identify gaps in an argument and prepare advisory meetings.

Reliable use of this support across a team requires shared working practices. Before the first prompt, establish which information may be used in which tool for which task, and who will review the result.

AI literacy starts before the first prompt

Here, AI literacy means the ability to use AI effectively, assess its limitations and review the substance of its results. ^[1] This includes recognising confidential proposals, personal data and research ideas that need protection. This assessment must not be left to the AI system.

Four questions provide the starting point:

• Material: What information is available, and how is it protected?
• Tool: Which AI system is approved for this material and task?
• Task: What specific support should the AI provide?
• Output: How will the result be reviewed and then used?

The AI Traffic Light and the four-question check support this initial assessment. They provide guidance; the applicable institutional and funder-specific requirements must also be considered in each case.

Why research support services need tailored AI workflows

Research support services and grants offices turn funding requirements into practical advice. They review drafts, explain criteria and help researchers present their scientific contribution, methodology and feasibility convincingly.

A general AI assistant does not automatically have the current funding call, internal advisory standards or the intended use of its feedback. The working context therefore needs to be provided explicitly. Shared roles, sources and review steps help align the assistance with these professional requirements.

May a grants office use AI for funding proposals?

This depends on the specific task, material, tool and applicable rules. Potential tasks in internal proposal support include reviewing structure, turning funding criteria into review questions or refining wording. Confidential documents may only be processed in a suitable, approved environment.

Internal advice before submission and formal peer review for a funder are different procedures. An assistance model for grants advisers therefore does not amount to general permission to use AI in peer review.

In formal SNSF procedures, reviewers first form an independent assessment. AI may then provide assistance only under the specified conditions, including confidentiality and disclosure. The current SNSF FAQ explains the tasks and boundaries. ^[2]

The problem with individual AI use

When everyone uses their own prompts, document extracts and review criteria, they work from different foundations. A convincingly worded answer may have been carefully checked or may rely on an outdated document, without this being apparent from the result.

A shared workflow records which sources and templates are used, which steps are repeatable and where professional review is required. Staff need subject-matter expertise and a shared foundation in AI literacy. They do not need to redesign each assistance role every time they use it.

The Sollberger AI-Literacy-Workflow

The Sollberger AI-Literacy-Workflow describes my approach to developing these processes. It is a practical guidance model, not a binding standard issued by a university or funding organisation.

Assess the material → Curate knowledge → Consolidate context → Use a focused AI assistant → Review the output → Take responsibility

1. Assess the material: Clarify public availability, confidentiality, personal data and protection requirements. Check whether the material may be processed in the intended tool.
2. Curate knowledge: Select current funding calls, guidelines, FAQs and internal guidance. Record versions and validity; flag outdated or unresolved information.
3. Consolidate context: Bring requirements, terminology, exclusions and open questions together in a manageable reference document. A suitable knowledge tool can help; source traceability and subject-matter review remain necessary.
4. Use a focused AI assistant: Choose a clearly bounded role, such as structural guidance, writing support or preparation for an advisory meeting. Provide only the information needed for that role.
5. Review the output: Check accuracy, completeness and suitability against the sources. Look for fabricated details, inappropriate weightings and unsupported claims of impact. ^[3]
6. Take responsibility: Decide who approves, reuses or communicates the result. Subject-matter responsibility remains with the designated person.

From scattered documents to a Single Source of Truth

Funding information is often spread across calls, regulations, FAQs, presentations and internal notes. A Single Source of Truth (SSOT) brings the relevant information about a funding scheme together in a reviewed reference document: requirements, criteria, deadlines, terminology and open questions, with their sources.

Such a document can provide temporary context for a general AI assistant for grants advisers. Its core role stays the same, while the supplied reference document defines the specific funding context. Not every piece of time-limited specialist knowledge needs to be permanently added to its knowledge base.

Even summaries of supplied documents can be inaccurate or misleading. Compare the result with the original sources. ^[3]

SSOT is an organisational principle. A compact document can contain errors and does not replace original sources. Ownership, version information and source references make it reviewable; changes to a call or FAQ need to be reflected in it.

Hypothetical example: From a funding call to a checklist

The following example illustrates a possible process and does not describe a client project:

1. Source material: Select a current public funding call and its official FAQs as sources.
2. AI task: Structure requirements by eligibility, deadline, length and documents to submit; identify the source location for each point and flag ambiguities.
3. Human review: A grants adviser checks every point against the original source and corrects omissions or misinterpretations.
4. Usable result: A dated internal checklist with source references. The applicable call documents remain authoritative.

How AI assistants can support quality across a team

A domain-specific AI assistant can include a defined role, typical review questions, terminology, output formats and boundaries. This gives the team a shared starting configuration for recurring tasks.

This makes agreed review criteria easier to apply, but does not guarantee correct results. Before wider use, the team should test an assistant on representative tasks: What errors occur? Which sources are missing? When should it ask a question or decline a task?

How to extend this approach across a team

Other staff can adopt documented roles and approved templates. Changes to sources, rules or assistant instructions can be maintained together rather than updating many personal versions individually.

This requires an accountable owner, accessible guidance and regular feedback from use. A pilot should first demonstrate whether the process is clear to the team and meets the required quality before it is extended. ^[4]

For the workflow described here, I recommend the following: after major changes to the model, tool or knowledge base, the team tests the assistant again using agreed examples.

Three practical building blocks for grants offices

The following assistance roles describe possible working methods in internal proposal support. Their suitability needs to be assessed within the relevant institutional framework.

Review assistant for grants advisers

An internal review assistant can raise questions about a draft proposal: Is the scientific contribution clear? Does the methodology fit the research question? Are the risks explained? Are claims of impact supported by the draft?

This critical reading perspective can help grants advisers prepare for advisory meetings. It is an internal working aid, not an independent expert review. In the workflow proposed here, it produces neither rankings nor funding decisions; a professional reviews and weighs each piece of feedback.

Proposal-writing assistant for researchers

A writing assistant can help researchers structure existing ideas and express them more clearly. For example, it can flag unclear transitions or check whether the research question and objectives are worded consistently.

The scientific idea, methodological substance and responsibility remain with the researchers. The assistant should ask for missing information and flag uncertainty rather than adding results, references or promises of impact.

SSOT workflow for specific funding contexts

Programme documents are first structured and consolidated into an SSOT document with subject-matter review. The reviewed document is then supplied as context for a specific assistance task.

This two-stage process separates source preparation from text work. It makes the information underlying feedback traceable. Checking against original sources remains part of both stages.

What AI can contribute to funding advice

Suitable tasks include structuring funding calls, turning criteria into review questions, flagging gaps in an argument, preparing advisory meetings or making the wording of information materials consistent. A team should select useful tasks based on its actual advisory needs.

What AI must not do

For the advisory workflow described here, I recommend clear boundaries: no automated funding decisions or rankings, no AI-generated texts presented as expert reviews, and no processing of confidential documents in unassessed environments. Fabricated evidence, scientific results or promises of impact must have no place in the work.

These boundaries describe my consulting approach. The rules of a formal assessment procedure determine whether and how AI may be used in it. Responsibility for subject-matter assessment cannot be fulfilled through a superficial check of an AI-generated text afterwards.

Boundaries: Material, tool, task and output

A public information sheet, a draft proposal containing personal data and an unpublished research idea require different safeguards. Tool approval needs to cover the specific use; institutional access alone does not settle this question.

Where personal data is involved, the relevant data protection requirements need to be checked. The FDPIC explains that the Swiss Data Protection Act also applies to AI-assisted data processing. Institutional rules, confidentiality agreements and funder conditions must also be considered, among other requirements. ^[5]

When preparing a specific task, the AI Traffic Light, four-question check and the article on effective AI communication complement one another: assess the use, clarify the conditions and formulate the approved task precisely.

Shared AI workflows are part of organisational development

Implementation affects task allocation, skills and collaboration. The team needs to decide who maintains sources, looks after assistants, reviews results and approves changes. Training and joint testing are therefore part of implementation.

A manageable pilot with a clear task makes these requirements concrete. Feedback on errors, effort and advisory quality provides the basis for improving the process or limiting its use.

Conclusion

Shared AI workflows combine subject-matter expertise with curated sources, clear assistance roles and reviewable results. The Sollberger AI-Literacy-Workflow offers a six-step approach. Its usefulness needs to be demonstrated within each team and through its actual tasks.