Task-focused, human-controlled: Useful AI in grants management

Artificial intelligence (AI) can save grant teams meaningful time, but it works best when it has a clearly defined purpose.

In Good Grants, AI can support tasks such as summarising applications, extracting information, identifying themes, drafting applicant feedback, working through calculations and analysing information across applications, reviews, allocations and grant reports.

AI can also assist with preliminary assessments and other repetitive analysis—all while keeping people responsible for the decisions that matter.

But these jobs do not all require the same thing from AI.

Whether AI sits within a grants management platform such as Good Grants or forms part of external workflows, the principle is the same: start by defining the job you want AI to perform, then shape the configuration around it.

In this article, we take a deep dive into how to make AI useful for grantmaking, focusing on the tasks at hand and building effective prompts, and how Good Grants AI tools can help.

Introducing Good Grants AI tools

Good Grants enables grant teams to create purpose-built AI configurations using supported large language models (LLMs). Each configuration can have its own name, model, temperature and pre-prompt instructions to give the AI context about its role, the information it should consider, and how it should respond.

The value is in the flexibility. You decide what role AI should play, where it belongs within your grantmaking workflow and what you want it to help you achieve.

AI is not there to run your grant program for you or make decisions better left to humans. Nor are you restricted to a small set of predefined AI functions. Instead, you can create focused, task-specific AI tools for the work your team actually needs to do.

Each configuration can be built for a specific task, such as application summaries, eligibility assessment, grant report analysis or applicant feedback.

Focus on the task

Good Grants allows you to choose which AI model you want to use. Models do differ, particularly for more complex tasks, and considerations such as availability, speed and token consumption matter.

But for most grantmaking use cases, choosing the model should not be your starting point.

Start with the task. What job do you want Good Grants AI tools to perform?

Take application summarisation as an example. You want the AI to identify the most relevant information in an application, represent it accurately, avoid introducing unsupported assumptions and return the information in a useful, predictable format.

That already tells you quite a lot about how the configuration should work:

• Purpose: Summarise applications
• Behaviour: Factual, concise and neutral
• Temperature: Low, to favour predictable responses (more on this below!)
• Context: Explain what information is available and which parts are most relevant to the task
• Output: Define the structure, level of detail and approximate length you need

You might call this configuration “Application summary” and use it wherever that summary is useful within your workflow.

Now consider a different job: identifying common needs or themes across a group of applications.

You could use exactly the same underlying AI model, but give it different instructions and a different temperature because this task calls for more interpretation and exploration.

The technology underneath has not changed. The job has—and so should the configuration.

Give AI the right instructions from the start

A useful AI configuration starts before anyone types a prompt.

This is where the pre-prompt comes in.

A pre-prompt provides the AI with standing instructions before the user’s own prompt is added. It can define the role the AI should perform, what information deserves attention, what it should avoid or never assume, and how its response should be structured.

Without one, users may need to explain the same requirements every time they ask the AI to complete a task.

With a well-designed pre-prompt and an appropriate temperature, much of that thinking is built into the configuration.

Return to our application summary example. Instead of repeatedly explaining what should be included, what should be excluded, how factual the response should be and how it needs to be formatted, the user might simply ask: Please summarise this application.

The AI already understands the rest, due to the configuration settings.

That saves time, particularly for jobs repeated across hundreds or thousands of applications. More importantly, it improves consistency.

Important instructions no longer depend on someone remembering the right prompt, copying it from another document or asking AI slightly differently each time. The configuration carries those instructions with it.

Use temperature to control variation

If the pre-prompt helps define what the AI should do, temperature helps determine how much variation you want in its response.

At a basic level, lower temperatures tend to produce more consistent and conservative responses. Higher temperatures allow more variation and creativity.

Neither is inherently better.

The appropriate temperature depends on the job.

For application summaries, calculations, information extraction, eligibility-related assessments or other tasks where consistency matters, a lower temperature will usually make more sense.

For drafting applicant feedback, exploring themes or working through tasks where different ways of expressing or interpreting information could be useful, greater variation may be appropriate.

The important thing is to choose deliberately rather than relying on the same setting for every task.

Name configurations for their purpose

Once you have several configurations, naming them becomes surprisingly important. A label such as Claude or Mistral tells the user which technology sits underneath, but not what the configuration is actually for.

Names such as “Application summary”, “Eligibility assessment” or “Grant report analysis” make the intended purpose clear and help users think about AI in terms of the job it needs to do.

It also gives grant teams flexibility behind the scenes. If a new model becomes available or an existing one becomes available in their region, they can simply swap the model without changing the configuration—or the task it supports.

Tips for using Good Grants AI tools

Keep each configuration focused

You can put a lot of information into a pre-prompt. That does not necessarily mean you should.

A useful configuration gives the AI enough information to understand:

• The role it needs to perform
• The information it should pay attention to
• What it should avoid—and what it should never assume
• How its response should be structured
• The tone it should use
• How much variation is appropriate

Instructions unrelated to the job can make a configuration harder to understand, test and maintain.

If one pre-prompt starts accumulating instructions for application summaries, eligibility reviews, applicant communications and financial analysis all at once, it is probably time to create another configuration.

One clear job is easier to configure well than several competing ones.

Test the whole configuration

Do not test only the prompt. Test the complete configuration.

AI configurations should be approached like any other part of your grantmaking workflow: try them against a realistic range of information before relying on them at scale.

For example, test an application-focused configuration with detailed applications, concise applications, unusual wording, missing information and genuine edge cases from your program.

Then compare the results. Does an application summary consistently focus on the information you care about? Does an eligibility assessment distinguish between a requirement that has not been met and information that has simply not been provided? Does applicant feedback maintain the intended tone? Does an analysis avoid making assumptions beyond the information available?

If not, adjust the configuration and test it again.

The prompt is only one part of the equation. The model, pre-prompt, temperature, context and desired output all influence the result.

Balance quality, speed and cost

The job should come first, but the underlying model still matters.

Different AI models vary in capability, speed and cost. A straightforward task to extract information from an application may not need the same model as a more complex analysis involving lengthy applications, reviewer feedback and grant reporting history.

The same applies to the response itself. There is little value in asking for a detailed essay if the grant manager using the output only needs five clear bullet points.

For jobs run frequently, small differences in response length or model cost can add up at scale.

Instead of asking which model is universally best, consider which combination of model, instructions and settings gives you the right balance of quality, consistency, speed and cost for the particular job.

Keep humans responsible for the decisions

AI can help grant teams process information faster, surface useful evidence and reduce repetitive administrative work.

That does not mean every part of grantmaking should be handed over to AI.

Decisions involving eligibility, application assessment, funding recommendations, allocations and applicant communications can carry significant consequences. They also raise important questions around fairness, governance, accountability and trust.

AI can support those processes. It can organise information, summarise evidence, identify inconsistencies, analyse reviewer feedback, bring relevant reporting information to the surface or help people navigate large volumes of data more efficiently.

But supporting a decision is different from making one.

Humans should remain responsible for consequential funding decisions and able to understand the information those decisions are based on.

The goal is not to take people out of grantmaking. It is to give them useful tools that reduce repetitive work and make the information they need easier to understand.

Build a practical AI toolkit for your grant program

The real opportunity with configurable AI is not access to an ever-growing list of models.

It is the ability to create a collection of clearly defined tools around your grantmaking work.

You might have one configuration to summarise applications, another to assess information against shortlist requirements, another to analyse grant reports, another to surface information from reviewer feedback and another to help prepare applicant communications.

Each has a particular purpose, its own instructions and settings, and a clear place in your workflow.

This makes AI easier for teams to use consistently, easier to test and easier to govern. It also means every person working on your grant program does not need to become an expert prompt engineer. Much of the thinking can already be built into the configuration before they use it.

That is where configurable AI becomes genuinely useful: not one general assistant trying to do everything, but a collection of focused tools designed around the jobs your grant team actually needs to do.

What grantmaking job can Good Grants AI tools help you with today? Test it out and let us know!