AI drafts and reviews. It never gets the last word.
Nonprofits have one fear about AI in grant writing, and it is the right one: a fabricated number reaching a funder. TGS was designed around that fear. Here is exactly what the AI reads, what it is allowed to say, and what stops it.
What the AI reads
Every AI feature in TGS is grounded in sources you can see. There are three layers of knowledge, and each reply shows which of them it drew on.
- The workspace. The parsed RFP is the rulebook: its requirements, scoring criteria, page limits, and attachments. Workspace tools (fit analysis, requirements, narrative, GrantCoach) read it and nothing else unless you add more.
- Your organization. The Knowledge Library: past applications, outcomes reports, budgets, and standard attachments you uploaded. Drafts and rewrites pull from these. Snippets you saved are never injected automatically; you insert them.
- Public data you turn on. Census ACS, CDC WONDER and PLACES, NSDUH, HRSA shortage areas, USAspending, ProPublica nonprofit records, and web search. Each returns its source and the source rides along into the draft.
TGS Studio conversations are private to the person who started them. Other people in your organization, including admins, cannot read them unless you save a document into the shared libraries.
What it is never allowed to do
Guess a number. If TGS cannot cite a figure from your sources or from public data you turned on, it leaves a placeholder instead of filling one in, and anything it cannot support is flagged for you to resolve before it goes anywhere.
Responsible AI
- Every request is checked three times. The AI is told to cite or leave a placeholder, code that does not trust the AI then checks every percentage, dollar amount, count, and rate against your sources, and a separate grounding check catches claims that sound right but are not supported. Anything that fails is marked
[UNVERIFIED: cite source before use]before you see it, the flag travels with the text if you paste it into a narrative, and Review cannot be marked complete until every flag is resolved with a source or removed. - Sensitive information stays out of prompts. TGS warns you if what you enter appears to include personal or protected health information before it is sent to the AI. You decide whether to proceed, and that decision is logged.
- Your data is not used to train AI models. TGS runs on models from Anthropic through Amazon Bedrock inside our AWS environment, under terms that prohibit training on customer data. Your narratives, uploads, and conversations are used only to complete your request.
- AI use is controlled and monitored. Each organization has a daily AI budget so a runaway loop cannot become a surprise bill, and TGS tracks which feature ran, token counts, and cost without storing the content of your work in those records.
What we do not claim
We do not claim the AI will write a winning grant on its own. It will not. It drafts from what your organization has already proven, it reviews against what the funder actually asked for, and it refuses to state as fact what it cannot source. The strategy, the relationships, and the final call are yours.
Watch it flag a number.
The fastest way to understand the guardrails is to see one fire on a real RFP. Send us one and we will show you.