An assistant that shows its working
Grounded AI is not a chatbot with a nicer tone. It is a retrieval step, a guardrail step and a citation step, each of which can be inspected. Here is the same assistant handling a question it can answer and a question it should refuse.
Passages retrieved from the document library
Guardrails
Illustrative example. The organisation, documents, passages and answers shown here are invented. The retrieval, guardrail and citation steps are how the assistant actually works.
The situation
DonateLife Victoria coordinates organ and tissue donation across the state. Its staff work to clinical guidelines, hospital procedures and operational protocols that run to hundreds of pages, stored as documents in SharePoint.
The knowledge was all there.
Finding it was the problem.
A nurse in the middle of a donation pathway needs one specific fact, quickly and correctly. Which hospital receives a liver going to South Australia, and at what address. Where to park at The Alfred. Who to contact at Barwon Health. Every one of those answers was sitting in a table on a page inside a long document, and the only way to get it was to know which document, open it, and scroll.
DonateLife Victoria wanted staff to ask a plain English question and get a fast, accurate answer with a link to the source. The rules were firm. Answer only from DonateLife's own approved documents, never from general knowledge. Always show where the answer came from. And leave the documents in SharePoint, where the clinical teams already maintain them.
In a clinical setting, a confidently wrong answer is worse than no answer at all. That principle shaped the whole build.
What we built
GravityLab designed and built Sal, an assistant on Salesforce Agentforce grounded entirely in DonateLife Victoria's own documents.
It only answers from approved documents
Sal searches the document library, answers from what it finds, and lists its sources. It cannot answer from general knowledge, and it cannot be drawn into answering questions the documents do not cover. When it does not know, it says so.
The documents stay where they already live
SharePoint remains the single source of truth. Clinical teams keep publishing exactly as they do today, and the assistant picks up changes on its own. There is no second copy of the knowledge to maintain and nothing to keep in step by hand.
It turns PDFs into answers, including the tables
This was the hardest part of the build and the one that made the difference. AI search reads narrative text well and tables badly, and DonateLife's most-needed facts (hospitals, addresses, contacts, thresholds) live in tables. We built a process that reads every table in every document and rewrites each row as a plain sentence the assistant can actually find, so the questions staff ask most often now get answered instead of missed.
Every answer links back to the source
Sal cites the documents it drew on and links straight to the file in SharePoint, so a clinician can open the original and check it. That matters in a clinical setting: staff are not asked to take the assistant's word for it.
It looks after itself
Adding a document means uploading it to SharePoint. Removing one means deleting it. The assistant keeps itself up to date automatically, with no ongoing administration inside Salesforce.
It was tuned on evidence, not impressions
Rather than judging quality by trying a handful of questions, we built a test set of real questions with verified answers and measured against it. That is what identified the table problem, ruled out the fixes that did not work, and proved the one that did.
The outcome
Sal went live in production in September 2026 and is rolling out across the DonateLife Victoria team.
- Staff ask questions in plain English and get answers in seconds, drawn from 26 operational and clinical documents.
- The information locked inside tables is now reachable. The liver to South Australia question that started the investigation returns the right hospital straight away.
- Every answer comes with its sources and links back to the original document.
- The document library maintains itself, so keeping the assistant current is a SharePoint task, not an IT task.
- DonateLife Victoria has a plain-language support handbook, so the assistant is supported by their own team rather than depending on developers.
Just as important is what Sal will not do. It will not answer from general knowledge, it will not stray off topic, and it is built to say the documents do not cover something rather than guess.
What's next
- Continued rollout across the wider clinical team, with adoption support and training.
- Ongoing refinement based on the questions staff actually ask.