Research Client Portal: A Safe Door Into the Study
A research client portal lets clients ask the study after the debrief. Three locks make it safe: client mode, permission per document and a log of each quest...
The debrief went well. Two weeks later the client's category manager writes: how did younger women in the north answer the switching question, and is that different from last wave? The analyst who ran the study is already on the next one. She opens the tables, finds that cut was never run, reruns it, pastes the chart into a slide, exports a PDF and sends it. Three days later the second question arrives. A research client portal promises to end that loop by letting the client ask the study directly. Opened carelessly, it also lets the study answer things it never should.
Key Takeaways
- After the debrief, client questions still reach the agency by email, and each answer means a new cut, a new slide and a new PDF from an analyst who has moved on.
- A portal vendor wrote in January 2026 that research delivery often still relies on email PDFs, downloads or static file sharing; Research Live argued for permanent online analysis platforms in 2015.
- OWASP's 2025 guidance for language model applications warns that when several classes of users share one retrieval store, context can leak between users or queries.
- A safe door into the study needs three locks: a client mode separate from the administrator mode, permission set per document, and a log of every question with the sources behind each answer.
- Under the 2025 ICC/Esomar Code, researchers keep overall responsibility for their research whatever technology is applied, so the agency answers for what the door says.
What clients ask after the debrief
A research client portal is a private space where a client logs in to see their own studies, reports and data, and, increasingly, to ask questions about them in plain language. The portal replaces the attachment as the place where the study lives after delivery.
PERSONAL EXPERIENCE The questions that follow a debrief fall into a few kinds, and I have answered all of them by email. A new cut: the same question for a subgroup nobody tabulated. A definition: what exactly counted as a lapsed user. A comparison: is this different from the last wave? A permission check: can I put this number in the board deck? And the hardest one, a question the study never asked, phrased as if it had.
None of these is unreasonable. Each one arrives because the client is finally using the study, which is the point of having commissioned it.
How the agency answers today
By email, through the analyst, one question at a time. The analyst reopens the project, reruns the cut, checks the base, builds a slide, exports a PDF and writes two lines of caveats that will not travel with the chart. Most agencies still do this by hand, and the hours land on account service or on nobody.
The vendors selling portals describe the same baseline. When Content Catalyst launched its Spark portals in January 2026, it said that research organizations' delivery infrastructure often relies on basic email PDFs, WordPress downloads, or static file-sharing systems. That is a vendor's claim about its own market, and it matches what I see.
The idea of a permanent door is not new. In 2015 Brian Kushnir argued in Research Live, the Market Research Society's publication, that agencies should give clients 24/7 online analysis platforms instead of a single big reveal. Some firms built them. Spherix Global Insights opened a client portal with personalized views of subscribed research in 2023, and one professional services network described in a Dovetail case study lets clients log in as viewers, with folder-level permissions controlling what they see.
What changed is the chat box. A portal that shows reports is a library. A portal that answers questions is a colleague, and a colleague needs rules.
Why a chat over the study is risky
Retrieval is the step in which a language model application searches a document store for passages relevant to a question and uses them to write the answer. It is what lets a chat answer from the study rather than from general knowledge. It is also the step where the wrong document gets in.
The agency's store holds more than one client. It holds the study for the bank and the study for its competitor, drafts that were never approved, the internal notes on what the analyst did not trust. OWASP's 2025 entry on vector and embedding weaknesses puts the risk plainly: in environments where several classes of users share the same vector database, there is a risk of context leakage between users or queries. Its prevention advice is fine-grained, permission-aware access control, strict logical partitioning of datasets, and detailed immutable logs of retrieval activity.
The second risk is the answer itself. NIST's generative AI risk profile names it confabulation: confidently stated but erroneous content by which users may be misled. In a research portal, that is a chat inventing a cut the tables do not contain, or reading a subgroup of eleven people as a finding.
| Lock | What goes wrong without it | How to check it before opening the door |
|---|---|---|
| Client mode separate from administrator mode | The client sees agents, instructions, other projects or settings | Log in as the client and list everything visible |
| Permission per document | An answer draws on another client's study or an unapproved draft | Ask a question only another client's study could answer |
| A log of every question and source | Nobody can say what the client was told, or from which page | Pull the log of one client conversation from last week |
| Answers that cite the page | The client quotes a number nobody can trace | Click every number back to the table it came from |
| A base size rule | A thin subgroup is read as a finding | Ask for a cut you know has a small base |
The risks come from OWASP and NIST. The checks in the last column are ours.
Three locks between the client and the study
A vertical flow showing what a safe research client portal checks before it answers. It starts with a client question: how did younger women in the north answer the switching question, and is it different from last wave. Lock 1, client mode: the client sees their own studies and an answer box, while agents, instructions and settings stay on the administrator side. Lock 2, permission per document: retrieval only searches the documents this client is allowed to see. Below it are three document stores side by side: this client's approved studies, which are searched; another client's studies, blocked and never searched for this client; and unapproved drafts, blocked and never searched until approved. From the approved studies the flow continues to an answer that cites the page: every number links back to the table it came from, and a cut below the base size rule goes to the analyst instead. Lock 3, the log: the question, the documents retrieved and the answer are kept for the agency. A closing band says the agency keeps responsibility for every answer the door gives. The risks are named in OWASP LLM08:2025 and NIST AI 600-1, and the arrangement of the locks is Cassi.ai's recommendation. The figure contains no measured numbers.
Two of the three stores are never searched for this client. The log records which one the answer came from.
Who answers when the door says the wrong thing
The agency does. The 2025 revision of the ICC/Esomar International Code, published in September 2025, keeps as its fifth principle that researchers hold overall responsibility and oversight for their research, "irrespective of the method, technique and technology applied". A chat that answers the client on the agency's behalf does not move that responsibility anywhere.
PERSONAL EXPERIENCE In practice it lands on two people. The account director, who promised the client faster answers, and the research director, whose name is on the study when a number from the portal turns up in a board presentation without its base. Most of the agencies I talk to keep the door shut for exactly this reason. The client is asking. What nobody can say is what the door would answer.
This is the gap we listed in why a general chatbot is not a research platform, seen from the client's side of the table.
Open the door in this order
A client mode is the view a client user gets: their own studies, an answer box and nothing that configures the system. An administrator mode is the view the agency keeps: which documents each client can reach, which agents run, what the logs show. An audit log is the record of each question, the documents retrieved to answer it and the answer given, kept so the agency can reconstruct any conversation later.
UNIQUE INSIGHT The order matters more than the features. Each step below only works if the one before it is already in place. This is the sequence we recommend.
- Separate the client mode from the administrator mode before loading any material.
- Give each client their own space, scoped document by document, with drafts kept out until approved.
- Load the approved deliverables together with the tables and the questionnaire, so answers can cite the page.
- Write what the door will not answer: cuts below the base size the agency already uses in its tables, questions the study never asked, and forecasts.
- Log every question, every document retrieved and every answer.
- Route what the door cannot answer to the analyst, with the client's question attached.
- Read the log with the account team every month, because it shows what the client wants to know next.
Open the door in this order
Seven steps for opening a research client portal, each with a note on the right saying what breaks if it is skipped. Step one, separate the two modes, client mode and administrator mode, before any material is loaded; if skipped, the first client login sees how the system is configured. Step two, scope each client by document, in their own space, with drafts kept out until approved; if skipped, retrieval can reach another client's study. Step three, load the approved deliverables, tables and questionnaire together; if skipped, answers cannot cite the page. Step four, write what the door will not answer: thin cuts, questions never asked and forecasts; if skipped, thin cuts and unasked questions get answered anyway. Step five, log every question and source: the question, the documents retrieved and the answer; if skipped, nobody can reconstruct what the client was told. Step six, route the rest to the analyst with the client's question attached; if skipped, the hard questions go back to email. Step seven, read the log every month with the account team; if skipped, the brief for the next study stays unread. A closing band says each step only works if the one before it is already in place. The figure contains no measured numbers.
The order matters more than the features. A log without per-document permission records leaks instead of preventing them.
Step seven is the one agencies do not expect. The log of what a client asks is the best brief the agency will ever get for that client's next study. Kept next to the archive, it also does what we describe in reusing past market research: it shows which past answers are still being asked for.
Where Cassi.ai comes in
@Cassi.ai is a software engineering company specialized in the pains of market research, innovation and insights, founded by Cassiano Albuquerque after twenty years in market research. We automate research projects with or without language models.
Deskor is a Cassi.ai product that connects company knowledge, AI agents and workflows, so the team builds proposals, decks, reports and dashboards from what the company already knows.
Deskor is our platform for putting a firm's own knowledge to work, and its Client Portal module is the door described here. Each client gets a branded, isolated space with their projects, their reports and a chat over their own account's material. The firm keeps the administrator side, with governance over which model runs each task and an audit trail of what was done, and answers point back to the source documents. When the question needs a dashboard rather than a sentence, the logic is the one in a market research dashboard built from the decision.
FAQ
What is a research client portal?
A private space where a client logs in to see their own studies, reports and data, and increasingly to ask questions about them. Vendors such as Content Catalyst sell it as a replacement for email PDFs and static file sharing.
Is it safe to let clients chat with research data?
Only with access control per document, separate client and administrator modes, and a log of every question. OWASP flags context leakage between users who share one retrieval store.
What should a portal refuse to answer?
Cuts below the base size the agency uses in its own tables, questions the study never asked, and forecasts. Those go to the analyst.
Who is responsible for what the portal tells the client?
The agency. The 2025 ICC/Esomar Code keeps overall responsibility with the researcher, whatever technology is applied.
Before the door opens
UNIQUE INSIGHT Clients do not stop asking the study questions when the agency stops answering them. They ask someone else, or they guess. A door with three locks keeps those questions inside the study, where somebody can see them.
Count the follow-up questions your team answered by email after the last three debriefs, and how many days each one took. If you want to see the door running on one of your own studies, we will set it up with your deliverables and show you the log.
About the author
Cassiano Albuquerque is the founder of @Cassi.ai. He spent twenty years in market research operations before building software for research teams. To talk about opening your studies to clients safely, contact him on LinkedIn.
Published by Cassi.ai. Read the full article at https://www.cassiai.com/blog/research-client-portal-access.