Agent & MCP Readiness
Last updated: November 18, 2026
Deciding whether your organization and your workspace are ready for AI-assisted research, and who should reach it first.
Organizational readiness signals
Readiness is less about enthusiasm than about whether the conditions for good research already exist. AI-assisted creation raises throughput. It does not inherently supply judgment that was missing. A well-tailored Rally workspace for AI-supported use cases can guide better judgment and configuration.
Are you adding capability, or substituting for it? Teams that are stretched and want to serve more demand tend to do well. Teams hoping an assistant will replace the review a researcher was providing tend to produce more studies and worse ones.
Signals you are ready
Research demand exceeds what your research team can serve, and the overflow is going unmet or going to tools you do not govern
Someone owns the Rally workspace and has time to maintain it
Leadership wants research across more teams, rather than tolerating it
Security and legal have a position on AI tools touching customer data, even if the answer is restrictive
Your research team is willing to review work they did not produce
Your team can name the questions they answer repeatedly, which is what the first round of guidance will hold
Signals to wait
No one owns the workspace, so nobody maintains what the assistant reads
The research team is the only group asking for this
Security and legal have not been consulted, and you plan to ask after rollout
The stated goal is fewer researchers rather than more research
An AI assistant can set up a study, find an audience, and draft the outreach. It cannot decide the study was worth running or notice that the findings do not support the recommendation. Organizations that treat it as headcount relief may discover this at the point where someone acts on weak research.
Workspace readiness signals
To see how ready you are to open up access to AI with Rally, ask:
Could a capable new researcher do their own study well using only what is configured, without asking you a question?
If the answer is no, find out which questions they would have to ask. That question names the gap. A new person asks a colleague. An AI assistant does not ask. It fills the gap from whatever your configuration suggests, and the result looks finished either way.
Common versions of that question, and where each points:
The question a newcomer would ask | Where the gap lives |
Do we have this information on participants? | Properties |
Which of these populations do I use? | Populations, Team-scoping |
Is there a specific setup for this kind of study? | Templates, Skills |
What level of consent applies here? | Consent |
What do we usually pay for this? | Workspace Knowledge |
Am I allowed to contact these people? | Governance Rules, Populations, Team-scoping |
What does the product name mean? | Workspace Knowledge |
Governance is the one exception to the test, and the reason the rest is safe to attempt. Contact rules, cooldowns, and limits are enforced by Rally on every path, so a gap elsewhere produces a study that is wrong rather than a participant who is harmed.
Readiness is also not a state you reach once. A workspace that passes this test today fails it as soon as a new program adds properties nobody described, or a template library grows past the point where the right one is obvious.
Choosing access pathways
Rally is reachable through the web app, through Rally’s own Agent, and through external AI tools connected over MCP with either read or write access. Each is granted separately on the role. Authoring the skills and knowledge that guide AI-assisted work carries its own permissions.
These are separate decisions rather than a single switch.
Read-only external access is the low-cost starting point. People can ask questions, check feasibility, and explore the database without changing anything. It also surfaces data quality problems while they are still cheap. A feasibility result that comes back far below expectations is telling you something about your properties before anyone has acted on it.
Rally’s own Agent carries read and write together and comes with any paid seat. All built-in roles in Rally have this capability. You may turn this off for custom roles, if needed.
External write access is the one to stage. It is the same capability through a tool whose behavior and data handling you do not control. The section below covers why that distinction matters.
Authoring skills and knowledge belongs with whoever maintains your standards. Both change how AI-assisted work behaves for everyone, which puts them alongside templates and governance rules rather than with study execution. See 📄 Skills & Workspace Knowledge.
Widen by group before widening by capability. A small pilot group with read & write access teaches you more than the whole organization with read-only. You learn what people actually ask for, where your configuration misleads, and what guidance is missing.
The gate between stages is evidence rather than elapsed time. What you are looking for: are the studies coming out of the pilot ones you would have approved? If yes, widen. If they are close but consistently wrong in the same way, that is a configuration gap, and fixing it is worth more than more training.
PII in connected AI assistants
This is the decision most likely to stall a rollout, and the one worth settling before anyone connects anything.
Where Rally’s protection stops
When someone connects an external AI Assistant to Rally like Claude or ChatGPT, they authorize it as themselves, and Rally’s permissions follow them. They see what their role allows and nothing more. That part is sound.
What Rally’s permissions do not govern is what happens to the data after it leaves. Participant names, email addresses, and any property marked as personal information travel to whichever AI Assistant that person connected, under whatever agreement exists between your organization and that provider. PII may be included in AI responses if the user was granted permission to see it. Once that PII is in the AI response in your connected tool, Rally cannot govern what happens to that information next.
Rally’s own Agent and an external MCP tool are different decisions, even though they look like the same capability. Work done through Rally’s Agent is covered by the agreement your organization already has with Rally. Work done through an external tool is covered by your agreement with that provider, which may be an enterprise contract your security team negotiated, or may be an individual’s personal account.
Use these questions to help evaluate whether PII should be masked for roles with permission to connect to Rally through an external AI Assistant.
Which AI tools has your organization approved, and does that approval extend to customer personal data?
Do the people who would connect know the answer, and will they check before connecting?
If the answer is no for some tools, is that a policy you can enforce, or a preference people will route around?
Masking PII impacts all paths
The strongest control available is turning off PII visibility for the roles that connect external tools. It applies everywhere rather than depending on anyone remembering, and it survives a person switching tools. PHI masking works the same way.
When PII is masked names, emails, and any other property that is flagged as PII in the Rally database will render as *** instead of the actual value in every pathway, not just AI. Users without permission to see PII properties won't be able to filter, target, or review that data to support recruitment selection, interview prep, or participant analysis.