Designing an AI agent for high-stakes, regulated work

How I designed Clio, the AI agent behind CiteMed's medical evidence platform.

citemed.com ↗

Role
Senior Product Designer
Sole designer
Timeline
Feb – Jul 2026
Worked with
CEO
Product manager

Engineering
Industry
Healthcare technology
B2B SaaS
Clio works across the whole platform: finding and listing studies, summarizing and extracting data, checking values against the paper and editing the report.

Overview

Overview

Before a medical device can go to market, its maker has to prove it is safe and works as claimed, largely from published clinical studies. CiteMed is where that work happens. Teams search for studies, screen thousands of abstracts, extract data such as sample sizes and outcomes, and compile reports for regulators. The work is repetitive, precise and audited. One error can delay a submission by weeks.

That makes AI both valuable and risky. Clio is built as an agent. It doesn't just answer questions. It plans and does the work with the platform's own tools, from filtering the library to drafting an edit to a report.

I designed what researchers see and, with the engineers, how Clio behaves underneath. Seven questions shaped both.

Tell me what you're building, and I'll ask a few questions to understand it.

/01. Where it helps

Where does Clio create value?

I started from the workflow, not the model: where researchers spend their time, what repeats enough to automate, and what needs their judgment. Clio turned out to help in six places.

  • Reference library: finds, summarizes, tags and shares studies.
  • Tables: adds columns that answer one question across every study.
  • Smart lists: turns a plain-language brief into filters that stay up to date.
  • Screening: recommends a decision and scores relevance against the protocol.
  • Extraction: pre-fills values for the researcher to check.
  • Reports: proposes edits and drafts replies to reviewer comments.

No single chat window covers that range. One-off and multi-step requests suit a conversation. Work that repeats on every study belongs inside the tool, where it runs at scale. Clio does both, with one agent behind them.

A one-off request in the chat: pick Summarize, set the scope and focus, and get a summary of each of the 12 studies.

/02. Context

What does Clio know before you ask?

Without context, every request starts from zero. I defined three layers of context Clio always holds: what the researcher is looking at, the project they're in, and what they've just done.

The last layer lets requests build on each other. A researcher summarizes 12 studies on continuous glucose monitors, then asks Clio to extract data from "the 12 summarized articles". Clio knows which ones. No searching again, no picking rows by hand.

Clio works out which studies “the 12 summarized articles” means, then extracts with the team’s Outcomes template.

What to extract comes from the client. Two companies can read the same study and need different data from it, because their devices and claims differ. So each client sets up its own extraction templates: the list of fields to pull from every study. Clio fills in those fields. It doesn't choose them.

/03. Autonomy

How much does Clio do on its own?

A summary that misses a nuance is cheap: the researcher notices and runs it again. A wrong value in an extraction table can slip into a regulatory report unnoticed. So how much Clio does on its own depends on what a mistake would cost. I sorted every task by that, from low risk to high, which gave four levels:

  • Easy to undo, like search and summaries: Clio acts directly.
  • Changes to a document: Clio proposes tracked edits.
  • Screening and extraction, where an error can reach a regulator: Clio recommends, the researcher decides.
  • Anything that reaches other people, like sharing a list: nothing happens without explicit approval.

Multi-step work shows this most clearly. One instruction (find trials of glucose monitors in type 1 diabetes that report severe low blood sugar, put them in a list and share it with the clinical team) becomes a three-step plan. Clio finds the studies on its own, then stops at each step that changes data or reaches people. I modelled where those checkpoints sit, and what Clio reports at each one, with the engineers.

One instruction, three steps. Clio finds 6 studies on its own, then waits for an OK before it creates the list and again before it shares it.

/04. Intent

How does Clio show what it understood?

A plain text box hides how the model reads a request. For a quick question, that's fine. For a task across hundreds of studies, a misread request is expensive, and often goes unnoticed until much later.

So every request becomes a form the researcher can check before anything runs. Typing # opens a list of actions. Picking one splits the request into parts they can edit, such as scope and focus. Anything missing stays visibly empty instead of being guessed. Where there's a choice, like an extraction template, Clio previews every field first. What the researcher sees is exactly what Clio acts on.

Smart lists take this furthest. A researcher describes the list in their own words: peer-reviewed studies since 2019 on non-invasive glucose monitoring for adults 65+ with type 2 diabetes. Clio turns that into rules they can read and edit. Clear-cut criteria, like age or publication type, become standard filters. Fuzzier ideas become AI rules. The original request stays above the rules, and every result explains why it matched.

Described in plain words, turned into rules, then changed in plain words: “include adults from 60” adds 6 articles.

/05. Verification

Where does the work go, and how is it checked?

Results land in the product, not in the chat. Extractions become tables, AI columns stay in the study view, and report edits arrive as tracked changes. Each comes with a short note on what Clio did and where the result is.

Defined once, filled in for all 18 studies. Where a study has no answer, the cell says so.

When a model is usually right, people start approving out of habit. In regulated work, approval is what makes data valid, so checking has to be quick enough that people actually do it. In extraction, every value links to the passage it came from and is approved or rejected on its own.

Each value opens its source in the paper. Where the paper is unclear, Clio flags the value and asks.

Reports already have a review process, so Clio works inside it. Its changes are tracked edits, and approved versions are locked.

Clio says what it changed and what it kept. Each edit is accepted on its own.

Reviewer comments go through the same process. The researcher sets limits, such as “no new references”, and Clio drafts within them.

One comment becomes one tracked sentence, accepted and resolved in one step.

/06. Failure

How does it fail?

Failure needs as much design as the expected path.

  • Unclear request: Clio asks a question instead of guessing.
  • Few results: it suggests how to broaden or narrow the search.
  • Sources disagree: it flags the conflict instead of settling it quietly.
  • Data missing: it marks the value “Not reported” instead of filling the gap.
  • Change of mind: researchers can edit any rule or value, stop long tasks, cancel plans and undo actions. Progress is saved, so nothing is lost halfway.

/07. Scale

Why one agent everywhere?

Researchers move between the library, studies, projects and reports all day. Adding AI to each area as a separate feature would have shipped sooner. But they'd have met a slightly different assistant in every place, and trust built in one wouldn't carry over to the next.

So Clio is one agent with one way of working, wherever it appears. Every request starts from the same actions, every choice uses the same picker, every change is a proposal to accept or discard, and AI content carries the same marker. Learn extraction, and you already know sharing and feedback. The team can add features without new patterns.

Outcome

Where it led

Clio is used across CiteMed's platform. Researchers hand it the repetitive parts of evidence work, from search and screening to extraction and drafting, and keep every decision that carries regulatory weight.

Most of the design happened before any screen existed: mapping the workflow, defining context, modelling how the agent behaves and setting autonomy by risk. The interface follows from those decisions.

Looking ahead

What comes next

Clio works on request, within a session, with clear checkpoints. That's the right start for regulated work. Three next steps are already in sight.

  • Memory: remembering a project's protocol, past decisions and the team's preferences across months, in a form people can see and correct.
  • Proactive help: Screening Assist already recommends without being asked. The same idea could flag a new study that fits a smart list, or a report section the evidence no longer supports.
  • Earned autonomy: as some kinds of output prove reliable, their checkpoints can relax, while staying firm wherever a mistake would reach a regulator.

That last one interests me most: agents that take on more as they prove reliable.

Tell me what you're building, and I'll ask a few questions to understand it.

← Back to Selected Works
Ask Elisey anything…AI chat

Go to

Selected WorksPage
Work ExperiencePage

Case studies

Clio AIAgentic AI assistant • 2026Case study
CiteMedAI-assisted clinical-evidence platform • 2026Case study
Health VisionAI-powered clinical intelligence platform • 2026Case study
CovisionEvent management SaaS • 2025Case study

Contact

Email Eliseyhello@eliseykolev.comCopy
Visit LinkedIn ProfileLink
Elisey Kolev
Ask Elisey TabAsk Elisey ↵