AI agents vs. analytics dashboards in value-based care.

I will say the quiet part first. Most value-based care organizations are not short on insight. They are short on hands. Their teams can already see the open care gaps, the slipping risk scores, and the members about to fall off a medication. They have dashboards for all of it. What they do not have is enough people to act on what the dashboards show, fast enough, before the quarter closes and the gap becomes a loss.

That is the gap between an analytics dashboard and an AI agent. One ends at the insight. The other ends at the closed gap. As you evaluate the wave of "AI for healthcare operations" landing in your inbox, that single distinction is the one worth holding onto.

What a dashboard does, and where it stops

A dashboard is a window onto your data. A good one is fast, accurate, and clear. It tells you that 1,400 care gaps are open across your panel, that a member's proportion of days covered (PDC) has dropped below the adherence threshold, that a provider's risk-adjustment factor (RAF) is trending below last year, that an ADT feed just flagged a hospital discharge. That is real value. You cannot act on what you cannot see.

But the dashboard stops at the boundary of the screen. The moment something needs to happen in the world, a person has to do it. A coordinator reads the alert, opens a second system, picks up the phone, navigates to the right payer portal, re-keys the data, waits on hold, leaves a voicemail, sets a reminder to follow up, and writes a note. The dashboard did not do any of that. It pointed at it.

This is why "we added analytics" rarely changes the outcome. The work that was always the bottleneck, the actual doing, is still entirely manual. You have made the to-do list more legible. You have not made the list shorter.

What an agent actually does

An AI agent takes the next action instead of describing it. Given the same open care gap, an agent pulls the relevant context, decides what to do, and then does it. Concretely, in value-based care operations that looks like:

• Closing gaps end to end. Not flagging that a gap exists, but completing the sequence that closes it: outreach, scheduling, confirmation, documentation.
• Making the calls. Outbound voice agents reach members and providers, leave structured voicemails, and capture what was said back into the record.
• Operating the systems people operate. Computer-use agents log into payer portals and EHRs and enter or extract data the way a person would, clicking through the same screens, instead of waiting for an integration that may never ship.
• Updating the record. Every action writes back to one canonical record, so the next agent or person sees current state, not a stale snapshot.
• Escalating only when needed. When a case calls for human judgment, the agent hands it to a named coordinator with the full context attached, rather than dumping a queue on someone's desk.

The difference is not cosmetic. A dashboard's output is a number a human then acts on. An agent's output is a completed action and an updated record. This is also the line the industry is now drawing for itself. Gartner predicts that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025, while also warning that over 40% of agentic AI projects will be canceled by the end of 2027, largely because what many vendors ship as an "agent" is really an assistant that still leaves the work to you.[1][2]

The math leadership gets wrong

Here is the calculation I see most often, and why it fails. A team of 8 coordinators already manages 100,000 members. Volume is climbing. The instinct is to hire: two more coordinators, maybe a quality analyst. The board approves the headcount. Six months later the backlog is the same size, because the population grew too, and the new hires were absorbed into the same manual workflow as everyone else.

Adding people scales the bottleneck linearly. It does not remove it. And the bottleneck is not a shortage of judgment. It is the enormous volume of work that should never have required a human in the first place: the third reminder call, the portal data-entry, the hold music, the chart packet assembled by hand. Prior authorization alone consumes an average of 13 hours of physician and staff time per week, and 40% of physicians employ staff dedicated exclusively to it, according to the American Medical Association's 2025 survey.[3]

The leverage is not more hands on the same work. It is removing the work that should not require a human, so the hands you have move to the cases that genuinely need them. At our flagship customer, a physician-led IPA in New York running risk on 175,000 patients, that shift recovered roughly 10 hours per user per week with 100% team adoption and no new headcount. Across Pelica deployments, outreach capacity has tripled per coordinator.

AI that answers questions vs. AI that does work

The word "AI" now covers two very different products, and conflating them is how organizations waste a year. The first answers questions about your data. The second does the work.

A chat or copilot feature bolted onto an analytics tool is the first kind. So, mostly, are general assistants like Microsoft Copilot and ChatGPT. They are genuinely useful: you can ask why a measure slipped, draft a provider email, or summarize a chart. But when the conversation ends, you are still the one who has to make the call, log into the portal, and update the record. The work did not move. It just got a better-informed starting point.

An execution agent is the second kind. It completes the task inside your workflows: places the call, enters the data in the portal, writes back to the record, follows up on the schedule, and escalates the exceptions. The test is simple. After the AI is done, did a job finish, or did you just get a better answer to act on later?

A short example: medication adherence

Take a member whose PDC is drifting toward the cliff that drops a contract from 4-star to a lower band. A dashboard flags the drop and waits. An agent monitors the same drop, calls the pharmacy to check the refill status, leaves a voicemail for the prescribing physician, waits, follows up automatically when there is no reply, and escalates to a live coordinator only when the case actually needs one. By the time the team looks, much of it is already resolved. We walk through that full sequence, step by step, in what an AI agent actually does when it closes a care gap.

What to ask a vendor

If you are evaluating "AI for healthcare operations," these questions separate execution from analytics-with-a-chatbox:

• After your AI runs, what has actually happened in the world? Name the action, not the insight.
• Can it place an outbound call and document the outcome, or does it hand a task to my staff?
• Can it operate a payer portal or EHR that has no API, the way a person would?
• Does every action write back to one record, or do my people re-key the result somewhere?
• How does it decide when to escalate to a human, and what context does the human receive?
• Is there a full audit trail of every action: what was read, decided, done, and when?
• How long to a live workflow? Weeks, or a multi-quarter data project?

If the honest answers are "we surface it," "your team does," and "we integrate eventually," you are buying another dashboard. That may be the right purchase. Just price it as analytics, and keep the headcount you were going to need to act on it.

The buyer is not asking for more AI. The buyer is asking for fewer open gaps at the end of the week, with the team they already have.

Sources and further reading

• [1] Gartner: 40% of enterprise apps will feature task-specific AI agents by 2026, up from less than 5% in 2025
• [2] Gartner: Over 40% of agentic AI projects will be canceled by the end of 2027
• [3] American Medical Association: 2025 prior authorization physician survey (13 hours per week; staff dedicated to PA)
• Shrank, Rogstad, Parekh: Waste in the US Health Care System (JAMA, 2019); administrative complexity estimated at $265.6B