Highmark Health’s work with Google Cloud is a long-running payer-provider transformation—not just a chatbot launch. It links a shared data and cloud foundation to internal employee tools, claims operations and efforts to bring payer information into clinical workflows. The practical lesson for healthcare leaders is to fix data access and workflow design first, then expand AI from search and drafting toward tightly controlled actions.
The evidence supports operational use cases and company-reported adoption and value figures. It does not establish that generative AI has independently improved clinical outcomes, or provide a public audit of claims accuracy, denials or processing time.
What Highmark Health and Google Cloud are building
Highmark Health brings together an insurance business and a provider system, giving it a chance to connect payer and care-delivery workflows. The entities have distinct roles: Highmark Health is the parent organization; Highmark Inc. is its insurance-services division; Allegheny Health Network (AHN) is its care-provider system; and enGen is the health-technology and administrative-services business referenced in Highmark’s annual report. Google Cloud supplies cloud, data and AI technology. These are not interchangeable names for one company or product.
The relationship predates generative AI. Highmark described its Living Health Dynamic Platform as an effort to connect clinicians, care managers, pharmacists, customer-service representatives, devices and digital tools around a more coordinated health experience. Its Living Health account provides that strategic context.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
Generative AI is one layer in a broader program. Highmark’s 2025 annual report describes Sidekick as a secure, dedicated internal generative-AI platform. Separate efforts include claims-operation automation and the movement of payer information into provider workflows. They should not be collapsed into a single claim that one AI system processes claims and delivers better care.
How the work developed
| Date | Development | What it helps explain |
|---|---|---|
| December 17, 2020 | Highmark described its Living Health Dynamic Platform and its six-year collaboration with Google Cloud. | The partnership began as a broader data and care-transformation effort, before the current generative-AI wave. Highmark’s account |
| November 2023 | Google Cloud discussed Highmark’s early generative-AI exploration for internal productivity and information access. | Employee assistance was an early route into practical use. Google Cloud’s account |
| February 26, 2024 | Highmark announced an Epic and Google Cloud collaboration to bring payer-derived insights into provider workflows. | This connects claims and coverage information to care coordination, rather than treating AI as a standalone chat interface. Highmark’s announcement |
| April 22, 2025 | Google Cloud described Highmark AI use in claims operations. | The vendor said AI was being used to automate and streamline the claims lifecycle and help detect and prevent fraud; it did not publish detailed performance measures in that account. Google Cloud’s recap |
| June 27, 2025 | VentureBeat published a recap of a VentureBeat Transform 2025 panel with Google Cloud CTO Will Grannis and Highmark analytics executive Richard Clarke. | The six lessons and some adoption and legacy-integration figures below are panel-reported, not an independent audit. VentureBeat’s panel recap |
| August 12, 2025 | Highmark announced a separate enterprise AI collaboration with Abridge, including ambient documentation and real-time prior-authorization work. | Highmark’s AI activity extends beyond Google Cloud; Abridge is an adjacent clinical-workflow collaboration, not a substitute for the cloud and data platform. Highmark’s announcement |
| 2025 reporting | Highmark’s annual report described Sidekick and other AI tools as part of its enterprise transformation; Google Cloud later reported increased Sidekick activity and calculated value. | These later figures show reported scale, but not a complete independent assessment of business or clinical impact. Highmark’s annual report and Google Cloud’s account |
Which use cases are reported—and what remains unproven
Sidekick: an internal employee assistant
Highmark describes Sidekick as a secure internal generative-AI platform for employees. Reported potential tasks include finding internal guidance, summarizing information, drafting member communications, supporting claims-related questions and providing a common entry point to approved AI capabilities. Its existence and strategic role are corroborated in the annual report.
At the June 2025 panel, Highmark was described as having more than 14,000 of its 40,000-plus employees using internal generative-AI tools. This is a panel-reported adoption figure, not an independently audited count, and the recap does not establish that every user used Sidekick regularly or completed a measurable amount of work with it. Google Cloud later said Sidekick interactions grew from 1 million to more than 6 million prompts in just over a year. Prompts indicate activity, not productivity by themselves. VentureBeat’s report; Google Cloud’s account.
Provider credentialing and contract verification
The VentureBeat panel described employees who had previously searched multiple systems manually. In the reported workflow, AI aggregates information, checks requirements and produces an answer with citations and contextual recommendations. That is a more informative example than a generic chatbot: it combines retrieval and cross-system synthesis, while showing the evidence behind the output. The public recap does not supply a measured time saving or error rate.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Claims operations
Google Cloud says Highmark is using AI to automate and streamline claims processing and support fraud detection and prevention. The public description does not specify whether each task is document retrieval, summarization, routing, decision support or automated adjudication. Nor does it report denial-rate changes, average handling time, straight-through-processing rates, accuracy, dollars recovered or an independent return-on-investment audit. A claim that AI is helping with the claims lifecycle should not be expanded into a claim that it autonomously decides complex claims.
Payer information inside provider workflows
In its Epic collaboration announcement, Highmark said payer-derived information could include conditions and history, in- and out-of-network visits, benefits and available programs, claims, acute-event alerts, care-management information and coverage information relevant to referrals and scheduling. The stated aims include reducing administrative friction, helping people make better-informed decisions, avoiding surprise out-of-pocket costs and supporting care coordination. That announcement documents a plan and intended use; it does not demonstrate a causal improvement in clinical outcomes. Highmark’s Epic and Google Cloud announcement.
Rank #2
- Book: deep medicine: how artificial intelligence can make healthcare human again
- Language: english
- Binding: hardcover
Claims data can add context, but it is not the same as a complete, current clinical record. It can be delayed, reflect billing rather than the full clinical picture, or omit care outside the available data. Clinicians need to see what a data point represents and how current it is before relying on it.
Grounded search and summarization
Google Cloud’s healthcare AI materials describe tools including Vertex AI Search for Healthcare, Healthcare Data Engine, Healthcare APIs and medically tuned models such as MedLM. Google says its healthcare search offering can ground answers in organizational information and cite source material. Citations can help a user check an answer, but they do not guarantee that retrieval found the right document, that the document is current or that the model interpreted it correctly. Google Cloud’s product announcement.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesAgents: a pilot direction, not proof of autonomous operation
The 2025 panel described a progression from chat interfaces toward agents that coordinate models and may take actions through backend systems. Highmark was described as piloting workflow-specific agents. That is not evidence of broad, unsupervised production autonomy. “Agentic” describes a system designed to plan or carry out steps; the actual permissions, approvals and deployment scope determine what it can do.
The six lessons for healthcare AI programs
1. Treat legacy modernization as part of the AI work
A model cannot reliably answer from data it cannot access or distinguish. Healthcare organizations need to connect old and new systems, govern the resulting data, and preserve the source and meaning of each record. The panelists reported up to 90% workload replication while connecting legacy systems, including COBOL-based systems. That is a panel-reported engineering result: it is not 90% automation, accuracy, cost reduction or a general modernization benchmark. The recap does not clarify exactly what workload was replicated, so the figure is difficult to compare with other programs. VentureBeat’s recap.
Before adding a generative model, map the integration boundary: which system remains authoritative, how data moves, how identities and permissions are enforced, and how changes propagate. For clinical exchange, use appropriate interoperability standards such as FHIR where they fit the workflow. Test for stale, duplicate, conflicting and missing records. AI can expose a data-quality problem more quickly; it cannot make that problem disappear.
2. Use foundation models; build the workflow advantage
Most healthcare organizations do not need to train a general-purpose foundation model from scratch. Their differentiating work is more likely to be connecting proprietary data appropriately, designing the workflow, maintaining evaluation cases and policy rules, integrating claims and EHR systems, setting escalation paths and measuring results. A specialized or tuned model may still be justified when privacy, task performance, latency, cost or control warrants it; the point is to make that decision against a defined need rather than prestige.
Rank #3
3. Build a governed platform instead of a pile of pilots
A shared platform can centralize model access, approved connectors, prompt management, logging, evaluation, security policy, usage monitoring and incident response. It should not force every problem into one model: a large model may suit complex synthesis, a smaller one may be faster for routine work, and deterministic rules may be the safer choice for a clearly specified calculation. The panel described choosing among model types and deterministic systems for different tasks. VentureBeat’s recap.
Platform consolidation also has a cost: cloud and inference consumption, data engineering, integration, evaluation, human review, monitoring and vendor dependency. Estimate the whole operating model, not just model access.
4. Start with a workflow task, not a model name
Define the outcome and the friction point before choosing technology. A practical sequence is:
- Set a business or care-delivery outcome that can be measured.
- Map the workflow step that creates delay, rework or avoidable effort.
- Identify the authoritative source data, its owner and how current it must be.
- Decide whether the task requires retrieval, summarization, classification, prediction, drafting or an action.
- Set error tolerances, review requirements and escalation rules.
- Choose the simplest model, rules engine or combination that meets those requirements.
- Test representative cases, including conflicts, missing information and unusual cases.
- Place the tool in the user’s existing workflow and monitor quality, safety, adoption and cost.
This sequence makes the model an implementation choice rather than the objective.
Recommended Free Tools
5. Treat adoption as a product and change-management problem
Highmark’s reported employee reach suggests that internal tools can attract use at scale. The panelists attributed adoption to training, prompt libraries, feedback loops and showing employees how a tool helped with a specific task. Adoption counts still do not reveal whether work became faster or better. Measure whether people return voluntarily, which teams use the system, how often outputs are corrected, whether staff spend more time reviewing than they save, and whether sensitive information is being put into unapproved tools. Involve claims staff and clinicians in design; they can expose workflow mismatches that a technical pilot will miss.
6. Expand from information to action with explicit controls
Finding a policy is lower consequence than changing a payment decision or sending a coverage notice. A useful maturity path is:
- Search and retrieval: find relevant approved information.
- Summarization and drafting: prepare a summary or proposed text for review.
- Recommendations: surface a possible next step with cited evidence and uncertainty.
- Human-approved execution: prepare a system action that an authorized person approves.
- Limited autonomous execution: permit only narrowly defined, reversible actions with monitoring and escalation.
In healthcare, “agentic” should not be treated as a synonym for unsupervised. Require explicit authorization, audit trails, bounded permissions, reversible changes where possible and a clear human fallback before increasing autonomy.
Where to draw the line in claims automation
The risk rises as a system moves from helping an employee find information to materially affecting payment or coverage. A safe starting point is to assist staff while leaving consequential determinations with the authorized process and reviewer.
| Lower-risk assistance to consider first | Higher-consequence use requiring stronger controls |
|---|---|
| Find policy or contract language and show its source. | Interpret ambiguous coverage and make a determination without review. |
| Summarize a claim file while preserving links to source records. | Automatically deny or pay a complex claim. |
| Draft provider correspondence for staff approval. | Send a consequential notice without authorized review. |
| Identify apparently missing documentation for a person to verify. | Accuse a provider of fraud based on an unreviewed flag. |
| Route a case to the appropriate queue. | Change adjudication logic or override contractual rules. |
| Compare a submission against stated requirements. | Make a final clinical or coverage decision from incomplete context. |
Claims systems already encode contractual and regulatory rules. A generative model should not silently replace those rules. Common failure modes include misreading clinical documentation, applying a stale policy, treating similar cases inconsistently, generating a denial rationale that does not match the actual decision, producing false-positive fraud flags, and increasing workload through extra verification. A decision-support tool needs clear boundaries, evidence and an appeal-aware process; an automated decision needs a substantially stronger case and governance.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Governance, privacy and trust are workflow requirements
Highmark’s discussion of the Living Health platform says Highmark controls access to and use of customer information, and that Google Cloud is contractually restricted from using data for unrelated marketing. It also distinguishes Google Cloud infrastructure from consumer Google Search. Those are statements about the described arrangement, not blanket guarantees about every product, configuration or deployment. Highmark’s privacy discussion.
For a healthcare deployment, verify the actual contracts, configuration and operating controls. A checklist should cover:
- HIPAA business-associate arrangements and applicable state privacy requirements.
- Minimum-necessary access, role-based permissions and identity management.
- Encryption, audit logs, retention and deletion rules.
- How prompts and responses are logged, who can review them and how long they persist.
- Vendor and subcontractor access, training-data restrictions and breach response.
- De-identified or synthetic data for testing where appropriate.
- Whether generated output becomes part of a clinical, claims or legal record.
- Member, patient and workforce transparency about data use.
- Human review, correction, escalation and incident-handling procedures.
Grounding answers in organizational documents and showing citations is a useful control, not a guarantee. Evaluate citation completeness, whether the cited source actually supports the answer, source freshness, conflicting documents, missing data and the model’s ability to say “not found” rather than inventing coverage or policy. For clinical use, test for missed conditions, inappropriate recommendations and alert fatigue. Payer-derived data should be labeled and presented with its limits so clinicians do not mistake it for a complete chart.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
How to evaluate the business case
Google Cloud reported 74 active Highmark AI use cases and $27.9 million in calculated AI-enabled value during 2025. The figure is company-reported, and the public account does not provide enough methodology to reproduce the calculation independently. It should be treated as a signal of claimed portfolio scale and value, not as an audited return, a per-use-case result or proof of improved care. Google Cloud’s account.
Keep distinct scorecards; a rise in prompts cannot substitute for a change in claims quality or patient outcomes.
| Measurement area | Useful measures | What it can establish |
|---|---|---|
| Usage | Active and repeat users, use by department, prompts per user, abandonment. | Whether the tool is being tried and retained. |
| Productivity | Time per case, search and drafting time, cases per employee, rework and escalation. | Whether workflow effort changes, including review effort. |
| Quality | Error rate, citation accuracy, retrieval precision, override rate, correction or appeal rate, fraud-alert false positives. | Whether the output is reliable enough for its intended role. |
| Claims and member experience | Cycle time, denial and avoidable-denial rates, provider abrasion, member satisfaction. | Whether administrative outcomes change; definitions and comparison periods matter. |
| Care outcomes | Clinician administrative time, care-gap closure, readmissions or other defined outcomes. | Whether a care effect is plausible; causal attribution needs an appropriate study design. |
| Safety and governance | Privacy incidents, unsafe outputs, policy violations, bias measures, incident detection and remediation time. | Whether risk controls work in actual operation. |
Set a baseline before deployment, define denominators and comparison periods, and track whether people correct or override results. If a project claims clinical benefit, measure that outcome directly and account for other changes that could explain it. The public Highmark and Google Cloud material cited here does not provide peer-reviewed causal evidence that these AI deployments improved clinical outcomes.
What other healthcare organizations can copy
Highmark’s integrated payer-provider structure, data assets, engineering resources and long-running cloud relationship are unusual. A smaller organization can still borrow the implementation discipline without reproducing its entire platform.
Quick wins for a faster PC:
Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →- Select one bounded workflow. Prefer a high-volume, measurable administrative task with an identifiable user and manageable consequences.
- Map the current process. Record handoffs, systems, exception paths, delays and rework before adding AI.
- Name authoritative data. Assign owners, confirm permissions and determine freshness requirements.
- Set a baseline and acceptance tests. Include ordinary cases, edge cases, missing records and conflicting sources.
- Start with retrieval and citations. Make it easy for the user to inspect the source and report an error.
- Pilot with frontline users. Train users, gather corrections and measure time saved alongside review and rework.
- Keep consequential decisions with an accountable process. Add approval and auditability before allowing a tool to initiate system changes.
- Scale through shared governance. Reuse evaluation, access controls, logging and incident response rather than launching disconnected pilots.
Google Cloud is a plausible platform for an organization building a substantial healthcare data and AI foundation, but the Highmark example alone is not a vendor recommendation. Buyers should compare the fit with their existing cloud footprint, EHR and claims connectors, interoperability needs, model portability, grounding and audit tools, permission controls, workflow expertise and total operating cost. Organizations without integration capacity, governed data or a sufficiently valuable use case may gain value sooner from a focused documentation, search or claims workflow product than from trying to recreate Highmark’s broad program.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
