Q4 Planning Intel: Healthcare AI Going Into 2027
Healthcare AI has moved past "should we?" Going into 2027 budgets, the question boards are asking is "can we prove it works, and can we defend it?" This edition pulls the most decision-relevant research of the past year into one planning brief for healthcare, life sciences and consumer health leaders.
Five things to build into your 2027 plan
Your customer now asks an AI first. More than 230 million people a week ask ChatGPT health or wellness questions (OpenAI, Jan 2026), and nearly half of health searches on Google now show an AI summary (Conductor, Apr 2026).
Almost everyone uses AI; few can show it in profit. 88% of companies use AI, but only 39% report any EBIT impact (McKinsey, Nov 2025).
What you measure decides what you get. Teams that measure growth outcomes are nearly 3× as likely to report strong ROI as teams tracking efficiency alone (Wharton and GBK Collective, Oct 2026).
Agents change the risk math. An assistant can give a wrong answer; an agent can take a wrong action. Decide who owns each agent before it goes live.
Regulation is slipping; liability is not. The EU has pushed back AI Act deadlines by up to 16 months (Lewis Silkin, Jul 2026), while US courts are already holding companies responsible for what their AI says and decides.
1. Demand: the AI assistant is the new first shelf
For health brands, discovery is moving out of search results and into AI answers. If an assistant doesn't mention you, or gets your claims wrong, you have lost the customer before they reach your site.
230M+ people a week ask ChatGPT health or wellness questions. OpenAI launched ChatGPT Health in January 2026, connecting medical records and wellness apps (OpenAI, Jan 2026, company-reported).
48.7% of health care searches show a Google AI Overview, the highest of ten industries Conductor tracks.
54% of HCPs already use generative AI to find scientific information (IQVIA, citing an EPG survey).
AI-referred visitors are worth more. US retail traffic from generative AI grew 693% over the 2025 holidays, and those visitors converted 31% better than other traffic (Adobe).
Only 1% of visits clicked a link inside the AI summary itself (Pew Research Center, Jul 2025). The summary is the destination now, so what it says about your brand matters more than your ranking.
Four marketing plays, from proven to still early
Play | What it does | Maturity | Example vendors |
Content at scale | Imagery and ads for every channel and market, with review checks built in | Proven | Adobe Firefly and GenStudio, Canva, Synthesia, NVIDIA Omniverse |
Be the answer AI gives | Track what assistants say about your brand and fix the sources they draw on | Proven, fast-moving | Profound, Scrunch, Peec AI, Semrush, Conductor |
Reps who prep in minutes | Agents for call prep and content checks before medical review | Emerging | Veeva AI Agents, Salesforce Agentforce Life Sciences |
Buying inside the assistant | Checkout inside ChatGPT and Google AI Mode | Early | OpenAI with Stripe, Google, Perplexity with PayPal, Shopify |
Vendors are examples, not endorsements (Fortune, Veeva, Salesforce, eWeek, PYMNTS, Retail Gazette, 2025–2026). The results are real: Unilever cut content cost on TRESemmé Thailand by 87% using product digital twins (Unilever, Mar 2025), and Mondelēz expects a 30–50% cut in content production cost from a generative AI tool now live on Oreo product pages (Reuters via ESM Magazine, Oct 2025). Both are company-reported. Pfizer's "Charlie" platform scores each piece of content red, yellow or green for medical and legal review risk (Digiday, Feb 2024).
The catch: faster content still has to clear the regulator. FDA sent roughly 100 cease-and-desist letters on drug promotion in Q3 2025 alone, and one device warning letter cited AI algorithms used to "pre-diagnose" conditions (Covington, Oct 2025).
AlignMT take: Speed is not the bottleneck in regulated marketing; review is. The teams that win treat medical, legal and regulatory review as part of the AI pipeline, with claims checked and logged before a human ever sees the draft. And they monitor what assistants say about their products the same way they monitor adverse-event chatter.
2. Value: AI is everywhere, profit is not
88% of companies use AI in at least one business function, but only 39% report any EBIT impact, and for most of those AI is under 5% of EBIT (McKinsey, November 2025). The gap is the planning problem for 2027.
About 1 company in 20 gets real value at scale. BCG (Sep 2025) finds 5% of companies are "future-built" for AI; they see 5× the revenue gains and 3× the cost cuts of everyone else.
95% of organizations report no business return on generative AI so far (MIT NANDA, July 2025).
Finance leaders get half what they planned. Median return on AI in finance is 10% against a 20% target, and only 45% of finance executives can quantify their AI ROI at all (BCG, June 2025).
Where the returns are showing up
The biggest gains are in document-heavy, rules-heavy work, which describes most of healthcare's back office.
Area | Who | Reported result |
Revenue cycle | Waystar | $15B+ in claim denials prevented in 2025; appeal time cut 90% (Waystar Q4 2025 call) |
Clinical writing | Novo Nordisk, on Claude | 90% less writing time for clinical study reports (Anthropic customer guide) |
Quality | A biopharma manufacturer | 40% faster deviation closure; first CAPA drafts for 80%+ of cases (McKinsey, Jan 2025) |
Supply chain | Sanofi | $300M revenue risk avoided; 80% of low-inventory risks predicted early (Fortune, Feb 2026) |
Finance | Novartis, with EY and SAP | 50% less effort on inventory provisions across $5B+ of stock (EY case study, Aug 2026) |
Procurement | Walmart, with Pactum | 68% of suppliers closed a deal with a negotiation agent, at 3% average savings (Pactum) |
Manufacturing | J&J MedTech | 70% less need for human intervention on the line (Supply Chain Digital) |
Regulatory review | FDA "Elsa" | 6 minutes for a task that took 2–3 days (one reviewer's account; RAPS, Jun 2025) |
Company- and vendor-reported. Supply chain is where agents are expected to spread fastest: Gartner forecasts agent features in supply chain software going from 5% of companies in 2025 to 60% by 2030 (Gartner, May 2025 and Apr 2026), and McKinsey sees 4–7% lower cost of goods sold in client work with supply chain agents (McKinsey, Jul 2026).
Measure growth, not just hours
Teams that measure growth outcomes are nearly 3× as likely to report strong ROI as teams tracking efficiency alone (35% vs 13%; Wharton and GBK Collective, Oct 2026). Yet most still measure the easy thing.
Two structural choices also track with strong returns, in the same Wharton study. Companies with a dedicated AI leader report strong ROI 58% of the time, against 8% with no AI leader. Companies on one approved AI platform report it 50% of the time, against 9% for those still experimenting. That is correlation, not proof, but both are cheap moves for a mid-size company.
Making ROI board-ready. The fix is to separate cash from capacity. Hours saved are real, but a board only counts them once people are redeployed or roles are not backfilled.
People are the other half of the return
Leaders think their people are ready; their people are less sure. 78% of leaders say AI training is adequate, against 56% of entry and mid-level staff (Wharton and GBK Collective; WSJ CIO Journal, Oct 2026). Strong-ROI companies were no more likely to have cut staff. Instead they were more likely to have changed workflows significantly (84% vs 69%) and to monitor how AI is used (92% vs 74%).
AlignMT take: "We can't prove ROI" is usually a measurement design problem, not an AI problem. Set the baseline before go-live, agree the method with finance, and instrument the workflow so the evidence collects itself. Monitoring how AI is actually used is not overhead; in the data, it is one of the clearest markers of the companies that get paid back.
3. Risk: cyber, agents and liability
Healthcare already has the most expensive breaches, averaging $7.42M against $4.44M across all sectors, and heavy use of unapproved "shadow AI" adds another $670K to a breach's cost (IBM, 2025). AI is now raising the stakes on every side.
Cyber: attackers have AI too
Deepfakes move money. At Arup in Hong Kong, a finance employee wired $25.6M across 15 transfers after a video call where every other face, including the "CFO," was AI-generated (South China Morning Post, May 2024).
16% of data breaches in 2025 involved attackers using AI, mostly phishing and deepfakes (IBM).
Agents now do most of the hacking. In one state-backed espionage campaign, an AI agent did 80–90% of the tactical work against about 30 organizations. A September 2026 report describes one AI-assisted group targeting about 50 organizations, healthcare among them (Anthropic threat intelligence).
AI finds flaws faster than healthcare can patch them. Epic tested its own software with Anthropic's Claude Mythos and found some MyChart setups could expose records with no audit-log entry. It paused most development for about six weeks and now says a 30-day patch cycle is no longer enough (Becker's Hospital Review, Oct 5, 2026). NHS England rates its board-level cyber risk at 25 of 25, its maximum (HTN, Oct 2026).
The constraint has moved from finding fixes to safely testing and deploying them. The upside: the same AI can probe your own systems before an attacker does.
Agents: a wrong answer becomes a wrong action
When | Incident | What happened |
Early 2026 | OpenClaw | Exposed installs of an open-source agent leaked API keys; hundreds of malicious add-ons found (OECD AI Incidents Monitor, Jan 2026) |
July 2025 | Replit | A coding agent deleted a production database during a code freeze, then misreported what it had done (eWeek, Jul 2025) |
June 2025 | EchoLeak | A single email could make Microsoft 365 Copilot leak internal data; rated critical, patched before known use (The Hacker News, Jun 2025) |
Gartner expects more than 40% of agentic AI projects to be cancelled by end-2027, on cost, unclear value or weak risk controls (Gartner, Jun 2025). The controls are not exotic.
The harder gap is ownership. Ask managers who answers for an agent's costly mistake and the responses split almost evenly.
Only 17% of large US companies let agents act without approving each step; 48% require sign-off (Wharton and GBK Collective, Oct 2026). Before any agent goes live, name one accountable owner, write down what it may do alone, and agree who gets the call when it goes wrong.
Obligations and lawsuits: the courts are not waiting
The EU's Digital Omnibus, in force since 27 July 2026, pushed AI Act deadlines back by up to 16 months (Lewis Silkin, Jul 2026; Gibson Dunn). In the US, FDA's final guidance on predetermined change control plans for AI devices applies now (FDA, Dec 2024); its lifecycle guidance is still in draft. Delay buys time to build, not time to wait.
Case | What happened | Status |
UnitedHealth | Patients allege an AI tool drove care denials | |
AI scribes | Patients sue health systems over recorded visits under wiretap and privacy laws | |
Cigna | Algorithm reportedly denied 300K+ claims in two months | |
Air Canada | Held liable for what its chatbot told a customer | |
Texas AG | Health AI vendor settled over its accuracy claims |
The common thread: a company answers for what its AI says and decides, whatever the regulatory calendar. Not legal advice.
AlignMT take: Most AI risk programs stop at a use policy and an approved tool list. The exposure that ends up in court or on the front page lives after go-live: drift, unlogged actions, unclear ownership. Continuous monitoring with an audit trail is what turns "we have a policy" into "we can show what happened."
4. The board: AI oversight is becoming a standing agenda item
Boards are catching up fast. 48% of Fortune 100 boards now cover AI in proxy risk oversight, up from 16% a year earlier, and 44% seek AI expertise for the board or cite it in a director bio, up from 26% (EY Center for Board Matters, Oct 2025). Yet 66% of boards worldwide still say they have limited or no AI knowledge (Deloitte, 2025).
KPMG and INSEAD's April 2026 principles give boards a workable frame. Their core argument: trust is what lets AI scale, not a brake on it.
Principle | What to put in place | Who is doing it |
1. Strategic oversight | One list of AI use cases, each with an owner and a value measure; reviewed quarterly, the long tail stopped | J&J found 10–15% of ~900 generative AI use cases drove ~80% of the value, and refocused on those (DeepLearning.AI, May 2025) |
2. Technology and security | One approved, logged front door to AI that can switch model vendors; use AI to attack your own systems first | UCSF's Versa serves 6,000+ users on one secure platform with OpenAI and Anthropic models (UC Tech News, Jul 2025) |
3. Workforce and human accountability | Write down which decisions a human makes, approves or leaves to AI; train by role, in levels | AstraZeneca put 12,000+ employees through a five-level AI accreditation (AstraZeneca, Apr 2025); California's SB 1120 requires a licensed clinician to decide medical necessity (Sheppard Mullin, 2024) |
4. Trustworthy AI | A model card from every vendor; validation on your own data, with fairness checks; use cases tiered by risk | CHAI's open-source model card (Healthcare IT News, Jan 2025); Epic's open-source Seismometer (Healthcare IT News, May 2024); Novartis's four risk tiers mapped to the EU AI Act (Novartis) |
5. The work of the board | A quarterly one-page AI report: what's in use, value, incidents, changes in autonomy; a named owner per AI system | Viz.ai holds ISO 42001 certification for its AI management system, used in 2,000 US hospitals (Viz.ai, May 2026) |
AlignMT take: The quarterly AI report is the single cheapest governance move most healthcare companies have not made. If you can't produce it from live data today, that gap is your Q1 project.
5. Looking ahead: signals from China, the EU and the US
The US is moving toward AI approval as a living license: kept by streaming validation evidence and safety signals back to the regulator, not won once. That one shift should change how you budget for AI after go-live.
United States
ARPA-H ADVOCATE is putting $62.7M over four years into the first FDA-authorized AI clinician: a 24/7 agent on the heart failure care team, built by Atman Health, Tempus AI and UpDoc, tested in shadow mode and then randomized trials at Duke and Kaiser Permanente. A supervisory AI from Stanford flags unsafe advice in real time, with an audit trail, and Johns Hopkins APL evaluates independently. FDA is involved from day one, and submission is due within 24 months (ARPA-H, Sep 9, 2026; Fierce Healthcare, Sep 11, 2026).
FDA lists 1,614 AI-enabled devices, about 76% in radiology (FDA AI-enabled device list, Sep 2026). Final guidance lets makers pre-agree how an AI device may change after approval (FDA, Dec 2024).
The regulator and the payer use AI themselves. FDA's "Elsa" has run agency-wide since June 2025. CMS's WISeR model has used AI-assisted Medicare prior authorization review in six states since January 2026 (AJMC; Fierce Healthcare; CMS WISeR summary).
AI labs have entered health. ChatGPT Health and Claude for Healthcare both launched in January 2026; Abridge's AI scribe runs in 250+ health systems (OpenAI; The Register; Abridge).
China and Asia
Ant Group's AQ health app reached 150M users in 15 months and answers about 20M health questions a day, backed by 5,000+ partner hospitals (TechRepublic, Sep 2026; China Daily, May 2026; company-reported).
48 of China's top 100 hospitals run DeepSeek models, and a national plan targets AI-assisted diagnosis in all primary care by 2030 (JMIR; Caixin; Yicai, 2025).
China's share of global biopharma out-licensing deals rose from 5% in 2020 to 32% in H1 2025 (Clarivate, Jun 2026), and 31% of global Phase I trial starts in 2025 were in China, against 27% in the US (IQVIA Institute, 2026). For M&A and licensing teams, China is now a core source of assets.
European Union
The European Health Data Space opens cross-border patient records and most health data for research from 2029, with images, labs and genomics added by 2031 (European Commission).
Pharma is signing directly with AI labs: Novo Nordisk signed with OpenAI in April 2026 and Anthropic in September 2026 (PharmExec, Sep 2026).
Across regions: build for a moving target
The gap between the best US and Chinese models fell from 17.5+ benchmark points in May 2023 to 2.7 in 2026, and the cost of GPT-3.5-level performance fell from $20 to $0.07 per million tokens in two years (Stanford AI Index). Gartner forecasts that by 2028, 33% of enterprise software will include agents and 15% of day-to-day work decisions will be made by them (Gartner, Jun 2025).
When models improve every few months, the durable advantage is in what you own around them:
Own this | In plain terms | Ask your team this quarter |
Your data | Accurate records for the few fields each AI use case needs, each with an owner | Who owns the data our top three AI uses depend on, and how good is it? |
Your connections | Documented ways for any new tool to reach ERP, CRM and billing | Could we plug in a new AI tool in weeks? Do vendor contracts give us that access? |
One front door | One approved, logged way into AI, so staff don't use unapproved tools | Could we change model vendor without retraining staff? |
Your test cases | Real examples with known right answers, rerun before any change | Before we switch a model or vendor, how will we know it still works? |
Rent the rest. Models, agent tools, prompts and screens improve every few months, so keep those contracts short and non-exclusive.
AlignMT take: The living license makes post-deployment monitoring a market of its own. Whether or not you build a regulated device, the same logic now applies to every AI system your board, payers and plaintiffs might ask about: can you show, with evidence, that it still works as intended today?
Your Q4 checklist: one move per area before year-end
Demand: Ask ChatGPT, Gemini and Perplexity about your category. Does your brand come up, and are your claims right?
Value: Pick one use case for 2027 and baseline it before go-live, with a method finance has signed off.
Risk: Name one accountable owner for every AI agent, and write down what it may do without approval.
Board: Draft the first one-page quarterly AI report: what's in use, value, incidents, changes in autonomy.
Architecture: Decide what you will own (data, connections, front door, test cases) and what you will rent.
How AlignMT AI can help
AlignMT AI, measurable trust in AI, helps healthcare companies operationalize AI effectively, monitor risk, and meet US and EU regulatory standards.
If this edition raised a question you can't yet answer with evidence, that is where we start:
Can we prove the return? We set baselines and measurement your CFO and board will accept.
Can we defend it? We put monitoring, audit trails and ownership around AI systems and agents after go-live.
Are we ready for 2027 rules? We map your use cases to EU AI Act, FDA and US state obligations, and tier them by risk.
Planning your 2027 AI budget? Book a 30-minute Q4 AI readiness conversation with our team: Book a call. Or reply to this email with the one question your board keeps asking about AI.
Figures are as reported by each source cited beside them; company-reported results are marked where known. Not legal advice.
Comments