Author: Ashish Datta

HIPAA-compliant LLM access really is not that hard.

I came to this conclusion after going fairly far down the wrong path.

It started with Kimi K3 and GLM-5.2. These open-weight models are getting close enough to proprietary frontier models that they need to be taken seriously.

My first thought was that healthcare companies could run models like these themselves and keep patient data inside their own environment.

That is an appealing idea. No protected health information sent to a model API. No dependence on a third-party inference service. A company controls the hardware, the model weights, the network, and the logs.

Then I looked at the hardware.

These are not small models. Kimi K3 has 2.8 trillion parameters, and the vLLM project describes an eight-NVIDIA-B300 deployment as its easiest way to run the model. GLM-5.2 has 753 billion parameters; NVIDIA’s current quantized release targets Blackwell hardware and supports runtimes such as vLLM and SGLang.

The exact footprint depends on the precision, context length, concurrency, and serving stack. That is the point. An organization considering this path needs to budget for more than a GPU server: high-speed GPU interconnects, storage for weights and logs, redundant infrastructure, monitoring, patching, model serving, and someone who can operate all of it. For serious production traffic, that can become a large infrastructure project quickly.

That does not make local inference impossible. For the right workload, volume, and organization, it may be a reasonable choice.

But it makes “we should run our own frontier model because HIPAA” a much more expensive sentence.

The local-model path has more than one vendor

The next option is a GPU cloud. Some providers will sign a business associate agreement (BAA), which is a necessary part of handling protected health information with a service provider.

But the details matter.

A BAA may cover specific compute, storage, and logging services rather than everything the vendor offers. The organization still needs to understand where prompts, responses, embeddings, backups, traces, and support data go. “The GPU is covered” is not the same as “the whole system is covered.”

That is real diligence, but it is also not unique to open models.

Then there is the more obvious question: what about the clouds most healthcare companies already use?

AWS lists Amazon Bedrock as HIPAA eligible. Microsoft offers a BAA for in-scope Azure services, and its Azure AI Foundry documentation describes its HIPAA compliance offering. Google supports HIPAA workloads through Vertex AI.

So a healthcare company can already get managed access to capable models from AWS, Microsoft, or Google without buying a GPU rack or operating an open-weight model.

The model is available. The work is deciding what to do with it.

HIPAA is a system property

A BAA does not make an application safe by itself. Neither does a local GPU server.

The company still needs appropriate access controls, encryption, audit logging, retention rules, and an architecture that keeps PHI inside covered services. It needs to know which users can ask which questions, which records the system can retrieve for them, and whether sensitive data is leaking into an observability tool or a debugging log.

Those are not optional details around the LLM. They are the system.

A useful healthcare assistant also needs to be more than a chat box pointed at a model. An end user should see an authenticated application that retrieves only the records they are allowed to access, shows where an answer came from, and makes it easy to correct or escalate an uncertain result.

Take prior authorization. An LLM could help assemble information from a patient record into a draft packet. That may save a person time. But somebody still needs to decide what information is relevant, validate that the draft is correct, submit it through the appropriate channel, and handle the exception when the case does not fit the usual pattern.

The same is true for chart summarization, intake, coding support, patient-message triage, and internal policy search. The question is not just whether a model can produce a plausible answer. It is whether the workflow makes a person faster without making a mistake harder to catch.

Start with the workflow

I suspect plenty of healthcare companies are treating HIPAA as the blocker when the larger issue is that they have not identified a narrow enough problem worth solving.

“Give our staff an LLM” is not a workflow.

“Help the prior-authorization team find the relevant clinical history, draft a packet, and flag missing information for review” is much closer. It gives the team something to evaluate: time saved, completeness, error rate, review burden, and the cases where the system should stop and ask for help.

Once that workflow is clear, the infrastructure decision becomes more practical.

A managed model service may be the right answer if it fits the company’s covered environment and the team wants to focus on the application. A locally hosted open model may be the right answer when the organization has unusually strict control requirements, enough sustained volume to justify the infrastructure, or a reason to operate the model as a core capability.

Neither option removes the need for careful design. Both can be part of a HIPAA-compliant system. Both can also be used carelessly.

The LLM is not really the hard part.

The hard part is picking a useful workflow, connecting the right data, evaluating the output, and deciding where a human needs to remain involved.

What healthcare workflow would you be comfortable giving an LLM access to today?

A while ago, a brochure website was a meaningful software project. Somebody needed to lay out the pages, create navigation, make a contact form work, and get it all deployed. Website builders did not make a good website automatic, but they made that particular layer of work cheap enough that it stopped being the main thing most companies paid for.

Something similar is happening to CRUD applications.

A competent engineer with current tools can get surprisingly far, surprisingly quickly: a schema, basic APIs, forms, table views, search, permissions, validation, an admin screen, and a handful of ordinary integrations. LLMs help produce that code faster and make the usual implementation details less expensive to iterate on.

That is real progress. It is also easy to draw the wrong conclusion from it.

The fact that it is getting easier to build a system of record does not mean the business problem is solved. It means the database-shaped part of the problem is less scarce.

The valuable question was rarely just “where do we put the records?” It was “what should we do next, who needs to do it, and how do we know it worked?”

From recording work to improving work

CRUD is still necessary. Organizations need a place to record customers, jobs, invoices, inventory, cases, and the rest of the nouns that make up their work. They need people to be able to find and correct those records.

But a record is not an outcome.

A useful distinction is between a system of record and a system of action. The first stores what happened. The second helps decide what deserves attention, coordinates action across people and systems, and learns from the result.

System of record	System of action
Stores customers, jobs, invoices, and cases	Prioritizes work and moves it forward
Lets people enter, search, and update data	Coordinates people, systems, and exceptions
Reports what happened	Forecasts, recommends, and optimizes what to do next
Uses broadly reusable patterns	Encodes domain-specific constraints and tradeoffs

The CRUD layer is often part of a system of action. It is just not usually the part that makes the system valuable.

The floor is moving to workflow

Take a customer-success tool. The commodity version has accounts, contacts, renewal dates, health-score fields, notes, and tasks. That is useful, and it is also a familiar application shape.

The harder version combines product usage, unresolved support issues, contract terms, champion turnover, and outcomes from similar accounts to answer a more useful question: which accounts need attention this week, and what action is most likely to change the outcome?

Then it needs to make that action practical. Perhaps a support issue needs escalation, a CSM needs a meeting, sales needs to be involved before a renewal date, and the team needs a shared view of what happened next. The value is not a nicer account page. It is reducing the chance that an important customer falls through the cracks.

This pattern appears everywhere. Field-service software can store work orders, technicians, addresses, and status updates. The differentiated work is scheduling and re-scheduling against technician skills, promised windows, parts availability, geography, overtime rules, uncertain job duration, and emergency calls. A credible schedule at 8:00 AM is not enough if the system cannot respond when a job takes twice as long as expected at 10:30.

That is workflow orchestration: the messy part involving handoffs, timing, exceptions, policy, and people. It is not glamorous, but it is where a lot of operational software earns its keep.

Not everything valuable is an LLM

LLMs are part of this shift, but “CRUD to AI” is too narrow a description.

LLMs are particularly useful when a workflow begins with unstructured information: an email, a document, a call transcript, an image, or a request written in normal language. They can help extract information, classify incoming work, summarize context, or give a person a natural-language interface to a system.

Other valuable systems may have no LLM in the critical path at all. They may use a rules engine, a forecast, a statistical process-control chart, a constraint solver, a simulation, or a carefully constructed report. Many systems will combine several of these approaches.

The common thread is not the model. The software does more than preserve a record of work. It helps make a better decision, execute it, and learn from the outcome.

Analytics turns data into a question worth answering

Consider revenue operations. A CRM stores leads, opportunities, stages, activity, and quotas. The useful analysis is often above that layer: pipeline coverage by segment, conversion rates between stages, typical cycle times, and the difference between a healthy-looking pipeline and one that is unlikely to close in time.

Those views support real decisions. Is a territory short on coverage? Is a segment converting differently? Does the organization need more sales capacity, a different territory design, or a different target? The implementation might be straightforward cohort analysis or a forecast based on historical data. It does not need to be generative AI to be valuable.

Product analytics has a similar trap. A dashboard can show that activation or retention moved. An experiment, with a clear metric and a credible comparison group, helps answer whether a product change caused the movement. That difference matters when deciding what to ship to everyone.

The system of record supplies the events. The analytical layer makes them useful for a decision.

Optimization makes tradeoffs explicit

Some of the highest-value software is not about generating text or predicting a label. It is about choosing among competing, constrained options.

A logistics application might store shipments, vehicles, drivers, stops, service windows, and delivery status. The difficult work is assigning loads and planning routes while respecting vehicle capacity, driver-hours rules, pickup timing, delivery promises, and cost. This is an optimization problem. There may be no chat interface and no LLM involved.

Inventory is another familiar example. A basic app can show that stock is low. A more useful system estimates demand and lead-time uncertainty, accounts for storage limits and the differing cost of stockouts, and recommends what to order, from whom, and when. It makes the tradeoff visible instead of leaving a person to infer it from a table of quantities.

Workforce scheduling has the same shape. Employee records, certifications, availability, and shifts are CRUD. Building a workable schedule means balancing coverage, labor rules, preferences, fairness, qualifications, and overtime. The value is a schedule that an operation can actually run.

Operations research, forecasting, and constraint solving have been doing this work for a long time. Cheaper application development does not replace them. It makes it more feasible to spend effort on the part that changes the outcome.

Reliable execution is part of the product

A recommendation that cannot be acted on is just another dashboard.

Useful systems need to connect to the places where work happens, create or route the next task, explain why an item was prioritized, and handle cases that do not fit the normal path. They need audit trails where the decision matters. They need safe fallbacks and a clear way for a person to take over.

This is especially important when an LLM is involved. A model can help read an invoice, summarize a case, or classify an incoming request. It should not turn uncertainty into an invisible decision. The system needs confidence thresholds, validation, exception queues, permissions, and a way to correct mistakes. Those are not incidental implementation details. They are what make automation usable in a real operation.

The same is true for non-AI logic. A routing optimizer needs to expose the constraints it used. A forecast needs to show when its assumptions no longer resemble reality. An approval workflow needs a path for the unusual case. Dependability is not separate from the product; it is part of the value proposition.

Start with the bottleneck, not the screen

For builders, the practical implication is simple: start with the recurring decision or bottleneck.

Ask what people are repeatedly deciding, what information they have to assemble to decide it, which constraints they are balancing, and what happens after they make the call. Then work backward to the data, integrations, analysis, and interface required.

That approach may still produce a CRUD application. Most useful systems need records. But the record pages become infrastructure for a more specific outcome: fewer missed renewals, better route utilization, more disciplined purchasing, faster resolution, or a decision that used to require several people and a spreadsheet.

LLMs have made the CRUD shell cheaper to produce. That should be good news. It lets teams spend more of their attention on the work that has always been difficult: understanding an operation well enough to remove delays, make tradeoffs explicit, and reliably move work forward.

Claude, Codex, and the rest are useful debugging partners. They can suggest hypotheses quickly, explain unfamiliar systems, and keep an investigation moving when you are stuck.

They can also send you on a very convincing rabbit hole.

We ran into that with a client’s ASP.NET application. After it had been up for a while, the first request for a static JavaScript file could be very slow. Requests after that were fast. It was the kind of narrow, intermittent behavior that invites a long list of theories.

Claude’s initial diagnosis was that SSL certificate revocation checking was holding up the first request. That is a real thing worth knowing about, and it sounded plausible in the abstract. But the application was using a self-signed certificate. There was no certificate authority revocation check to perform. A small piece of domain knowledge ruled out a direction that otherwise could have consumed hours.

The problem was not that Claude mentioned certificate revocation. The problem would have been treating a confident, technically detailed answer as evidence.

Start with what the system is doing

Rather than follow the SSL theory, we tested the behavior we could observe. We read the assets directly from the filesystem and fetched the asset through the application. The pattern was consistent:

Fetch an asset and the first request is slow.
Fetch it again immediately and it is fast.
Edit the file, then fetch it again, and the next request is slow again.

That is a much more useful description of the issue than “static JavaScript is slow.” The expensive path was associated with first access to changed file content. It was not ordinary request handling, and it did not fit the TLS explanation.

The experiment strongly pointed to endpoint scanning. SentinelOne was running in that environment and was the most likely cause: changed content was likely being scanned on its first access, while the next read benefited from the result already being available. We did not treat that as a definitive vendor-level attribution, but it fit the observed behavior far better than revocation checking did.

Plausible is not proven

LLMs are especially good at producing plausible explanations. They have seen the vocabulary around a symptom, and they can connect it to a real mechanism. That is useful for generating a list of things to investigate.

But a diagnosis has to survive the details of the actual system:

Does the proposed mechanism exist in this deployment?
Does it explain the timing and repeatability of the symptom?
What inexpensive test could distinguish it from the other hypotheses?
What observation would prove it wrong?

A self-signed certificate was enough to make us stop and question the revocation theory. The cold-read, warm-read, and modified-file test gave us a better hypothesis to pursue. Neither step required an encyclopedic knowledge of every possible cause. They required knowing enough to check the assumptions and to design a small experiment.

Use the model as a partner, not an authority

Claude still helped with the investigation. The right use was not to ask it for the answer and implement the first response. It was to use it as a partner while we compared theories against the environment and the measurements.

A practical debugging loop looks like this:

State the observation precisely, including what changes between a slow request and a fast one.
Ask the model for competing hypotheses and a test that would separate each one.
Check its assumptions against the architecture, configuration, and operational environment.
Run the smallest useful experiment.
Feed the result back in and repeat.

This is also why domain expertise still matters when using LLMs. If you cannot tell whether an answer fits the system you are operating, confidence and detail are easy to mistake for correctness. Bring in someone who knows the domain, or slow down enough to validate the model’s premises before chasing its conclusion.

The model can make a good investigator faster. It cannot replace the judgment needed to decide whether a theory belongs in the investigation at all.

Everybody likes an API. It is the clean version of an integration: documented endpoints, structured responses, credentials meant for software, and hopefully somebody else’s problem when the implementation changes.

But plenty of useful systems do not have one. Or they have an API that covers part of the product but not the screen a team actually needs. The data is there, behind a normal login, and somebody is opening a browser and copying it into a spreadsheet.

That is a good place for browser automation.

We have been looking at a practical combination of Playwright and Codex for this kind of work. Playwright drives a real Chrome browser. Codex can work with a browser exposed over the Chrome DevTools Protocol (CDP), which makes it useful for exploring an application and helping build the automation. The important piece is that the login stays human: someone who is authorized to use the site signs in and completes MFA themselves. The automated job uses the resulting browser session; it does not try to get around the login.

A concrete example is recording live odds from sportsbooks such as FanDuel or BetMGM. If an organization is allowed to collect and use the data, that is a much better job for a computer than for someone watching pages and updating a sheet all day.

The browser is sometimes the integration

This is not an argument to scrape everything. If there is an official API that does the job, use it. It is almost always less fragile and easier to support.

But sometimes the browser is the only interface available to the user. A person can log in, look at a live market, and see the numbers, but there is no supported endpoint for getting the same information into an internal system. In that situation, the browser can be the boundary between the site and your workflow.

There are obvious limits. The account needs to be authorized, and the intended use needs to comply with the site’s terms, contracts, and applicable law. That matters especially for sportsbooks, where access and permitted use can vary by operator and jurisdiction. This is not a way to bypass MFA, CAPTCHAs, rate limits, or other controls.

Getting a logged-in browser session

Playwright is a browser automation framework. It can launch Chromium or Chrome, navigate pages, click buttons, fill forms, and read what the page renders.

The feature that makes authenticated automation workable is a persistent browser profile. Instead of starting fresh every time, Playwright launches Chrome with the same profile directory. That lets it retain browser state such as cookies and sessions when the site permits it.

The first run is simple: launch a visible browser, log in normally, handle the two-factor prompt, and make sure the page you need is available. After that, the job can reopen that same profile. It can run visibly while somebody is building or debugging it, then run headlessly when it is ready to collect data on a schedule.

import { chromium } from "playwright";

const userDataDir = "/secure/path/to/browser-profile";
const context = await chromium.launchPersistentContext(userDataDir, {
  channel: "chrome",
  headless: false, // first run: let the account owner log in
});

const page = await context.newPage();
await page.goto("https://example.com/login");

// The authorized user completes sign-in and MFA in this browser window.

That profile directory is sensitive. It may contain an active session, so treat it like a credential: keep it in approved storage, restrict who can access it, and do not put it in source control or logs. Also plan for it to expire. The site decides how long a session lasts. If the job finds a login page again, it should stop and let a person reauthenticate.

Asking Codex to do the tedious part

The page is usually the part that makes these projects annoying. Modern web applications load data after the page appears, change the DOM as markets update, and use selectors that are not obvious until you can inspect the live application.

This is where Codex helps. With the authenticated browser connected over CDP, you can ask it to inspect the page, find the market and selection elements, and build the Playwright script around what is actually there. It is much faster than guessing at selectors from a screenshot or trying to reverse engineer an undocumented backend.

A reasonable first prompt is something like:

Set up Playwright with a persistent Chrome profile and launch it visibly so I can log in. Once I have an authenticated session, use that session to inspect the live odds page and build a script that records the event, market, selection, displayed odds, and collection time.

Codex is helping with the implementation; it is not replacing the account holder. Keep that line clear. It should not be asked to find credentials, solve MFA, or work around controls the site has put in place.

A live-odds collector

Say the job is to keep an internal, timestamped record of odds for a set of games and markets.

First decide what an observation is. At a minimum, it will probably include the sportsbook, event, market, selection, displayed odds, the source page, and the time the value was seen. That last field matters: a number collected at 2:00 PM is not the same thing as a number collected five minutes later.

Next, use the visible browser session to get to the right market and let Codex inspect the rendered page. The job needs to know when the market is actually loaded, how a suspended or unavailable price is represented, and which labels or attributes are stable enough to use as selectors. Prefer user-facing labels where possible over a long chain of generated CSS classes that will disappear in the next redesign.

Once that is understood, the collector can launch the same persistent profile in headless mode, visit only the pages it needs, validate what it finds, and write normalized records to a database or queue.

const context = await chromium.launchPersistentContext(userDataDir, {
  channel: "chrome",
  headless: true,
});
const page = await context.newPage();

await page.goto(targetMarketUrl, { waitUntil: "domcontentloaded" });
await page.getByRole("heading", { name: /live odds/i }).waitFor();

const collectedAt = new Date().toISOString();
const observations = await page.locator("[data-market]").evaluateAll((markets) =>
  markets.map((market) => ({
    market: market.getAttribute("data-market"),
    text: market.textContent?.trim(),
    collectedAt,
  }))
);

The code above is deliberately generic. The real selectors should come from the target site and should be tested against its actual states. The useful outcome is not just a script that reads a page once. It is a small internal data source that dashboards, reports, or models can rely on without each one having to understand the sportsbook’s UI.

The unglamorous stuff is what makes it work

A browser job will change when the website changes. That is normal. The difference between a useful integration and a fragile script is how it behaves on a bad day.

Keep the browser profile locked down. Save where each observation came from and when it was collected. Alert when the job suddenly gets no results, sees a login screen, or returns far fewer records than normal. And give somebody a straightforward way to rerun the visible browser and refresh the session.

That is enough to turn a manual task into something dependable without pretending the website is an API.

For companies sitting on useful data behind logins, this is a practical option: a person handles the authorization once, Playwright keeps the browser state, and Codex speeds up the work of turning what is on the screen into structured data.

There has been a lot of discussion recently about companies using AI to build internal tools that replace SaaS licenses. That is interesting, but it misses a big category of software: the web apps you cannot replace.

Sometimes the constraint is technical. More often, it is not. An insurance company may require you to use its verification portal. A specialty vendor may only accept orders through a clunky ecommerce site. Or a marketplace may be where all of the demand for your product or service lives.

You can build a better internal tool, but you still have to use those sites.

The browser is the integration point

Chrome extensions have always been a way to change the experience of a site you do not control. An extension can read information from a page, add controls to it, and help guide a user through a workflow.

Historically, that was possible but often not practical. You needed to write and maintain custom code for every awkward workflow, and the payoff had to be large enough to justify it.

The latest AI models change that calculation. It is now much easier to build an extension that augments a legacy site, whether that means changing a workflow, extracting information from a page, or adding LLM capabilities directly where people are already working.

Instead of asking someone to copy information from one system into another, you can put the assistance in the browser tab where the work already happens.

A bike search on Facebook Marketplace

I recently had a good excuse to try this out. I was looking for a bike on Facebook Marketplace with a specific set of requirements. The hard part was not finding listings. It was reviewing the photos for each listing to determine whether a bike was actually a fit.

Doing that manually meant opening and reviewing dozens of listings every day. That is exactly the sort of repetitive visual task that an AI model can help with.

So I built a Chrome extension that uses OpenAI to review listing photos and flag the listings that match what I was looking for. Rather than replacing Facebook Marketplace, the extension improves the part of the Marketplace workflow that was taking the most time.

The result is not a fully autonomous bike buyer. It is a faster way to narrow down a large list of listings so I can spend my time looking at the promising ones.

Where this approach works

The Marketplace example is personal, but the pattern applies to business workflows too. Look for web-based processes where a person repeatedly has to review, classify, summarize, or move information before they can make a decision.

A Chrome extension can be a practical place to add help to:

an insurer’s required portal
a vendor ordering site
a marketplace your team depends on
an internal legacy application that is difficult to change

The goal is not necessarily to replace the site. It is to remove the tedious steps around it while keeping people in the workflow they already need to use.

See it in action

Check out the demo below to see the bike finder at work:

Watch the video on YouTube

Interested in building something similar for a workflow your team cannot avoid? Get in touch with Setfive.

HIPAA-Compliant LLM Access Is Not That Hard