How automatic approvals reduce benefits administration

Wellness receipts are the paper cuts of benefits administration: small, numerous, and almost always routine. Here is how our AI approves the clear-cut ones in seconds, why friskvård came first, and how it applies your company's own policy.

Between us we have spent years in HR roles, and we still see the same thing in the companies we work with that approve by hand today: benefits programs go to lose their shine in the receipt queue. An employee does the right thing, buys the gym card, uploads the receipt, and then waits. Not because anyone doubts the receipt, but because approving it is nobody's favorite task and it competes with everything else on HR's desk.

The irony is that almost none of these decisions need a human. A gym membership receipt with a clear provider, amount, and date is not a judgment call. It is pattern matching against rules that Skatteverket has already written. So we gave an AI exactly that part of the job, and only that part.

What actually happens when a receipt arrives

The employee uploads a receipt in the webapp, the same flow as always. Within seconds, the AI reads it, checks it against the rules for that benefit category, and makes one of three moves:

• The clear cases get an answer immediately. A recognizable gym membership, an amount inside the allowance, a receipt that shows what it needs to show: approved, with the reasons written down.
• Receipts that are clearly not eligible get a decline, and a reason. A pair of training tights bought at the gym counter is gear, not a membership. The employee is told exactly that, in plain language, together with the confidence figure behind the decision. Never a silent rejection.
• Everything else goes to a human. Blurry photos, gift cards, bundled purchases, unusual providers. The case lands in HR's normal queue exactly as it would have before, with the AI's notes attached.

The 85% confidence threshold is the design decision we care most about. The AI does not get to be almost sure before it approves something on its own. At 84% it does not approve, and the case flows to a person instead. Deliberately boring, deliberately conservative.

Why friskvård came first

Every benefits program in Sweden has one category that dominates the receipt queue: the wellness allowance. It is high volume, low amounts, and governed by unusually well-defined rules. That combination is exactly where automation earns its keep.

So the first ruleset is built on Skatteverket's friskvård guidelines. It knows that a gym membership qualifies and that the protein bars bought at the same gym do not. It knows a receipt needs an identifiable provider, a cost, and a date. It knows the difference between a subscription and a single visit. When something falls outside those lines, like a gift card for a massage studio, it does not guess. It hands the case to HR with its reasoning attached.

The rule engine is versioned and built per category, so new categories can be added without touching anything else. Wellness is where the volume is today; it will not be the last.

It reads the law, and your policy

This is the part we find genuinely exciting. Skatteverket's rules are the floor, but most companies have their own layer on top: which activities you have chosen to reimburse, caps below the legal maximum, documentation you require.

If you let it, the AI reads that too. HR can grant it access to the company's own uploaded policy documents. It reads the expense and wellness rules out of them into a short summary, keeps that summary alongside your company's other settings, and checks every later receipt against both the law and your policy. Upload a new version of the policy and the summary is rebuilt from it. If your policy says no to padel memberships, the AI says no to padel memberships, and writes down that it was your policy that decided.

The human always has the last word

Automation in something as personal as an employee's benefits only works if trust is built in from the start, so this is how it is built:

One thing worth being precise about, because it is usually the first question HR asks: an override corrects that case, not the model. Every revert is stored with the reason it was given and stays in the audit trail, but it does not quietly teach the AI a new precedent. If you want different answers next time, you change the policy it reads or the rules for that category, and the change is visible and deliberate. Nothing about how it decides drifts behind your back.

This is also what keeps the feature on the right side of GDPR: a human can always review and override, and the AI reads only what it needs, the receipt and the form.

What it adds up to

For a mid-size company, the wellness queue is hundreds of small decisions a year. Each one is a couple of minutes of reading plus the context switch that costs more than the minutes. Automating the routine majority means the genuinely ambiguous receipts get faster and better attention, and employees stop experiencing their benefits through the medium of waiting.

That is the actual product idea: not replacing HR's judgment, but reserving it for the cases that deserve it. It is a big part of why companies choose CLVR Benefits.

Data and privacy, in plain terms

Trusting an AI with receipts is only reasonable if you know exactly what it sees. So here it is, plainly. When a feature is on, we send only what that feature needs: the receipt image or PDF, the benefit category names it picks between, and, if you have enabled it, the company policy documents you uploaded. Nothing else.

We never send employee names, emails, or any other personal data. We never send company information beyond those category names, and we never send data from anyone else's expense reports or benefits. The AI sees the receipt and the form, and nothing around them.

The features run on Claude, from Anthropic. We send the minimum each one needs, we do not use your data to train any model, and we keep only what the feature and audit compliance require. That holds for your policy documents too: the AI reads them into a summary we store against your company and use as context when reviewing your receipts. It is not training data, and it never informs anyone else's decisions. The full details are in our privacy policy and Trust Center.

And it is never all-or-nothing. Each AI feature is its own switch, off by default, that HR turns on deliberately, and letting the AI read your policy documents is a separate, explicit choice on top. You decide which parts of this you want, one at a time, or none at all.

The short version

• Routine wellness receipts are approved in seconds, with written reasons. Everything ambiguous goes to HR, exactly as before.
• The AI only approves on its own above 85% confidence. Below that, the case goes to a person.
• Receipts that are clearly not eligible are declined with a written reason and the confidence figure, and the employee can correct and resubmit the same report.
• It works from Skatteverket's friskvård rules, and it can read your company's own policy documents and apply those too.
• An override corrects the case, not the model. To change how it decides, change the policy it reads.
• HR sees every decision, can revert any of them, and the whole feature is opt-in.