"How do you protect personal data?" sounds like one question. It's really four, because companies have settled on four very different strategies — and most of the confusion in this market comes from comparing tools that were never meant to solve the same problem. This is a map. It says, plainly, what each approach is good at, where it falls apart, and the one specific situation promptShield is built for.
We'll get the punchline out of the way: promptShield is not the bulk PII-removal provider. If you need to strip identifiers from millions of records flowing through a data pipeline, we are the wrong tool and we'll tell you so. We're built for the opposite shape of problem — a relatively small number of high-value documents, the kind where one missed name is a real incident, and where a human has to be able to put their name behind the result. Here's how the whole landscape lays out.
The four strategies companies actually use
| Strategy | What it is | Good at | Where it falls apart |
|---|---|---|---|
| Governance platform | Classifies and labels data at rest across the estate | Enterprise-wide inventory and policy | You need one workflow protected this quarter, not a multi-quarter rollout |
| Cloud redaction API | Programmatic detection, billed per call | Volume and automation | The documents cannot leave your control, and no human ever sees the result |
| Managed redaction | You send the files out; a vendor redacts and returns them | Work you have no staff to absorb | The transfer is itself the risk, and turnaround runs in days |
| Manual blackout | A person with a marker or an annotation tool | Absolute control and clear accountability | Past a handful of pages: no detection, no safety net, and misses are silent |
Why this market is so confusing
Notice that these four don't sit on a single line from 'worse' to 'better.' They're built for different problems. A governance platform and a manual blackout aren't competitors any more than a city water system competes with a kitchen tap. The mistake — the one that produces endless apples-to-oranges comparisons — is treating 'PII protection' as one market when it's at least four.
The axis that actually matters isn't volume on its own. It's consequence-per-document × the need for certainty. A pipeline scrubbing ten million log lines can tolerate a statistical recall rate — the cost of any single miss is tiny and the next layer of defense catches the rest. A law firm sending fifty discovery PDFs cannot. There, the cost of one leaked Social Security number is a sanction, a blown deposition, or a malpractice exposure. At that stakes level, "99.5% recall" is not a feature. It's a liability with a number attached.
The gap: high-value documents that a person must vouch for
Look back at the four strategies and you'll find a hole right in the middle. The governance platforms are too heavy. The cloud APIs and managed services both require your raw documents to leave your control, and the cloud APIs remove the human entirely. Manual blackout keeps the human and the data but has no detection and no safety net.
What's missing is a tool for the situation that's actually most common in professional practice: a relatively small library — call it under a thousand documents — of files that matter enormously, one at a time. The clinic sharing two hundred records with a researcher. The accountant preparing a handful of files for an external audit. The lawyer assembling a data room. The HR team releasing an investigation file. These aren't bulk problems. They're high-consequence, low-volume, zero-tolerance-for-error problems — and they're exactly where automation alone is most dangerous, because anonymization fails silently. When a name is missed, nothing breaks. The document just leaves the building with a real identity still in it, and you find out only when it's somewhere you can't recall it from.
That's the gap promptShield is built for. Not bulk. Not a platform. The last mile of high-stakes redaction, where a human has to be able to stand behind the result.
Why a human must fundamentally stay in the loop here
This is the part that's easy to mistake for a limitation, so let me put it as the principle it actually is: at this stakes level, automation should propose and a human must dispose.
It comes down to what responsibility means. To be responsible for something is to agree to pay the price if it goes wrong — and a machine has no stake to forfeit. You can't fine it, sue it, or revoke its license. So when a document leaves your office "anonymized," someone is vouching for that, and it can only be a person. But you cannot vouch for something you never looked at; that isn't responsibility, it's gambling with your own name. Vouching carries a minimum: you have to at least be able to see what's about to be redacted, and confirm it.
A fully automatic anonymizer asks you to trust, at volume and without looking, that it caught everything — and the asymmetry is brutal: ninety-nine correct redactions don't compensate for the one name that slipped through, because a single miss is the whole failure. That trade is fine when the consequence of a miss is small. It's unacceptable on a high-value file. This is why, for the documents we're built for, responsible anonymization has to be semi-automated — not because the automation is weak, but because the responsibility cannot be automated away. (We wrote a whole essay on this; see "A machine can't be responsible. You can.")
And here the low-volume shape is what makes the human gate possible, not just necessary. You can actually read every page of fifty documents before they leave the building. You cannot read every page of fifty thousand — which is precisely why bulk pipelines have to be statistical, and precisely why they're the wrong answer for files where you can't afford to be.
How promptShield stacks up
Within that gap, four things define how we compare to the alternatives — and they're deliberate consequences of the problem we chose.
It's offline. Fully. The detection pipeline — pattern matching, named-entity recognition, and an optional local model — runs entirely on your machine. No document, no fragment of one, ever leaves the device. This single fact removes the central objection to the cloud APIs and managed services at once: there is no third party, because there is no transfer. For documents that exist precisely so they don't travel, the right architecture is one where they don't.
It deploys in minutes, with no upfront dev adventure. There's no platform integration, no data-ingress plumbing, no professional-services engagement to schedule. It's a desktop application. You install it and you're redacting the same afternoon — the opposite of the multi-quarter governance rollout.
The cost is flat and perfectly predictable. No per-call, per-character, or per-page metering. A busy month costs exactly what a quiet one does. For a small practice, that predictability is often worth as much as the privacy: you can put the line in a budget and forget it.
The human stays in the loop by design. The pipeline proposes every detection as a reviewable region — it links recurring entities across pages and filters obvious noise, turning hours of manual scanning into minutes — but it never decides. You see exactly what's about to be redacted, you confirm, adjust, add, or remove, and only then is the tokenized result produced. The machine does everything it's good at; you do the one thing only a person can do — be accountable for the result.
| Governance platform | Cloud redaction API | Managed redaction | Manual blackout | promptShield | |
|---|---|---|---|---|---|
| Documents stay on your machine | No | No | No | Yes | Yes |
| Time to first redaction | Quarters | Days of dev work | Days per batch | Minutes | Minutes |
| Cost shape | Platform licence | Per call or page | Per document | Staff time | Flat |
| Automated detection | Yes | Yes | Yes | No | Yes |
| A person vouches for the result | No | No | No | Yes | Yes |
| Built for | The whole data estate | Millions of records | Overflow batches | A handful of pages | Under ~1,000 high-stakes files |
Who this is for — and who it isn't
It's for you if: you're a regulated professional or a small-to-mid team that personally handles sensitive client documents — law firms, clinics, accountants, financial advisors, HR — and you're accountable for what leaves your hands. Or you're a privacy or compliance lead who needs a fast, defensible answer for a specific high-stakes workflow without a platform procurement. Or you're inside a larger organization and need an air-gapped, offline option for one sensitive process — not an enterprise rollout.
It's not for you if: you need to scrub PII from millions of records in an automated pipeline with no human review. That's a real and important job — it's just a different one, and the cloud APIs and governance platforms are built for it. We'd rather tell you that than sell you the wrong shape of tool.
The honest summary is this. Most of the PII-protection market optimizes for scale, and at scale the human has to drop out. We optimize for the opposite: the documents where the human can't drop out, because the cost of a single miss is too high and someone has to be able to vouch for the result. If that's the situation in front of you, the volume is low enough to read every page — and that's exactly why it's the situation we built for.
Common questions
What are the four ways companies protect personal data?
Governance platforms that classify data at rest, cloud redaction APIs that process it in bulk, managed services that redact on your behalf, and a person with a marker. Each is built for a different shape of problem.
When is a cloud redaction API the wrong tool?
When the document itself is the sensitive thing. Sending it to a service to have its secrets removed means the secrets arrive at the service first, which is what you risk when the document itself is the sensitive thing.
Where does promptShield fit?
A small number of high-value documents that a professional needs to hand to an AI today, on their own machine, with the identities restored afterwards. That is the engine we wrote for exactly that job.