Specialised agents on every request, from the moment it arrives to provisioned.
Real historical decisions a draft policy is replayed against before it can go live.
Of IGA market revenue is professional services - the tax you pay to make legacy software usable. Agents remove it.
Can an agent loosen a control you set. Enforcement only ever tightens.
The product
Every screen on this page is the real product, captured from a live deployment. This is a single access request mid-flight: the agents have already read the policy and built the approval chain, and each level on it carries the rule that put it there and the person it went to.
Why Pato
Every governance vendor now says "AI-powered". Look underneath and you find two camps - and neither answers the question your auditor will actually ask: what, mechanically, stops the AI from granting too much?
Camp one - the legacy suites
SailPoint and Saviynt implementations genuinely take 1.5–2.5 years: every access profile defined upfront, custom scripting, a connector project per application. Analysts put 57–70% of IGA market revenue in professional services - which says something uncomfortable about the software. Many deployments never leave phase one.
Camp two - the AI-native rivals
Fast to deploy, strong on integrations, genuinely automated. But the AI is contained by prompts and good intentions, not by mechanism. Ask them to show you the thing that prevents over-granting - then watch the slide change.
Our agents decide inside a mathematically enforced envelope - a pure, tighten-only function you can unit-test, replay against history, and hand to your auditor. And because nothing is modelled upfront, you're live in weeks, not years. The mechanism is the product.
Governed autonomy
Machine speed is worthless if you cannot stop the machine. Hard rules are enforced in code, after the agents have spoken, in a fixed order: deny beats escalate, escalate beats the agents' proposal, the proposal beats auto-approve. When data is missing, the request falls to a human - never through the cracks. The AI can make an outcome stricter. It cannot make it looser.
Some access is refused whatever the case for it. Matching requests stop dead, regardless of who asked or what the agents concluded.
Crown-jewel systems meet a person every time, however confident the agents are. Automation speeds the path to the approver; it never replaces them.
Sensitive scope raises the bar automatically - more approvers, more seniority - without anyone having to remember that it should.
Combinations that must never sit with one person are blocked at the moment of request, instead of being discovered by an auditor a year later.
Routine access clears in seconds - but only where the agents already agreed and nothing tightening has fired. A fast lane can never overrule a red light.
Compose new rule types in the console from a fixed set of enforcement building blocks. No release cycle to wait for, and nothing given up on the proof side.
Text to policy
Access rules usually end up in a consultant's spreadsheet and a rule engine two people understand. Here you write what you mean - and the platform turns it into something it can execute, test and prove.
Step 01
Type the rule the way you would say it to a colleague. "Production database access needs the data owner and the requester's line manager, unless there's an active incident." That is the whole authoring step.
Step 02
Replay it over your last thousand real decisions before it goes anywhere near production, and see exactly which outcomes it would have flipped. Contradictions with your existing rules are caught as you write.
Step 03
Activate, version, and roll back in one click if it reads worse than the last one. A policy change is an afternoon's work, not a quarter's programme with a statement of work attached.
Driven by agents
Most governance tools automate the approval click and hand everything else back to your service desk. Pato Identity puts a specialist on each stage - and they hand off to one another without a person in the middle.
The coordinator
Runs the floor. Dispatches work to the other agents, watches their health, retries whatever stalls, and keeps every request moving to a conclusion instead of quietly ageing in a queue.
The rulebook
Reads your plain-English policy and applies it before anything else runs. Clear refusals stop here. Clear approvals complete here. Only the genuinely ambiguous travels further.
The judge
Weighs the request against your policy and its context, then returns a decision, a confidence score and a written reason. Below the confidence you set, it defers to a human rather than guessing.
The dispatcher
Works out who genuinely needs to approve - line manager, system owner, risk - and whether they go one after another or all at once. No more requests sitting with the wrong person.
The chaser
Watches the clock. Approver on holiday, SLA about to breach, request untouched for a week - it escalates on your rules, rather than on somebody noticing three weeks later.
The messenger
Tells the right person at the right moment, and stops once the request is closed. Approvers get a decision to make, not a daily digest they have learned to ignore.
The hands
Carries the approval into the target system and actually grants the access - directory, application or database. The request does not end at "approved". It ends at done.
The translators
Resolve what your sentences point at: "the requester's line manager", "anyone in Finance", "the owner of that system". They turn the words into the real people and groups in your directory.
Each agent's instructions are plain English too - versioned in the console, testable before you activate them, reversible in one click. You never wait on a vendor release to change how an agent behaves.
The business case
Access requests are a tax every department pays in small change - a form here, a chase there, a spreadsheet at audit. This is where the hours and the money come back.
Compliance & audit
Every decision arrives with its reason, its confidence, the policy it was judged against and the rules that fired - in a tamper-evident log. When the EU AI Act asks how your access AI is overseen, the answer is architectural, not a slide. "Why did this person have that in March?" has a timestamp.
IT operations
Routine requests clear without ever reaching the service desk. The ones that need judgement arrive with context attached and the right approvers already assigned - no triage, no chasing, no working out who owns that system now.
Engineering
Developers wait as long as it takes to justify access, not as long as it takes to find an approver. Emergency access during an incident follows a policy you agreed in advance, rather than a message to whoever happens to be awake.
HR & business operations
A new starter's access follows from their role on day one. A mover loses what their old role gave them when the role changes. A leaver's access goes when the record does - automatically, and with proof it happened.
Finance
Legacy IGA runs $200K–$500K in year one - mostly consultants encoding policy you could have written as a sentence and modelling roles before a single request clears. Here policy is a sentence, connectors configure in days, and access patterns are learned from live decisions instead of defined upfront. The services tax disappears from the bill.
We would rather not quote you somebody else's ROI. Bring a month of real access tickets and we will show you which of them close without a human touching them - and which ones never should.
Deploy anywhere
Your cloud, your data centre, or an estate with no route to the internet at all. The platform ships as containers, and the model behind the agents can be one you host yourself - so no request, no identity and no policy has to leave your perimeter.
Containers on your own infrastructure. With a self-hosted model in place, a request can be judged end to end without a packet leaving this box.
Use them for the extra capability, or not at all. Personal data is stripped before any prompt is assembled, and the provider never sees your policy store, your directory or your audit log.
Connects to the systems you already run
Model providers - swapped without a deploy
Providers sit behind one abstraction with priority ordering and automatic fallback, so a self-hosted model can back a hosted one - or replace it entirely.
Questions
Short answers, in the order they usually come up. If yours is not here, write to us - a person replies.
Identity governance and administration in which AI agents make the access decision itself, rather than a rules engine with a language model bolted on to summarise it. It is also called agentic identity governance. In Pato Identity, eight specialised LLM agents read the policy, judge each request, identify the right approvers and provision the result. The legacy model runs quarterly certification campaigns; an AI-native one governs continuously, request by request.
Enforcement is deterministic code that runs after the agents have spoken, and it can only tighten their decision. The ladder is fixed: deny beats escalate, escalate beats the agents' proposal, and the agents' proposal beats auto-approve. An agent can escalate a request or add approvers; it can never remove a control you set. That guarantee lives in code, not in a prompt, so it cannot be argued out of the model.
Every identity type in the estate: employees and contractors, plus the non-human and machine identities most IGA suites treat as an afterthought - service accounts, API keys, certificates, cloud workloads and the AI agents now joining the workforce. They are governed through the same policy and the same approval path as people.
As sentences. You write "Each critical access right needs to be approved by three different levels" and that is the whole authoring step - no rule-engine configuration and no consultant. Before a draft policy goes live it is replayed against real historical decisions so you can see exactly which outcomes it would have changed, and contradictions with existing policy are caught as you write.
Yes. Pato Identity deploys into your cloud, your data centre or an air-gapped environment, and can run against a self-hosted language model, so no identity data or access decision has to leave your perimeter. It connects to the identity providers and directories already in place rather than replacing them.
Each request carries its own provenance: which policy matched, why each approval level exists, who was asked and why they were the right person, the routing confidence, and the timestamps. The audit answer is written at decision time and attached to the request, rather than reconstructed from logs months later.
Thirty days beside your current approval process. The agents judge every real request in parallel and touch nothing. At the end you get the divergence report: what would have cleared safely in seconds instead of days, what still needs a person, and what the audit trail looks like. Evidence, not a promise.
Rather just write to us? contact@pato-identity.ai