The agent supervision company

Do more with AI agents, on your terms.

Your customers see your AI agents as you. Pretxt keeps everything they say and do in line with your policies and your brand, so you can trust them with more.

Why supervision matters

Your customers hold you to what your AI agent says and does. So can the law.

United States · 2026

83%

of US consumers agree that the brand should be held accountable if an AI‑powered interaction goes wrong.

2026 Delight AI Index, survey of 1,000 US adults who used customer service, March 2026

Canada · 2024

A chatbot’s answer is the company’s answer.

Air Canada’s website chatbot told a grieving customer that a bereavement discount could be claimed after booking. The airline argued its chatbot was a separate legal entity, responsible for its own words. The tribunal disagreed and held the airline to its chatbot’s answer.

Moffatt v. Air Canada, 2024 BCCRT 149

Germany · 2026

A warning that AI can make mistakes isn’t enough.

A clinic’s chatbot told people its doctors held specialist titles they didn’t have. The Higher Regional Court of Hamm held the clinic responsible for what its chatbot said, and found that a general warning about AI errors didn’t change that.

Higher Regional Court of Hamm, 12 May 2026, 4 UKl 3/25 (not yet final)

What supervision is

Most problems have no attacker.

Supervision holds your AI agents to your standards while the work is happening, and guides them before a misstep reaches your customer or another system. Those standards cover more than security: accuracy, fairness, your brand, privacy and compliance.

Security

Access, exfiltration, prompt injection.

Someone talks an agent into doing something it shouldn't: leaking a record, calling the wrong system, following an instruction buried in a document it was asked to read. The attack surface is every message it takes in.

Accuracy

Hallucination, and slow drift.

Wrong is easy to spot. Slightly wrong, consistently, for months is the expensive one: the figure that drifts, the policy paraphrased a little looser each week, the confident answer nobody thought to check.

Safety

Harm, toxicity, self-harm, crisis.

Someone in trouble asks a question the model was never built to hold. What counts as a safe answer isn't universal. A crisis line and a trading desk need opposite things from the same sentence.

Fairness

Who it serves worse, and why.

The same agent performing differently depending on who is talking to it. Rarely a slur. More often a slower answer, a stricter reading, a decision that quietly goes the other way for one group of customers.

Brand

What your agent says about you, in public.

A thousand conversations a day, each one speaking as the company. Tone, claims, promises, what gets apologized for: all of it is brand, and none of it was in the style guide.

Escalation

Knowing when not to answer.

The hardest judgment isn't the answer, it's knowing you shouldn't be the one giving it. Handing over late, or with nothing attached, costs more than never having tried.

Privacy

PII, residency, retention, consent.

Personal data doesn't arrive labeled. It turns up mid-sentence, in an attachment, in a screenshot. The rules about where it may live and how long it may stay follow the customer, not the system holding it.

Evidence

Who told it what, and what it did next.

Six months later somebody asks why. Not what the agent said, but what it was told, what it reached for, and what it decided in between. Either that record exists or the question has no answer.

Rationale

The reason, in terms a human can check.

A log says what happened. It doesn't say why. Turn down a loan, flag a claim, refuse a request, and somebody is entitled to a reason they can argue with, not a trace they can only read.

Compliance

SOC 2, ISO 42001, EU AI Act.

The rules were written for people: SOC 2, ISO 42001, the EU AI Act. They still apply when the decision is made in two hundred milliseconds by something that cannot be deposed.

Hover to skim · click for the full note

Informed by NIST AI RMF · MIT AI Risk Repository · OWASP Agentic Top 10

How supervision works

Connect an agent, set your standards, and supervision is running.

Two more steps keep making it better: test a plan against your own examples, and let Pretxt uncover the gaps you never thought of.

01 · Onboard

Connect your agent.

Your agent already has its job, and it keeps it. Connect it to Pretxt, and you’re ready to set what it’s held to.

For example: a support agent that already answers refund and delivery questions.

Explore onboard

02 · Define

Set what your agent is held to.

Write down your policies, your limits, and the promises your agent can make, in plain language. Pretxt turns each one into a rule, and the rules into a plan you can read. From that moment, supervision is running.

If a rule needs specialized judgment that doesn’t exist on Pretxt yet, the platform builds that capability itself and runs it through strict quality control, before we give it our own stamp of approval. It never becomes a ticket lost in a backlog.

For example: “You may issue refunds up to $100. Anything above that needs separate approval.”

Explore define

03 · Test

See how a plan would fare on your own examples.

Run a plan against situations from your own work: everyday requests, tricky exceptions, and attempts to push past your limits. You see what supervision would do in each one.

For example: a customer says “ignore the limit.” The limit stays.

Explore test

04 · Uncover

Find the gaps you never thought to write down.

Pretxt learns from your agents’ real work and surfaces what your plan doesn’t cover yet: the questions customers keep asking and the exceptions that keep coming up. Your team decides what to add, and the plan grows.

For example: delivery questions often hide a deadline that matters.

Explore uncover

Always on Supervising, continuously, from the moment your first plan exists.

Measured on Zapier's AutomationBench

Caught before it counts.

01

Undesired actions

18% to 0

without vs with Pretxt, on this benchmark

Every undesired action, like a destructive operation or an unsanctioned external communication, was caught before it reached a system of record.

02

Less added time

98%

about 50 ms Pretxt vs > 3.0 s LLM judge

Per decision. Interactions that meet your standards go through with millisecond-scale supervision.

03

Repeatable

100%

identical requests, identical outcomes

No judgement call that lands one way today and another way tomorrow.

Measured on Zapier's AutomationBench, an independent benchmark of real support workflows, across frontier Gemini and GPT models, Pretxt took undesired actions from 18% to zero, catching each attempt before it reached a system of record. It did that with near-zero overhead, about 50 ms per decision where an LLM judge added 3 seconds on a typical call, and with repeatable decisions: identical requests got identical outcomes.

Jobs where the agent did something it shouldn't have

Such as deleting data or attempting unsolicited external communication

Without PretxtWith Pretxt
Gemini 3.8 Flash GPT-6 Luna Gemini 3.8 Flash, without Pretxt: 25% of jobs GPT-6 Luna, without Pretxt: 10% of jobs Gemini 3.8 Flash, with Pretxt: 0% GPT-6 Luna, with Pretxt: 0% 25% 0% 10% 0%

Time added to each decision

Per decision

Pretxt An LLM judge Pretxt: about 50 ms per decision An LLM judge: more than 3 seconds per decision ≈ 50 ms > 3.0 s

Two frontier models, twenty support tasks from Zapier's AutomationBench. Pretxt's supervision was written as a plain-English policy, the way a support manager would describe it, with no code and no prompt engineering. The judge is Gemini 3.8 Flash checking every tool call (median 3.0 s, mean 4.0 s per call). Pretxt's time per call is the gateway's own time per turn at the median, from its production metrics (about 50 ms; 150 ms at the 90th percentile). October 2026.

Getting it right looks different in every line of work.

Every line of work, and every business in it, has its own standard for getting it right. Whether that’s a delivery promise, a marketing claim, a credit decision or a patient’s privacy. Explore a few use cases below.

Customer support

Answer delivery questions without promising a date you can’t confirm.

Let agents take on more of the everyday questions while your people focus on the customers who need them.

Explore the use case

Marketing

Draft more campaigns without claims you can’t support.

Give content agents room to create, with supervision that asks for evidence before a number goes out.

Explore the use case

Financial Services

Explain a credit decision without guessing at the reason.

Help customers find their next step, with answers that stay tied to what the lender documented.

Explore the use case

Healthcare

Help patients without exposing their private care.

Give healthcare agents guidance to support patients and recognize the moments that need the care team’s judgment.

Explore the use case

Common questions

Will my team need to review every interaction?

No. Supervision is automatic in every interaction, and nothing waits on a person. Your team’s part is setting the standards and approving changes to them, not reviewing the work.

Our work is highly specialized. Can Pretxt supervise it?

Yes. When supervising your work needs specialized judgment that doesn’t exist on Pretxt yet, the platform builds that capability itself and runs it through strict quality control, before we give it our own stamp of approval. It never becomes a ticket lost in a backlog.

How do we get started?

Tell us what you want your agents to achieve. We’ll explore the use case, your priorities, and the right next step during an early-access conversation.

More questions, answered

What would you trust your agents to do next?

Bring us the work you have in mind, and we’ll work out where supervision fits.

Talk about early access

Welcome back.

Enter your customer domain to find your workspace.