pakt
Built and operated by BOB
No reviewer
Approvals run end to end inside limits you set once, not a queue that scales with volume
< 200ms
Decision time, so automation stays as fast as the agent's own tool calls
3 prompts
To add a payment rail, exchange or system of record we don't already support
Take the human out of the loop without taking on the risk
Today
Throughput capped by reviewer headcount
Agent proposes
Refund, payout, order, invoice
Review queue
Every action waits for a person
Human approves
Judgement, under time pressure
Action executes
Hours later, if at all
With pakt
Throughput capped by nothing
Agent proposes
Refund, payout, order, invoice
pakt evaluates
Limits, exposure and eligibility, in code, under 200ms
Within policy → executes
No person involved
Outside policy → refused
Escalated only if your rules say so
Evidence recorded
The rule and totals behind every decision
Wherever an agent can
already move money.
Function
What goes wrong
What Pakt enforces
Accounts payable
Invoices paid to a bank account changed by whoever emailed last, above PO value, or twice for the same document.
Three-way match against PO and receipt, duplicate-invoice detection, payee-change holds, approval tiers by amount.
Customer support
Refunds issued outside policy, twice on the same order, or well past what a customer is entitled to.
Order and window checks, no double refunds, per-customer annual and lifetime caps.
Paid media
Runaway spend overnight: a feedback loop puts six figures behind a campaign that converts nothing.
Daily and monthly ceilings per campaign, per account and in aggregate; automatic freeze on spend anomalies.
Disputes
Provisional credit granted on a claim narrative written by the person who benefits from it.
Credit ceilings per dispute and per customer, mandatory review thresholds, new-account holds.
Claims
Payouts above the policy limit, on uncovered perils, or on cover that had already lapsed.
Limit and coverage checks, deductible always applied, human adjuster above a set severity.
Trading desks
Agent judgement drifts from its strategy parameters as the context window grows: positions creep past mandate, concentration builds, and a drawdown runs with no floor under it.
Notional, leverage and concentration ceilings held outside the model, plus loss limits and kill-switch thresholds checked on every order regardless of how long the agent has been running.
Exchanges
Most of their AI product stops at advice. Users get a bot that suggests a trade and then asks them to go press the button, because nothing stops an autonomous agent from doing real damage to a real account.
Per-user mandates the exchange defines and the user sets: size, asset, leverage and loss limits enforced on every order, so the agent can trade for the user instead of advising them.
Deployment and control
Run Pakt wherever your security review requires.
Self-hosted. Runs entirely inside your perimeter, so no customer record leaves your infrastructure.
Hybrid. We evaluate policy; the credentials that authorize actions stay with you.
Fully managed. Fastest path to production, with enforcement isolated from our operators.
Fails closed
If Pakt is unreachable, nothing is authorized. An outage stops approvals; it never invents one.
Injection-proof by design
Enforcement reads facts from your systems of record, not text the claimant wrote. There is nothing in the prompt to attack.
Immutable records
Exports your auditors and clients can re-check without asking your engineers.
Escalation by exception
A person sees only the cases your policy says require one, not the volume the agent could have closed itself.
Pakt is built and operated by BOB