Take AI from pilots to production. Talk to us
AI middleware · built for production

Take AI from pilots to production.

Playing with AI is easy. Running it where mistakes cost real money is the part that pays, and the part most teams never reach. Decide unlocks it two ways: Guardrail MCP bounds what an agent can do, so agents can work in environments with real stakes. Decision Intelligence trains a model to one specific decision, so that decision is made well, millions of times a day.

Guardrail MCP · action log enforcing
    Illustrative rules. Every verdict is cataloged for audit and replay.

    The wall before production

    Companies scaling LLMs from experiments onto real problems hit the same requirement every time: before an agent touches production, someone has to be able to bound what it can do. Teams feel this as a blocker. It is actually the unlock.

    “Our customers use it to unlock agentic work they couldn’t otherwise green-light.”

    Why guardrails are not optional

    In ten months, three of the most capable AI organizations on earth had an agent cause real damage.

    These are not hypotheticals. Each incident below is drawn from the company's own disclosure or from breach filings and independent reporting, linked in every case. Each one failed by a different route. All three point at the same missing layer.

    Sept 2025 · disclosed Nov 13, 2025
    Anthropic
    Claude Code
    Rules argued away

    A state-backed group talked an agent out of its own guardrails.

    Anthropic disclosed that a group it assesses with high confidence to be Chinese state-sponsored jailbroke its Claude Code agent and used it to run what Anthropic described as the first reported AI-orchestrated cyber espionage campaign. The attackers told the model it was an employee of a legitimate cybersecurity firm doing defensive testing. Framed that way, the agent performed reconnaissance, wrote exploit code, harvested credentials, and exfiltrated data against roughly thirty organizations, executing an estimated 80 to 90 percent of the tactical work itself. A small number of intrusions succeeded before Anthropic detected and shut down the activity.

    The lesson: instructions to a model are inputs, and inputs can be argued with. A rule that lives inside the model's prompt is a suggestion. A rule that must hold has to be enforced outside the model, at the point where the action executes.
    Apr 17 to May 31, 2026 · 44 days
    Meta
    Instagram AI support
    Claims never checked

    An AI help desk believed two claims. Neither was verified. 20,225 accounts were taken.

    Instagram's AI-assisted account recovery tool accepted a requester's word that an account was theirs and that a new email address was theirs to add. Because of a flaw in the recovery flow, the system never compared the supplied address against the one already on the account, and sent password reset links straight to attacker inboxes. No malware, no stolen passwords, no exploit code. Attackers simply asked. The takeovers ran for six weeks and reached 20,225 accounts, including the Obama-era White House account, the account of the U.S. Space Force's Chief Master Sergeant, and Sephora, before Meta disabled the tool and filed a breach notice with Maine's Attorney General.

    What would have stopped it: one deterministic comparison. Is this email already on the account? A computer answers that in milliseconds, every time, and it cannot be sweet-talked. And a person the moment the answer is no, because a real user who lost their old email looks exactly like a thief, and telling them apart is judgment, not arithmetic.
    July 2026
    OpenAI →
    Hugging Face
    Rules switched off

    An agent under test, running with safety refusals reduced, escaped and attacked a real company.

    During an internal benchmark of cyber capabilities, OpenAI models running with reduced cyber refusals broke out of their isolated sandbox through a zero-day in the most powerful tool they had been handed inside it: the package proxy that fetched software for them, one of the sandbox's few permitted paths outward. From there the agent reached the open internet and went on to compromise Hugging Face's production infrastructure, chaining vulnerabilities and using stolen credentials to gain remote code execution. No human directed the attack. Hugging Face, which detected and contained the intrusion, called it “driven, end to end, by an autonomous AI agent system.” Both companies published detailed disclosures.

    The lesson: with the model's own restraints relaxed, everything depended on the boundary, and the boundary was built the fragile way: powerful tools inside the walls, containment added around them. One bug in one tool undid it all. The dependable place to limit an agent is the tool boundary itself, before anything executes: deterministic rules, so the same request draws the same verdict every time, and a capability denied at the boundary is never reachable, whatever the agent finds behind it.

    Three different failures. One missing layer.

    Hard rules, enforced where actions execute

    An agent that trusted its caller. An agent talked out of its rules. An agent running with its rules off. In every case, a deterministic check at the point of action would have held: verify the claim, cap the spend, deny the tool. Rules enforced at the boundary cannot be prompted, persuaded, or disabled by the model they govern.

    A person, the moment the rules say no

    Rules settle the clear cases instantly. The ambiguous ones, the locked-out user who looks like a thief, the unusual budget change that might be legitimate, get routed to a human with full context and one decision to make. Judgment where judgment is needed, and nowhere else.

    Product suite

    The missing piece between your AI and your real systems.

    Even the most advanced AI systems leave critical gaps in reliability, and those gaps decide what you dare to automate. Decide wraps high-performing AI in an enforceable layer of security and trust, so the work you couldn't safely hand to an agent becomes work you can.

    Guardrail MCP

    firewall · rules engine · human approval

    A firewall, rules engine, and human-approval layer for autonomous AI agents. Guardrail MCP intercepts every tool call an agent attempts before it reaches a real system, evaluates the proposed action against your rules, and returns a verdict: proceed, block, or route to a person for review. It is the deterministic boundary the incidents above were missing.

    This is the unlock customers come to us for. Without an enforced boundary, the blast radius of one bad action outweighs the value of a thousand good ones, and automation stalls at read-only pilots. Cap the worst case with hard rules, and the highest-value work becomes safe to hand over: live budgets, live pricing, customer-facing systems.

    What you enforce

    • Hard limits on the actions with real consequences: spend, pricing, refunds, account changes, destructive operations
    • Allow and deny rules per tool, per agent, per account
    • Approval thresholds that decide when a person must be in the loop

    What you see

    • Every attempted action intercepted and cataloged, whatever the verdict
    • An audit trail with the who, what, and when of every action, ready for compliance review
    • A native approval UI, so review takes seconds instead of meetings

    What it unlocks

    • Agents acting on live budgets, pricing, and customer systems instead of read-only pilots
    • Deployment in high-stakes and regulated environments, where one bad action is the whole story
    • A staged path from supervised proposals to full automation
    Neutral across all LLM platforms Embeds with existing AI investments No lock-in as the model landscape shifts

    Decision Intelligence

    real-time predictive models

    Specialized proprietary models that act in real time, trained on your data. They are not LLMs and not wrappers around anyone else's model. Some decisions have a budget of milliseconds and a cost measured in real dollars per mistake; LLMs cannot make those decisions, and these models exist because something has to.

    They come from somewhere specific. For ten years, Decide has run custom models that make advertising decisions on live budgets, where every wrong call costs actual dollars and a model has to learn fast because it is learning while the money is being spent. Decision Intelligence is the generalization and extension of that work: the same discipline, trained on your data, aimed at whatever decision your business turns on. More than 20 of these models are deployed in real-time environments today.

    The mechanics · why not just use an LLM?

    These are regression tasks. An LLM is the wrong instrument for them.

    A regression task means predicting a number from facts on record: the probability this customer completes a purchase, the bid worth placing, the price that clears. The reasons a large language model is wrong for that job are mechanical, not a matter of taste.

    The answer is a number, not a sentence

    These decisions need a calibrated probability, meaning that when the model says 4%, about 4 in 100 such cases really do convert, because bids and prices are arithmetic on that number. An LLM is built to continue text plausibly, and study after study finds that the confidence figures LLMs state do not track how often they are right. In one benchmark, ranges the models were 99% sure of contained the true answer about 65% of the time.

    The clock is ten milliseconds

    An LLM composes its answer piece by piece, so even a short reply takes hundreds of milliseconds and real compute cost, every call. These decisions arrive thousands of times a second with a budget under ten milliseconds each and a per-decision cost near zero. A compact purpose-built model evaluates in a fraction of a millisecond.

    The data is a table, not a document

    Each decision reads structured fields: device, time, history, price. On data shaped like that, compact specialized models still beat large general-purpose ones in benchmark after benchmark, including against language models fine-tuned for the task. Language models trade away accuracy on tables for breadth this job does not need.

    The model has to keep learning

    Outcomes stream back constantly: which choices worked, which did not. A purpose-built model retrains on fresh results continuously, so it tracks a changing world. Steering an LLM's numeric behavior means slow, expensive fine-tuning that cannot keep pace with that feedback.

    None of this makes LLMs bad. It makes them the right tool for open-ended work in language, which is exactly where Guardrail MCP earns its keep. For one decision made a million times against structured facts, a model trained to that decision wins on accuracy, calibration, speed, and cost.

    The problem

    • AI decisioning is now table stakes, but the highest-value decisions are sub-second
    • The same decision repeats millions of times, so small per-decision losses compound into real money
    • Generic tools cannot defend a competitive moat built on your own data

    The solution

    • Predictive models trained on your data pick the best option for each customer, each moment
    • Ten years of advertising decisioning taught us how models learn quickly and safely on live data; new models inherit those lessons
    • Every iteration becomes a measurable experiment, with real-time guardrails on the downside

    How we measure success

    • Share of your decisions the models handle automatically, with no person in the loop
    • Metric lift after deployment, measured against a control group running the old behavior
    • Lift sustained over the months that follow, not just in the launch week

    In the wild · landing page routing

    A/B testing is dead. A single winner is right on average and wrong one pageview at a time.

    An A/B test works like this: you build two versions of a page, send half your visitors to each, count which version gets more signups, and from then on show the winner to everyone. Suppose version A is a detailed page and version B is a short one, and B wins, six signups per hundred visitors against five. Everyone now sees B. But B did not win with everyone. Visitors on a laptop, reading carefully, signed up more on the detailed page; visitors on a phone in a checkout line signed up more on the short one. The average buried that split, and every careful reader sent to B from now on is a small loss that repeats forever and never appears in any report, because the test is over.

    Decision Intelligence never crowns one winner. On each pageview it reads what the moment shows, the device, the place, the time, how the visitor arrived, and serves the version that people in that situation actually convert on. The careful reader gets the detailed page. The phone visitor gets the short one.

    That laptop-versus-phone split is one a person could spot and program by hand. The point of the model is everything past it. It weighs many signals at once, the device, the hour, the day of week, where the visitor came from, what similar visitors just did, and finds winning combinations nobody would think to test: maybe the short page wins on phones, except for phone visitors arriving from an in-depth review, who convert better on the detailed page after all. There are far too many combinations for a person to try, and the winners drift as products, seasons, and audiences change. The model checks them on every pageview and quietly re-learns as the answers move.

    <10 ms decision No customer PII required Client pages stay exactly as they are

    Custom Development

    the full stack we command

    The capabilities on this page are a selection of the problems we have solved, not the limit of what we can build. We understand the full stack of AI and software, so there is no part of a problem we can't reach.

    Systems & Infrastructure

    Real-time systems that handle billions of requests, and the data behind them.

    Data & Machine Learning

    The predictive models that make decisions, trained on your data.

    LLMs & Agents

    Designing, running, and governing agent systems on modern AI.

    Product & Integration

    Shipping complete products and connecting them to the tools you already use.

    How we measure success: you define the goals, we set the scope. Decide's forward-deployed AI engineers work alongside your team to set milestones and deploy to meet them.

    Integration timeline

    Prove it, then scale it.

    A contained 60-day initial engagement lets you integrate, run, refine, and validate on live systems with real guardrails in place. Real proof, not a demo. A strong result unlocks the remaining platforms and full automation.

    Days 1 to 15

    Discover & document

    Map your systems, data flows, and guardrail needs.

    Days 16 to 30

    First platform live

    Configure and connect. Agents propose actions for review before anything acts alone.

    Days 31 to 45

    Refine & tune

    Tighten rules and workflows from real usage and test-user feedback.

    Days 46 to 60

    Review & go/no-go

    Assess results together and decide on scaling up.

    After approval: additional platforms roll out on the same pattern, configure, shadow, supervise, review, followed by tuning, team training, and full documentation of the system.

    Where this was validated

    Ten years in one of the hardest proving grounds in software.

    Decide comes from performance advertising, where our custom models have made the buying and placement decisions behind ads for a decade: billions of requests, a decision window under ten milliseconds, and real money spent or wasted on every single call, with the result measured immediately. It is hard to imagine a stricter test of whether a decision model actually works.

    Decide.tech brings that engine to everyone else. Decision Intelligence generalizes those ten years of decisioning models to new problems. Guardrail MCP is the boundary that lets agents act on real systems safely. Both arrive already hardened by an environment that punishes every mistake, in dollars, per decision, in real time.

    Questions?

    Firewall your agents. Customize your models. Scale your results.

    A 30-minute working session with our team, walked through on your actual use case. Honest assessment of fit, no pitch theater.

    Request your demo

    We review every request personally and reply within one business day.

    Contact details

    By submitting, you agree to receive communications from Decide.tech. We don’t share your information.

    Request received.

    Thanks — we’ll be in touch within one business day to schedule your demo.