A tool, a project, or the job done.
Three ways to put an agent in a role. Fourteen questions buyers actually ask about all three.
We answer yes on all 14. Buying and building each answer partly on 10 and no on 4. Open any row and hold us to it.
01What am I actually buying?A filled roleA toolA project
A managed agent in a defined role, supervised by an existing senior on your side. It works to a job description and a performance contract, which you cannot hand to a tool. We deploy it, govern it and operate it.
A general tool your team prompts or an app that serves one person. It is useful from the first day and the work still sits with whoever opens it.
A project your team owns outright. Total control over what it does and total responsibility for it.
02Who carries the outcome?We doYou doYours alone
We carry it, and the agent measures itself against its performance contract. That sits under SLAs, quarterly reviews and incident response included in the fee, with a senior on your side who can say stop.
You carry it, because a tool vendor stands behind its software rather than behind the work that software produces.
You carry it completely, which is the point of building. Anyone telling you otherwise is selling something.
03How soon is the job done?Live in 24 hoursMinutes to installMonths
Your agent is live within 24 hours. Onboarding then spends three weeks calibrating it against your existing data and your way of working, with us alongside the whole time.
The software is running within minutes, which is the fastest start of the three. It does not do the job on its own, so the clock to the work being done runs until your person picks it up and uses it.
Months before anything does the job, because the cost is counted in engineering time rather than in software licenses.
04Who teaches it our business?We doYou do, foreverYour engineers
We do, during three weeks of calibration against your existing data, and you get a clear method for working with it. It then keeps learning inside the role it fills.
You teach it yourself, every time you use it. Most tools start each conversation fresh and know only what you paste into them.
Your engineers teach it continuously, which is a job that never quite finishes. What it knows is exactly what somebody found the time to build into it.
05What if it gets it wrong?Cited to sourceVaries by vendorYou fix it
Every answer comes from your own data and cites its source, which greatly reduces hallucination. If a fact is not on file the agent says so instead of guessing. Incident response is part of what you pay for.
How it fails depends on the vendor. A general model with no grounding in your data is more prone to inventing an answer.
How it fails depends on your build. You own the detection, the fix and the test that stops it happening twice.
06Will it remember any of this?Yes, and finds it fastRarelyOnly what you built
The memory of your business is persistent and grows over time, because the agent stays in the role and runs against its own governed vault. Many tools have memory, so the difference is retrieval: the agent pulls the right piece of it when the work needs it.
Rarely in any depth, because memory tends to be general or tied to one person's account rather than to the role.
It knows only what you built and only for as long as somebody keeps maintaining it.
07Who decides what it can do?Tiered, A1 to A4Admin settingsYou build it
An A1 to A4 model defines exactly when an agent may act alone and when it must seek human approval. Every deployment ships under that standard.
Admin settings decide, rather than role limits. The permission model is the vendor's rather than yours and it is usually all or nothing.
You decide everything, which also means you build the controls your compliance team is going to ask for.
08What will my auditor say?Logged and citedVendor logsWhatever you built
Controlled and logged access, answers that cite their source and a 10-phase lifecycle with review gates, under a standard delivered in editable Word so your own compliance team can sign it off.
Whatever the vendor chooses to log, in the vendor's format, scoped to the vendor's system rather than to your process.
Exactly what you built and nothing more. Compliance tends to say no to system access without autonomy limits, an audit trail and a kill switch.
09Who keeps it working?We doThe vendorYour team, weekly
We do, because keeping the agent working is our job rather than a ticket competing with everything else on a sprint board.
The vendor updates the product on their own schedule. You usually find out what changed after it has changed.
Your team maintains it every week. Models move underneath a home-built agent and it degrades quietly if nobody owns it.
10What do I get for the money?The work itselfCapability, salary staysUpkeep, forever
A filled role for a fee you can predict: 50% of the median annual salary for the role, with a $50,000 floor and sub-agents from $15,000. The money buys the work rather than a tool plus the salary of somebody to use it. For a role paying below that floor, the floor is the number to check.
A tool subscription is the cheapest line item of the three and it stays cheap. The money buys capability rather than output, because the work still needs the person you are already paying.
Engineer salaries every year, for which you get a system you own and have to keep alive. A build is never free, because it is a salary you are already paying spent on maintenance instead of on the work.
11What if our expert leaves?The record staysTool stays, knowledge goesKnowledge walks out
The agent carries the workflow, the documentation and the operational memory, so the record of how the work is done stays in the role rather than in one person's head.
The software is entirely unaffected, which is a real advantage worth weighing. What it does not hold is how your process actually runs, so that knowledge still leaves with the person.
The knowledge of how it works walks out of the building with them, usually at the worst possible moment.
12Can I stop it fast?Sixty secondsNext renewalBuild your own switch
A 60-second agent disable standard with tested circuit breakers, so control is always one action away. On contract term, our fee is annual from $50,000 a year, so ending the agreement is a slower conversation than stopping the work.
Stopping mid-task means revoking access or logging out. Cancelling ends the next renewal rather than today's work. On contract term a monthly tool is the cleanest exit of the three.
You can stop it whenever you like, because it is yours, though the stop button is one more thing you have to build and test. Stopping the project does not stop somebody owning the code.
13Where do its skills come from?Built in-houseThe vendor decidesOften public code
Every skill is built by us from scratch and shared between our agents through our own gated skills store. Nothing is pulled from public code libraries, where attackers plant malicious code.
Skills come from the vendor and, in some tools, from add-ons anyone can publish. Checking what goes in is the vendor’s job, on the vendor’s terms.
Teams often start from open-source code to move faster. Every piece pulled in is code you now have to vet and keep safe.
14Does it work where my team works?Teams, Slack, WhatsApp, TelegramInside its own appWherever you connect it
It works inside Microsoft Teams, Slack, WhatsApp and Telegram, so your people reach it where they already talk.
Most tools live inside their own app or product suite, so your team goes to the tool rather than the tool coming to them.
It works anywhere you connect it, as long as your engineers build and maintain each connection.
Run the numbers before you talk to us.
Put in one salary to see the agent fee, the setup cost and what you save over five years.
Run your numbers