The base eight agents handle the spine of Intake-to-Pay, the work every finance team does. Custom agents handle the work only you do. The thing you've stopped asking vendors about because they always say “that's on the roadmap.”
We built it.
It took an afternoon. It saves nothing. It reduces no risk. It has never once appeared in any other business case.
It also plays beautifully at 4:47pm on the last Friday of the quarter, when forty-one approvals clear in ninety seconds and the AP team gets to go home.
Here's why we're telling you this. Not because we think you want Pachelbel piped through your ERP. Because most finance software has trained you to ask for very small things. You've learned to want what's on the roadmap. You've learned that “can it do X?” is a question with a disappointing answer, so you stopped asking.
We'd like you to start asking again.
If we'll build the Canon in D agent, imagine what we'll build when you ask for something that actually matters.
Every finance platform is built for the average finance team. Which is useful, because roughly 80% of what you do genuinely is average: invoices arrive, approvals route, POs match, payments go out. That's the spine. We built nine agents for it and they're good at it.
Then there's the other 20%.
The 20% isn't small. It's where the money leaks, the auditors linger, and your most experienced person spends their week. It's also, almost always, the part that never made it into a system, because it's too specific for a vendor to build and too important to ignore. So it lives in a spreadsheet. Or a person's head. Or a 40-tab workbook named FINAL_v6_USE_THIS_ONE.xlsx.
Put simply: a custom agent is one of ours, pointed somewhere new. Same reasoning engine. Same explainability layer, it tells you why it decided, in a sentence a human can read. Same simulation mode. Same principle of sitting above your ERP rather than replacing it.
The difference is what it's looking at. Our nine look at the standard spine. A custom agent looks at your thing, the deduction rule, the grant restriction, the 2019 acquisition that's still on its own chart of accounts.
The lowest form of automation notices a problem and hands it to you. A custom agent makes the call: approve, reject, escalate, hold, within the boundaries you set. Judgment is the product. Notification is what you already have.
Every decision comes with human-readable reasoning attached, visible to whoever needs it and permanently in the audit log. If a custom agent can't explain itself, we don't ship it. That's not a policy. That's the whole architecture.
Point it at your last twelve months. See exactly what it would have decided, transaction by transaction, against what actually happened. Tune it. Run it again. No custom agent goes live on a hunch and a demo.
A custom agent isn't a bolt-on in a corner. It receives handoffs and passes them along like any other agent, Intake can route to it, it can hand to Route or Commit, its outputs feed Signal. It joins the pipeline. It doesn't sit beside it.
A sample. Yours will be different, which is the entire point.
A distributor deducts $14,000 for a promotion. Is it real? Was it authorised? Does it match the trade agreement? Was it already claimed last quarter under a different name? The agent validates every deduction against the contract before anyone pays it, and it doesn't get tired in month eleven.
Restricted funds have rules, and the rules are not the same rules. This agent knows which cost centres a grant can touch, flags the ones it can't before the spend commits rather than during the audit, and produces the funder report as a by-product instead of a project.
Two hundred locations, one royalty formula, and a spreadsheet that three people are afraid of. The agent calculates, validates against reported sales, and flags variance, monthly, in minutes, with the working shown.
That entity you acquired in 2019 is still on its own chart of accounts and everyone's made peace with it. The agent hasn't. It reconciles across the mismatch, applies the mapping, and stops month-end close from being a hostage negotiation.
You earned the rebate. Did you collect it? The agent tracks tier thresholds against actual volume, tells you when you're $40k of spend from the next tier, and chases the ones you're owed, which, statistically, is most of them.
Plays Pachelbel's Canon in D when a batch of approvals clears. Provides zero measurable ROI. Absolutely worth it.
We just told you we'd build a musical agent, so this may seem inconsistent. It isn't. There's a difference between frivolous and wrong, and we only object to one of them.
If a decision can't be explained in a sentence a Controller can read and an auditor can accept, we won't ship it, regardless of how well it performs. Accuracy you can't defend isn't accuracy. It's a liability with good metrics.
If your requirement is “when amount > $500, route to Priya”, that's a rule. Your ERP does it. It doesn't need reasoning, and charging you for an agent to do it would be theatre. We'll tell you that, for free, on the call.
Some processes are broken because the policy underneath them is broken. Automating that just makes you wrong faster, at scale, with an audit trail. We'll usually spot it, and we'd rather say so than build it.
Restraint is a design principle. It's in every one of our agents, and it applies to us too.
Five steps. No RFP required, though we notice the irony.
You describe the thing. Usually badly at first, because it's lived in someone's head for years and has never needed words. That's normal. We ask until it's precise.
We write down what the agent must decide, what it needs to know, what it's allowed to do on its own, and where it must stop and ask. You approve it before anything is built.
The agent gets built on the same architecture as the nine, reasoning, explainability, guardrails, handoffs. Not a script. Not a prompt in a trenchcoat.
We run it against your last twelve months. You see every decision it would have made versus what actually happened. This is where it gets tuned, and where you find out whether your policy was ever right in the first place.
It goes live with a human in the loop on every decision. As confidence proves out, you widen the autonomy. You decide the pace. Never us.
Most custom agents go from conversation to simulation in under three weeks. The Canon in D one took an afternoon, but it also doesn't have to be right about anything.
If it involves a financial decision, a policy, and data we can reach, we can probably build it. The constraint is almost never technical, it's whether the decision can be defined precisely enough to defend. If you can explain the rule to a new hire, we can build the agent.
You do. It's your policy, your rules, your process. We don't take your competitive edge and resell it to the company down the road. The architecture is ours. The judgment is yours.
Extremely common, and honestly one of the more useful things about this process. Half of step 01 is discovering that three people apply the rule three different ways and nobody's noticed. You'll learn something about your own operation whether or not you build anything.
Conversation to simulation, usually under three weeks. Simulation to supervised live depends entirely on how fast you get comfortable — which is a decision we deliberately leave with you.
Custom agents get the same simulation mode as the nine. Change the logic, replay your history, see what changes, then ship it. Same-day changes are routine, and nothing goes live untested.
Yes. But we'd rather build the one that saves you $200k first, and then, as a treat, the Canon.
Bring the workaround. The rule that lives in one person's head. Thirty minutes, and we'll tell you honestly whether it should be an agent, including if the answer's no.
Canon in D requests accepted. Judged, but accepted.
EXPLORE MORE