Back-office agents
Agents for your back office
Agents that clear the routine work — approvals, credit, collections, reorders, and more. Each one has written limits. Inside them, it acts. Outside them, a person decides.
Ten agents, each with one job
Live in the admin consoleEach agent follows written rules, not guesswork. Below is what each one may do on its own today — everything else waits for a person.
Purchase-order approvals
Reviews pending purchase orders.
Approves orders up to $20,000. Larger ones go to a person.
Credit underwriting
Reviews trade-credit requests.
Opens credit lines up to $50,000 for a credit score of 680 or more. Everything else is reviewed.
Collections
Follows up on overdue accounts.
Sends a payment reminder on low-risk balances up to $2,500. Larger or riskier accounts go to a person.
Replenishment
Watches stock against reorder points.
Drafts supplier POs when up to 25 items run low. A bigger shortfall — a possible disruption — is reviewed.
Cash application
Matches incoming payments to open invoices.
Applies a payment when there is one confident match, within 1%. Short-pays and unclear matches go to a person.
Quoting
Turns requests for quote into quotes.
Sends a quote up to $25,000 when every line matches your catalog and margin is at least 25%. The rest are reviewed.
Order exceptions
Clears stuck orders.
Retries a failed payment or a stuck fulfillment on orders up to $1,000. Address and oversell problems always go to a person.
Sourcing
Compares supplier bids on price and lead time.
Awards the bid only when it wins by 5% or more. Near-ties go to a person.
Marketplace pricing
Adjusts listing prices for overstock and scarce items.
Changes a price by up to 10%. Deeper markdowns go to a person.
Seller activation
Nudges marketplace sellers who are stuck in sign-up.
Sends low-risk reminders on its own.
Build your own watch-list agent
In Agent Studio, pick a data source and write a rule — for example, overdue accounts over a set amount. Try it in practice mode, then publish it. It flags matching records for your team to review.
How they stay safe
You can see what every agent did, why, and who approved it.
A person approves anything bigger
Each agent has written limits. Inside them it acts; outside them it puts the item in an approval queue, and a person approves or rejects it.
Try it before it acts
Run any agent in practice mode first. It shows what it would do — and changes nothing.
Only the permission it needs
Each agent can use one tool for its one job. It goes through the same permission check as every other caller, and it can never authorize spending.
A full record
Every decision is logged with the rule it followed, and every approval with who made it.
A dashboard you can watch
See how many items each agent handled, how many it acted on, how many wait for a person, and its success rate.
Technical detail
- Bounded authority: per-agent guardrail policies; candidates outside a policy are queued (pending_approval) rather than executed. The policy snapshot is recorded on every action.
- Dry-run / sandbox: POST /agents/:name/dry-run returns would_auto_execute vs would_queue per candidate with zero side effects (no execution, no approval-queue rows).
- Scoped principals: each agent acts as {act: agent, sub: agent:<name>} with one catalog scope and one tool; every side effect goes through the tool-policy point (scope, tenant, and policy checked; audited like an MCP tool call). mandate:issue is never an agent scope.
- Idempotency: a candidate is not re-proposed within 6 hours.
- Observability: GET /agents/metrics — live vs dry-run runs, scanned / auto / queued / failed, success rate, per-agent breakdown, pending-approval depth.
- Runs on demand from the Agent Console or on a schedule.
- The agents are rule-based: their decisions follow the written policies, not a language model.
Finance co-pilots
Available on requestRules-based helpers for your finance team. The co-pilots find the problems and suggest the fix, and a person makes the call. The two accounting engines follow fixed accounting rules. AI never posts to your books on its own. There are no screens for these in the admin console yet — we turn them on for you and connect them to the tools your team uses.
AP exceptions & 3-way match
Lists supplier invoices that don’t match the purchase order or the receipt, ranked by dollars at risk, and suggests what to do — approve, ask for a credit, hold, or escalate. A person applies the decision; it never approves on its own.
Margin leak & cost of goods
Compares the cost you book for each item with what you actually pay suppliers, and flags items sold below cost, missing costs, and cost drift. Read-only: it recommends, it doesn’t change your prices or costs.
Collections worklist & cash-apply preview
Ranks overdue invoices by balance and days late, and previews how a payment would be applied across a customer’s invoices. A receipt is posted only when a person confirms it.
Continuous controls scan
Scans supplier invoices for likely duplicate payments and unusually large amounts, with the dollars at risk. Read-only.
Inventory valuation
An accounting engine that keeps a running weighted-average cost for your stock and reconciles cost of goods sold to it, following fixed accounting rules.
Revenue recognition
An accounting engine that spreads revenue over time under ASC 606 rules and shows a deferred-revenue roll-forward that ties out to the ledger. Recognition runs when your finance team starts it.
Technical detail
- Co-pilots: deterministic, side-effect-free decision engines over the native AP and AR sub-ledgers; operator-only API routes (JWT + tenant + actor-scope guards, @OperatorOnly).
- Inventory valuation: perpetual weighted-average cost (WAC) on every stock movement, with a COGS true-up journal posted by the engine when an order is paid (reversed on cancellation) and a GL tie-out report. Revenue recognition: RATABLE / USAGE / MILESTONE / POINT_IN_TIME schedules, deferral to Deferred Revenue (GL 2300), idempotent recognition postings, and a waterfall tied out to GL 2300. Both are rule-based accounting in the native ERP — no language model is involved in a posting.
- Writes (apply an AP disposition, post a receipt, recognize revenue) run only on an explicit operator call; operator overrides of a recommendation are recorded.
- Published as MCP tools for operator credentials: 7 read tools (finance:read) and 3 write tools (finance:write, confirmation required). Back-office scopes are stripped from buyer and AI-agent credentials, and write tools are never offered to an AI assistant.
- A read-only finance analyst assistant can answer plain-English questions over the read tools; it is switched on per customer.
Operations co-pilots
Live in the admin consoleHelpers your team opens when it needs them. Each one does the checking and the math, and your team decides.
EDI exception co-pilot
Checks incoming EDI 850 purchase orders against contract prices, units of measure, and minimums, and drafts the 855 acknowledgement — clearing clean orders and flagging only the real exceptions.
Supplier invoice capture
Reads a supplier’s invoice, checks the arithmetic, and fills in a draft payable for your team to review and correct before it is created.
Should-cost negotiation
Benchmarks what you pay each supplier against peers or a cost target, sizes the annual savings, and drafts a negotiation brief.
Account onboarding
Checks a new business account’s details, catches duplicates, applies your terms and credit policy, and returns ready, needs review, or blocked — with the setup plan.
Own-fleet rate-shopping
If you run your own delivery fleet, compares it with the carriers on each lane and shows whether to ship it yourself or buy a label — and what that saves.
On the roadmap
Planned · not available todayAn autonomy dial per customer
Choose how much the agents may do on their own — from “suggest only” to “act within limits” — for the whole company, a team, or a person.
Limits you set yourself
Today every agent uses the default limits shown on this page. Changing them for your business in the console is planned.
Finance co-pilot screens
Screens in the admin console for the finance co-pilots, which today are available on request.
How the buyer-facing assistant works: How our AI works. Letting outside AI assistants order: Ready for AI shopping assistants.
See the agents on your own data
Book a demo and we’ll run the agents in practice mode against your orders, invoices, and stock — so you see what they would do before they do anything.