Guide // MandateGuard
How does duplicate-commitment detection stop agents paying twice?
By the Ingenious AI team · Published 2026-11-14 · Reviewed 2026-11-14
▸ Run the free live demo See pricing
Free covers 1 agent and 1 policy with the full authorize flow, signed receipts and the approval inbox — no card required to start.
Short answer
Duplicate-commitment detection watches for an agent taking on the same obligation twice: the same vendor and the same amount authorised again inside your window, while the first commitment is still in force. Every single-request check can pass on a duplicate — right vendor, normal amount, room in the day — so MandateGuard compares the request against the workspace's own receipt history and flags the second one to a human before money moves twice. On Pro and Team it turns the most silent shape of agent waste, the doubled subscription or licence, into an approval decision with the original receipt attached.
What is duplicate-commitment detection?
A duplicate commitment is a second authorisation for the same obligation: same vendor, same amount, inside a window where the first is still in force. MandateGuard's duplicate-commitment detection, on Pro and Team, watches the workspace's own receipt history for that pattern and flags the second authorisation to a human before the spend completes. It judges commitments, not clicks — the question is whether the business is about to owe the same money twice.
Why agents duplicate so easily
Humans duplicate purchases through forgetfulness; agents duplicate through architecture. Two workflows share one workspace and neither knows the other already bought the dataset. A retry after a network timeout re-runs the checkout step because the first success response never arrived. A subscription converts from trial to paid while a second agent, tasked with ensuring coverage, buys the same plan outright. In each case every individual check passes: the vendor is allowlisted, the amount is normal, the day has room. Duplication is invisible to per-request controls because the error is not in the request — it is in the relationship between this request and history. That is why the detector reads the ledger, not the request.
What the detector compares
MandateGuard's duplicate-commitment detection evaluates a new authorisation against the workspace's receipt history, looking for the same vendor and the same amount inside your window — the signature of a commitment being taken on twice. The window matters: the same 49-dollar dataset bought monthly is a duplicate on day 14 and a renewal on day 31. Because the ledger is hash-chained, the history being compared is verifiable; the detector is not reasoning over a log that could have been tidied. When the pattern matches, the second authorisation is flagged to a human before money moves twice, with the original receipt available for the decision.
Escalation, not a silent block
The detector escalates rather than refusing outright, and that restraint is deliberate. Some repeated vendor-and-amount purchases are legitimate: a second licence for a second team, a re-order of a consumable, a top-up that matches last month's by coincidence. A hard block would train operators to bypass the control. Escalation keeps the human where judgement lives — is the first commitment still in force, and does it cover this need? The approval flow then resolves it: approve and a one-time token releases exactly that purchase, or deny and the request closes. Both outcomes are receipted, so the next duplicate review starts from a richer history.
The subscription shape of duplication
The most expensive duplicates are rarely single purchases; they are commitments that recur. An agent that re-subscribes creates a second monthly charge that will repeat quietly until someone reconciles statements. Detection at authorisation time is the one moment where the second commitment is still free to prevent. Teams should support it operationally: keep a simple register of subscriptions the agent legitimately holds, with renewal dates, so a reviewer answering an escalation can confirm coverage in seconds. The register and the ledger check each other — disagreements between them are exactly the findings worth having.
What duplicates reveal about your workflows
Every escalation is a free audit finding. Recurring duplicates between two workflows mean the workflows need a shared source of truth about what has been purchased — even a task-reference convention helps, because receipts carry the reference and reviewers can see which job bought what. Duplicates following tool errors mean retry logic treats checkout as idempotent when it is not; the fix belongs in the agent, and the detector buys you the time to make it safely. Duplicates near renewal dates suggest the agent is compensating for invisible coverage — often a sign the humans, not the agent, lack a register of what the business already pays for.
Plans and related controls
Duplicate-commitment detection is part of the paid depth on Pro (29 dollars a month) and Team (79 dollars a month), alongside vendor trust, the kill-switch, evidence export and the intent service. It composes with the free controls rather than replacing them: caps still bound size, the allowlist still bounds counterparties, and approvals still supervise the top band. Together they cover the realistic accident taxonomy of agent spending — too big, too often, wrong seller, twice, and unsupervised — with each incident class leaving a signed receipt instead of a surprise on a statement.
Five actions to take now
- Define duplicates for your shop: same vendor, same amount, inside your window
- Turn on duplicate-commitment detection on paid plans
- When an escalation lands, check the first receipt before deciding
- Track renewal dates for subscriptions the agent legitimately holds
- Fix retry logic in agents that generate accidental duplicates
Worked example: the second licence that was not bought
A research agent buys a 49-dollar monthly dataset licence on the 3rd, receipt recorded, task reference data-licence-march. On the 17th, a second workflow in the same workspace — testing a new enrichment path — reaches the checkout step and requests the same vendor for the same 49 dollars. The vendor is allowlisted, the amount is under the cap, and the day has room, so caps alone would allow it.
Duplicate-commitment detection compares the request with the workspace ledger, finds the matching commitment inside the window, and escalates it to a human. The reviewer opens the first receipt, confirms the licence already covers the team until the 3rd of next month, and denies the second. The genuinely new purchase the workflow needed — a different 19-dollar dataset — proceeds normally the same afternoon.
Frequently asked questions
Does it block legitimate repeat purchases?
It escalates for a human decision rather than silently blocking, because some repeats are legitimate. The reviewer sees the original receipt and decides with context.
What counts as the same commitment?
The pattern is vendor plus amount within the detection window, read from your workspace's own receipt history.
Can agents trigger duplicates without retry bugs?
Yes — parallel workflows, overlapping jobs and subscription conversions all duplicate without any single agent malfunctioning.
Is duplicate detection on the free plan?
No — it is part of the paid depth on Pro and Team, with vendor trust and the kill-switch. The free plan records the receipts that make manual review possible.
What should I do when an escalation arrives?
Open the first receipt, confirm whether its coverage is still active, and approve or deny. Repeated escapes mean a workflow needs redesigning.
Keep reading
- What is vendor trust scoring from your own receipt history?
- How do you prevent duplicate subscriptions from AI agents?
- How do you stop an AI agent from overspending?
Put the policy in force: spend envelopes · intent registry · approval inbox · how the proof works · all MandateGuard guides