the challenge
Finance wants AI. The data question says no.
Most AI tools send your invoices to a model company and ask you to trust its policy. That is where adoption stops. Loopclose removes the question.
the difference
Ask where your invoices go
For finance teams that refuse to choose between AI and control of their data.
Most AI invoice tools
A policy page, a compliance badge and a promise not to train on your data. Whether your invoices are stored, and by whom, is left to the fine print.
Loopclose
Open-weight models inside AWS. Nothing stored at the model layer, nothing used for training, no OpenAI, Anthropic or Google in the path. A person approves every decision.
a sample run
Four invoices, one PO list
Demo data, not a customer. Each exception comes with a reason in plain English and a drafted supplier email.
the data path
Short enough to read to an auditor
Your invoice
A PDF, a scan or a CSV. Pilots start with anonymised files.
An open-weight model in AWS
US region. Nothing is stored at the model layer and nothing is used for training. No OpenAI, Anthropic or Google in the path.
A result for a person
Findings, a reason in plain English and a recommended step. Your team decides.
the engine
Matches every invoice against its purchase order and flags price, quantity, duplicate and missing-PO problems. The arithmetic is done by code, never guessed by a model.
Each exception comes with a plain-English reason and a recommended next step.
Drafts the supplier email with the exact numbers and opens it in your mailbox. You press send.
Every decision goes into a tamper-evident log, with an evidence pack you can hand to an auditor.
the method
The first loop
Offer a sample
10 to 20 paid invoices and their PO lines. Anonymise them first: remove bank details, addresses and personal names.
The loop runs
Within 48 hours, on the data path above. Nothing connects to your systems.
The result returns
A one-page summary of mismatches, duplicates and overpayments, plus drafted supplier emails. It’s yours either way.
the path
In the limit, finance closes its own loops.
Invoice exceptions are the first loop. Contracts, supplier credits and spending policy come next, each when a customer asks for it. Every loop runs on the same data path, and nothing is kept.
The bottleneck is adoption.
Actual adoption
The models are no longer the constraint. The constraint is that firms have not reorganised around what the latest models can do. Buying a chatbot licence is not adoption. Adoption is when the work itself is redesigned: which tasks belong to an agent, which belong to a person, and who answers for the result.
Adoption is not a ceiling
How little a company uses AI today says almost nothing about how much it will matter. A low number means the gap between what is possible and what is deployed is wide. It does not mean the impact is small.
We think AI is still underpriced
We think the value of the work these models can take over is far larger than what companies pay for it today. That is a belief, not a forecast, and it is the reason we build.
Why adoption stops
In finance, adoption rarely stops at what the model can do. It stops at one question from IT or the auditor: where does the data go? Most vendors answer with a policy page. We answer with a data path. Remove that question and the rest of adoption can begin.
Why finance
Finance is full of work software never touched: reading a supplier’s reply, checking an invoice against a contract clause, chasing a credit note. It is judgment, language and paper, spread across inboxes and folders. It is where reorganising around the models pays off first, and where the proof is a number you can point at.
How we adopt with you
Loopclose closes one loop at a time. The agent does the work, your team directs it in plain language, and every instruction stays in the system. Nothing is kept at the model layer, numbers come from code, nothing is paid or sent without a person, and each decision leaves evidence.
Kacper
Four commitments
The agent proposes, you decide
It never pays, posts or sends anything on its own.
Numbers come from code
Matching and arithmetic are deterministic. The model reads documents and explains results.
Every decision leaves evidence
What the agent saw, what it decided and why, with the source file’s fingerprint.
Nothing is kept
Open-weight models in AWS, US region. Nothing stored at the model layer, nothing used for training, no model company in the path. SOC 2 Type I is in preparation, and we complete your security questionnaire before touching live data. Read the security page.
Open the first loop.
Send 10 to 20 paid invoices with their PO lines, anonymised. Within 48 hours you get a one-page exception summary and drafted supplier emails. Yours either way. Three pilots are open.
Request the 48-hour review