We find the savings.We prepare the next step.You knock them off.
Cureman reads what your company buys: software licences, AI usage, and invoices for goods and services. It computes what can be saved, puts the next step ready for every item, and hands you one list.
Nothing leaves your organisation before a person has approved it.
Member of the NVIDIA Inception program for startups.
Potential savings
Each one is prepared. You switch it on or approve it in the dashboard.
- Open:AI$11,422a yearTurn on model routing for simple requests
Ready: measured on your own traffic without changing it. One switch turns it on.
Activatein your dashboard, after sign-in - Open:Licences$4,500a yearReclaim 15 unused seats in ChatGPT
Ready: the question to each seat holder, whether they still need it.
- Open:Invoices$4,431onceMeridian Freight billed invoice ME-2026-1081 twice
Ready: the message asking the supplier to cancel the duplicate.
- Open:Licences$2,621a yearRemove 1 departed employee from Salesforce
Ready: the removal as a step of an action plan, shown in full.
- Done:Licences$2,160a yearCancel the unused Loom workspace
Done. The saving counts from now, not from the day it was found.
- Done:Invoices$1,850onceClearview Cleaning Services billed the same period twice
Done. The request for a credit note was drafted for you.
- Open:Purchasingno figure until someone quotesGet competing offers for what Meridian Freight supplies
Ready: a draft request for quotation with items, yearly quantities and today's prices from your invoices.
An example. The findings and amounts are what the engine computes for the fictional demo organisation that comes with the product, largest first. Which ones are ticked off is our doing. Your own list is in the dashboard.
The platform in just over a minute
What waits for you, what it is worth, and how you approve it.
Narrated by an AI voice. The data shown belongs to an invented demo organisation.
Named findings, not a percentage of your spend.
Each item rests on something of yours that can be pointed at: a seat, a request to a model, a line of an invoice. These are the kinds it looks for.
Seats, plans and renewals
- A seat nobody has used
No sign-in for longer than the period you set, or the holder said they no longer need it.
- A seat of someone who left
Still paid for after the person was marked as gone.
- Licences bought and never assigned
More bought than handed out, beyond the reserve you want to keep.
- A plan above what is used
People on a tier whose extras they do not touch, or a billing cycle that costs more than the other one.
- A price above the vendor's own list
What you pay compared with what the vendor publishes.
- Two tools for the same job
Counted by the people who hold a seat in both.
Booked as the module: Licence optimisation.
Models, tokens and seats
- Simple requests on the most expensive model
Measured on your own traffic first; then routed to a cheaper model of the same provider.
- Long conversations sent in full every time
Sent upstream with only the history the request needs, and only when that is computed to pay. Your conversation is never changed.
- The same system prompt paid for on every request
Prompt caching that is available and not used.
- Identical requests paid twice
A retry loop or a missing cache in the service that sends them.
- Background jobs at interactive prices
Work nobody waits for that could go through a provider's batch interface.
- A seat where pay-per-use is cheaper, or the reverse
Light users of a chat licence, and heavy users of an API who would pay less with a seat.
Booked as the modules: AI spend audit, AI gateway, Context optimiser, Prompt insights, AI workspace.
Goods, services and suppliers
- An invoice billed twice
Same supplier, number and amount, or the same amount and date under another number.
- A period billed twice
The same item and quantity charged again for a time already paid.
- A price above the agreed one
Each invoice line held against the price you agreed with that supplier.
- An invoice that does not add up
Lines and totals that come to more than the invoice's own figures.
- The same item at different prices
From two suppliers, or from one supplier at two of your sites.
- A supplier nobody has asked for a second offer
A request for quotation is drafted from your invoices. No saving is shown until someone has quoted.
Booked as the modules: Invoice audit and spend analysis, Sourcing and negotiation.
It does not wait to be asked.
A finding is not a report to read. For each one the next step is put ready before you open the list, and it is kept moving until somebody has answered.
- By itself
Prepares
The question to a seat holder. The message to a vendor or a supplier. The request for quotation, with items, yearly quantities and today's prices taken from your invoices. The action plan at the vendor, every step shown in full.
- You
Waits for your yes
Nothing leaves your organisation until a person has approved it: no message to an employee, none to a supplier, no change at a vendor. If you want a second person to approve, it asks for one. Everything that waits for a yes is in one list, with what it is worth.
- By itself
Follows up
Every message belongs to a conversation with a clock. When the working days you set have passed without an answer, the follow-up is ready, in fixed words and without a new figure. A reply stops the clock.
- By itself
Passes it on
An employee who does not answer gets one reminder, and then the question goes to their manager, exactly as the plan you approved said it would. A supplier who stays silent comes back to you as unanswered. Nobody loses a seat for staying silent.
A figure you can show to finance.
Every amount can be followed back to the seat, the request or the invoice line it rests on.
Computed by code
Each amount is calculated by code and shown with its calculation. A model never produces a figure; a draft that states one of its own is discarded.
Unverified is labelled
A list price from the catalogue or from research is marked unverified until a person has checked it. Unverified figures are shown and not added up.
What a model reads is checked
An invoice read by a model is checked by arithmetic and labelled until someone has reviewed it. An offer in a supplier's reply counts once a person confirmed it.
Nothing is assumed
A saving nobody has quoted or billed is not counted. An amount recovered once is kept apart from what recurs every year.
15 of 38 paid seats have not been used in the last 45 days.
From the same fictional demo organisation as the list above.
Where your data is, and where the AI runs.
Procurement data is invoices, contracts and the names of people. Below is what happens to it, said concretely, with what is so today kept apart from what you can have on request.
- HostingToday
- Cureman runs on AWS in Frankfurt, Germany. What it stores, the invoices you upload included, is stored there.
- Access to your vendorsToday
- Connectors only read; they change nothing at a vendor. Their credentials are stored encrypted. Every query is tied to your organisation, and every change is recorded under the name of whoever made it.
- Your mailboxToday
- From a connected mailbox, only replies to conversations started here are stored. Of any other mail, only the sender is looked at.
- Your employeesToday
- Nobody is written to without a person approving the message, and whoever is on your do-not-contact list is never asked about their licences.
- Data processing agreementOn request
- Processing on your instructions under a data processing agreement (Art. 28 GDPR) that names every sub-processor.
- Where the AI runsToday
- Models read documents, interpret replies and draft text. In the standard installation these calls go to the API of a model provider, today Anthropic. The gateway forwards your own AI requests to the providers you have configured, and to nobody else.
- AI on your premisesOn request
- The models run in your own data centre or in your private cloud instead. Then no invoice, no reply and no prompt is sent to an AI company. Which model runs there is agreed with you and tried on your own documents before it goes live.
- What a model is trusted withToday
- No figure depends on which model runs. Amounts are computed by code, and what a model reads is checked by arithmetic and labelled until a person has reviewed it. That is why the figures do not get less reliable when the model runs on your own premises.
One platform, and the modules you book
The platform holds what every module shares: findings, decisions, conversations, people, the price catalogue and the audit log. Each module adds one area of spend and states what it measures.
The licence optimisation module reads seats, plans and activity from vendor APIs, finds unused seats and oversized plans, and asks seat holders after approval.
The AI spend audit reads usage reports from model providers, prices every token against a model price list and shows spend by provider, model and project.
The AI gateway sits in front of every model provider, records tokens and cost per request and, under your policy, serves simple requests from a cheaper model.
The context optimiser leaves earlier exchanges that a long conversation no longer needs out of the upstream request, and only when that is computed to pay.
Prompt insights classify gateway traffic by task, attribute it to users and tags, and recommend a better-suited tool where the numbers support it.
The AI workspace is a chat front end for employees on the AI gateway. You pay the provider for the tokens used instead of a seat licence for everyone.
The module takes invoices as PDFs, e-invoices or CSV, checks each line against its arithmetic, earlier invoices and agreed prices, and adds up the spend.
The sourcing module turns what you buy into a request for quotation, compares the offers by arithmetic and prepares each negotiation round for your approval.
How the work is done, step by step
The levers that lower a model bill, in the order to pull them. Measure first, then model choice, output length, prompt caching, context, repeats and budgets.
What a language model does reliably in purchasing today, what belongs to plain code, and which steps stay with a person. With a table by task.
Seven places to look when software costs have to come down, ordered by how strong the evidence is and how soon the money stops flowing.
When an early-payment discount is worth taking, how to compare it with the cost of paying early in plain amounts, and how to find the invoices that offer one.
Which Microsoft 365 licences can go or move to a lower plan, found from purchased licences, assigned accounts, last sign-in and a question to the holder.
How to find software that departments bought themselves, using invoices, card statements and the application inventory, and what to do with each find.
What buyers, IT and finance ask first.
What is Cureman?
- Cureman is an AI procurement platform. It finds savings in software licences, AI spend and supplier invoices, runs requests for quotation and negotiation rounds with suppliers, and prepares each next step up to the point where a person has to approve it.
Which costs does it look at?
- Software licences (seats, plans and renewals read from vendor APIs), spend on AI models and chat seats, supplier invoices for software, services and physical goods, and the prices suppliers offer in a sourcing event. Each area is a module that can be booked on its own.
Does the AI decide anything?
- No. A model researches prices, reads documents and replies, and drafts messages. Every figure is computed by deterministic code, and every message to an employee or a supplier is approved by a person before it is sent. A draft that states a figure of its own is discarded.
Where do the savings figures come from?
- From the organisation's own data: seats and activity from vendor APIs, token usage from provider reports or the gateway, line items from invoices, prices from offers suppliers made. A saving that nobody has quoted or billed is not counted, and unverified prices are labelled and kept out of totals.
Can our own AI assistant work with it?
- Yes. The buyer's tools are open to an organisation's own assistant over MCP (Model Context Protocol). An assistant key can read and propose. It cannot send a message, approve, award or change anything.
Can it run under our own name?
- Yes. An organisation can set its own product name, logo and accent colour and serve the dashboard and the model endpoint on its own host names, after proving control of each name with a DNS record.