Ask a question nobody prepared a report for, and get the answer in seconds, with every figure traced.
In 2026 we rebuilt Spendkey from scratch around a team of AI agents. They read your contracts and invoices, call tested tools for every sum, and give verdicts a CFO can sign. This page follows one question from the moment it is asked to the answer on the screen.
The CFO asks which suppliers billed above the price on the purchase order. The board meets at nine tomorrow. Nobody has a report for it.
A year ago this was a week of analyst time and a spreadsheet nobody could check. Pick the question below and watch what happens instead.
Pick a question and watch how Cassian answers it on the demo workspace.
Each answer shows the tools Cassian called and marks every figure the check traced.
Billed above the unit price agreed on the purchase order: £1.044k ✓ traced across 169 invoices ✓ traced. This is money to claim back from the supplier.
| Supplier | Invoices | Above PO price |
|---|---|---|
| Garthby Waste Services Ltd | 16 | £548.49 |
A separate £48.130k ✓ traced was billed above what the purchase order supports in total. That may be legitimate, so it is kept apart and never added to the first figure.
The model planned the work, tested code did the sums, and a final check removed anything it could not trace.
The language model decides what to ask of your data and how to explain the answer. It never decides what a number is. That is why we only show figures that are 100% traced to your own data.
Your question
Asked in plain English, in any language.
The agent plans
The language model picks which tools to call and with what inputs.
no sums, no memoryTested tools run
Totals, shares, price comparisons and matching are fixed code.
10,000+ unit tests per releaseYour data answers
The tools query your database directly, read-only.
reconciled to the centEvery figure traced
Each amount is traced to a tool result, or removed before you see it.
runs as code, every answerThe agents never do arithmetic
They do not answer from memory. They choose a tool, give it inputs and read back what it returns.
The tools are tested code
The tools query the database directly. Every calculation they perform, from totals and shares to price comparisons and matching records, is covered by unit tests. More than ten thousand run before every release.
A check traces every figure
Before an answer reaches you, each amount or share is traced back to a tool result from that same answer. Anything it cannot trace is removed, so a number the model invented or misremembered is never shown.
We try to break it first
Before any change goes live, we test the agents with long adversarial conversations built to make them misquote, overstate or calculate on their own, and fix what those find.
A team of agents works across the platform, each with one job, on the same cleaned data.
Some prepare your data, some find the money, some hold the margin on the next deal, and Cassian is the one you talk to. Every one of them calls tested tools for its sums.
Prepare the data4 agents
- Normalisation agentCleans supplier names, currencies, units and dates across every ERP you send.
- Parent research agentFinds who really owns each supplier from public filings and company websites.
- Classification agentPuts every spend line in your own category tree, with the reason it sits there.
- Document reading agentReads supplier and customer contracts, invoices and purchase orders, in any language, down to the clause.
Find the money4 agents
- Market data agentLinks each part and each price to the published index behind it, and projects where it goes next.
- Opportunity Assessment agentSets out every savings opportunity with its evidence, range and next step.
- Recommendation agentRuns more than 200 checks in 20 groups, such as regulation, country and category, on each opportunity.
- Audit agentChecks supplier and customer invoices against their contracts, and keeps each kind of money separate.
Hold the margin4 agents
- Policy agentTurns what was found into rules in your own words, with tolerances, for the function head to accept.
- Simulation agentTests a rule on the last 12 months, on live data and on the market forecast, and shows what a price, index or clause change does to margin.
- Verdict agentTests each purchase order and contract against your rules before it is committed, and names who decides.
- Recovery agentPrepares each claim with its evidence and chases it until the supplier repays.
Talk to it1 assistant
Which customer contracts lose money if copper rises 10%?
Seven contracts. The largest is the busbar supply agreement, which reprices once a year. I have traced every figure to your invoices.
- CassianAnswers any question, explains any verdict, and builds the dashboard you ask for.
The agents follow a playbook our analysts wrote by hand, one spend line at a time.
Anyone can connect a language model to a spreadsheet. Knowing what a correct answer looks like, for a cleaning contract in Rotterdam or a fuel clause in a haulage agreement, is what takes the time.
Every judgement was written down: which category, which parent supplier, and why. Then it was checked by a second analyst.
Document reading and generative AI first, then Cassian and the Opportunity Assessment Agent, built on the answers our team had already checked.
Every line carries the reason it sits where it does, and every figure is traced, so your team can check the agents the way we checked our analysts.
A general AI model can read your file. It will not do the three things a margin decision needs.
Your own Copilot, Gemini or ChatGPT can still use the same cleaned data through our MCP server, so every tool gives the same numbers.
Enterprise security, and a number you can check.
Your files reach Spendkey through one encrypted transfer, and every connection after that is encrypted as well. Your data stays in a region you choose on Microsoft Azure, on secure in-house hosting, or on your own premises.
ISO 27001Certified information security
SOC 1 Type 2Controls over financial reporting, tested over timeSOC 2 Type 2Security, availability and confidentiality, tested over time
GDPRCompliant, with a data processing agreement
Microsoft AzureHosted in the region you choose
Microsoft partnerSingle sign-on with your identity providerAsk every AI vendor these questions, including us.
Most vendors, most internal teams and most consultancies fail at least two of them.
What is in production today, and with whom?
The platform is live with clients in Europe, North America, Africa, the Middle East and Australia, on more than 100 million transactions. The results page names them.
If I ask it something nobody prepared for, does it answer in seconds?
Yes. Cassian plans the steps from the tools it has, so a new question does not need a new report. Try the four questions above.
Who on your team has carried a savings target or signed a supplier contract?
Our founders and analysts come from procurement, finance and consulting. The platform was built by people who did the job it now does.
Can the model make up a number?
No number reaches you unless it traces back to a tool result from the same answer. The check runs as code after the model has written its answer, and removes anything it cannot trace.
Is Spendkey AI-native?
Yes. In 2026 we rebuilt Spendkey from scratch as the Margin Control Platform, replacing the original platform rather than adding AI to it. The agents are the product: they plan the work, call tested tools for every figure and explain the results, and Cassian is how you talk to them.
Bring a question nobody has prepared a report for.
We will answer it live in 30 minutes, on the demo workspace.
