- Home
- Articles
Articles
How the work is done: finding unused licences, pricing tokens, checking invoices, asking suppliers for offers. Written without benchmark figures, because a saving is only known once it is computed from your own data.
Software licences
Seats, plans and renewals: who holds a licence, who uses it, and what the right plan would cost.
What SaaS licence management is, which records to keep for every application (seats, plans, renewals, owners) and the yearly rhythm that keeps them true.
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.
A step-by-step method to find unused software licences from the vendor's own data, ask the seat holders, and remove seats before the renewal date.
When a cheaper SaaS tier is enough, how to set a usage threshold per tier, when to ask the seat holder, and how to compute the saving of a downgrade.
How to ask seat holders whether they still need a licence, with the wording, the answers, one reminder, the manager step and the people who are never asked.
How to prepare a SaaS renewal. Work out the notice date, the seats you need and the plan mix, and decide what to ask the vendor for before the deadline.
How to find two SaaS applications that do the same job, count the people who hold a paid seat in both, and consolidate at the renewal of the tool you drop.
How to find paid seats still held by people who have left by joining each seat list with the directory, and how to close the gap in the offboarding process.
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.
AI costs
What models cost per token, where the spend goes, and how a gateway meters and routes it.
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.
Audit AI spend from the providers' own usage reports. Price every token, reconcile with the invoice, and break the total down by provider, model and project.
What input, output and cached tokens are, why output tokens cost more than input tokens, and how to compute the cost of a single LLM request step by step.
An LLM gateway is a proxy between your applications and the model providers. What it meters per request, how upstreams and keys work, and when it is worth having.
How model routing works. Classify each request, set thresholds in a policy, serve simple requests from a cheaper model, and measure the price difference per request.
How prompt caching and context trimming lower the input cost of long LLM conversations, how to compute each, and when trimming costs more than it avoids.
How to compute, per person, whether an AI chat seat licence or paying per token is cheaper, with the break-even formula and a worked example.
A method to attribute LLM spend to teams, features and people with keys, tags and user identifiers, price every request and reconcile with the provider.
How monthly budgets per API key and per user and model allow-lists work for LLM spend, what happens at the limit, and how to set the amounts.
Invoices and spend
Checking invoices, finding duplicates and price increases, and adding spend up by supplier, item and site.
Six checks for every supplier invoice before it is paid: formal details, arithmetic, price, duplicate, billing period and bank details, each with its rule.
Matching rules for exact and near duplicate invoices: normalised invoice numbers, amount and date, credit notes, and what to do once a duplicate is found.
Build a unit price history per item and supplier from invoice lines, spot the increase, compute what it costs per year and check it against the agreement.
How to build a spend analysis from invoice lines: clean suppliers, items and units, assign categories, and read spend by category, supplier, item, month and site.
What an e-invoice is under EN 16931, how XRechnung, ZUGFeRD, Factur-X and Peppol differ, what German VAT law requires, and which invoice checks change.
How payment diversion fraud works, the warning signs on an invoice, and the control that stops it: compare the account with the record and verify by call-back.
What tail spend is, how to measure it from invoices by supplier, invoice count and category, and how to bundle the purchases that repeat.
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.
Sourcing
Requests for quotation, comparing offers, negotiation rounds and agreed prices.
What a request for quotation must contain so that offers come back comparable: items, yearly quantities, delivery terms, a deadline and one answer format.
Compare supplier offers by arithmetic: same unit and quantities, shipping per year, incomplete offers kept apart, and a comparison table that shows every step.
How to run negotiation rounds with suppliers: target prices taken from the comparison, what may be said about other offers, and when to stop and award.
How to buy for several delivery sites: quantities per site from invoices, one request or several, comparing offers site by site with freight, and site prices.
When and how to follow up with a supplier who has not answered: timing, wording, what counts as an answer, when to escalate, and how to keep the thread.
How to check later invoices against agreed prices line by line, claim back what was billed above them, and measure the saving on what was actually billed.
AI in procurement
What an AI may do by itself, what a person has to approve, and how savings are proven.
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.
What an AI in procurement may do alone, what needs a person's approval, and how to design the approval point so that it is a real check.
A language model writes plausible numbers, not computed ones. How to keep model-made figures out of savings reports, supplier messages and invoice data.
What the Model Context Protocol is, how it connects a company's own assistant to procurement data, and why the assistant should read and propose only.
A method for proving procurement savings. Keep in question, agreed and realised apart, separate one-off from yearly, and measure on invoices.
How to let an AI agent work in a vendor's or supplier's website. Approved values, allowed hosts, hidden logins, a stop before submit and a full record.
One list of everything in procurement that waits for a person, ordered by deadline and worth, with decide, waiting and automatic kept apart.