Reducing the IT bill: the levers that work, AI included

Published 24 July 2026 · updated 16 August 2026

An IT bill holds more than two cost categories. The per-seat licences and the metered consumption are the visible ones, but hardware, networks, support, managed services, security, projects, training, internal labour, downtime and exit also belong in the total. And each category needs its own lever.

When AI is a material expense and the workloads repeat, three changes pay for themselves within a few invoices: caching, batch processing and model routing. All three still need measurement and quality tests, and sometimes development. Changing tools comes later, where licence cost or lock-in dominates.

AI: four levers before changing model

Measure before cutting

An AI bill arrives as one global amount. Until you attribute the spending by team and by function, nobody knows whether this month’s 4,000 francs came from customer service or from a forgotten script. A gateway like LiteLLM, hosted on your side, gives each team a key with its own ceiling; Langfuse adds visibility per request.

Caching: a frequently missed saving

When several requests begin with identical content, a regulation or a catalogue for example, some providers discount the cached input. As at 14 August 2026, OpenAI lists a ratio near 0.1× for certain models. The ratio isn’t universal: cache creation, storage, TTL, minimum length, exact prefix matching and the output price all matter. The saving on the second request only arrives when those conditions are met. And a changing date at the start of a prompt may stop the match without producing an error.

Batch processing

Some providers currently discount asynchronous batch processing by as much as 50%. Sorting the night’s mail, summarising the week’s minutes, extracting fields from two thousand invoices: none of that needs an answer second by second. Check the model, the region, the deadline, the limits and the error policy: real-time pricing isn’t always exactly double.

The right model in the right place

Within one provider, the price gaps between model families may be large, and they change with the catalogue. Use the expensive model only where testing proves you need its quality. Use a smaller model, or deterministic code, everywhere else. Deduplication, sorting and counting often need no model at all. The next step, open-weight models, changes the order of magnitude again: a dedicated article has the figures.

The rest of the bill

Virtualisation

VMware’s licensing and offers changed after Broadcom bought the company. Proxmox may cut the licence cost in some environments, but it isn’t a cost-free equivalent. Compare the compatible hardware, the backup, the networking, the high availability, the support, the skills and the migration. An estimate of eight to twelve weeks means something only with a stated scope, an inventory and acceptance criteria.

Per-seat licences

Nextcloud may replace one SaaS seat licence with infrastructure you have to operate, potentially with Enterprise support on top. The honest comparison is use against use: a company that lives in Excel and Teams doesn’t replace them with file sharing.

E-signature, with a Swiss caveat

Documenso and DocuSeal may cover certain routine signature workflows, depending on the configuration and the evidence. Don’t assign a legal signature level without checking the exact version and the service provider. Under Swiss Code of Obligations Art. 14(2bis), equivalence to a handwritten signature requires a qualified electronic signature with a qualified electronic timestamp, based on a qualified certificate from a recognised provider, subject to statutory or contractual exceptions; see OFCOM. Classify the documents by their formal requirements, then pay for the level each one needs.

Where free costs more

None of the tools above is free: the licence disappears, the operating cost appears, and you pay it in hours, in installation, in updates, in tested backups, in the Friday-evening call. Three situations often make the switch a bad deal: a small team whose licence saving can’t fund reliable operations; a tool whose compliance depends on certification or on a process that needs specialist review; and a package the whole team knows, when retraining costs more than the expected saving. Ten users isn’t a universal threshold. Calculate it.

Five questions before deciding

I don’t touch the model before the four levers above are measured. Changing the model first is the expensive way to discover that the bill was never about the model.

What share of the bill does this line represent? An 80% gain on 3% of the bill goes unnoticed.
Does the cost follow headcount or use? Per-seat is fixed by changing tools, consumption by changing method.
Who operates it after the switch? If the answer is “the person who proposed the change”, check they’ll still be there in two years.
What does going back cost? An open format reduces the risk, but export, re-import, metadata, permissions, history and dependencies still need a test.
On AI, is measurement in place? Until the spending is attributed, everything else is steered blind.

The comparison that decides

Compare three years of licences, infrastructure, internal labour, support, security, migration, training, downtime, compliance, exit, residual risk and resale value. Add a low case and a high case for volume, price and staff time. A saving that disappears with two hours of monthly administration isn’t structural.

Where the quoted prices come from

I took the prices and the product properties on 14 August 2026. Cache, batch and model ratios vary by provider, model, region and contract. Calculate the durations and the savings for the actual scope.