Calculators & tools
Six working tools to size the problem before you try to fix it — what your habits cost, which model a task needs, and how disciplined your team actually is.
The waste calculator
Waste doesn't add up — it multiplies. Using a premium model for routine work is expensive; doing it inside a 15-turn thread, twice, asking for long answers each time, compounds that several times over. Set how your team actually works and see the real gap.
Company AI spend & waste estimator
The rates used
Illustrative £ per million tokens, input / output: light 0.20 / 0.80, standard 2.40 / 12, premium 12 / 60. Chosen to reflect the shape of real pricing — output costs roughly 4–5× input, and premium runs roughly 50–60× light — not any one vendor's list price.
Where it stays conservative
Reasoning models also bill hidden "thinking" tokens as output, which can dwarf the visible answer — not counted here. Failed attempts usually stay in the thread too, inflating later turns. Both push the real figure higher.
Where it may overstate
Prompt caching can cut the cost of re-sent history substantially where a tool supports it, and some tools trim old context automatically. On flat per-seat plans this waste shows up as usage caps and throttling rather than an invoice.
Which model does this task need?
The question is never which brand you're using. It's whether the task needs speed and convenience, or judgement and reasoning. Answer three things and get a recommendation.
Model selector
Prep step first
Your input is still messy. Use a lighter model to clean it into a precise prompt, then send that to the model above. Don't spend premium capacity working out what you meant to ask.
Two ways to do the same job
The saving never comes from avoiding AI — it comes from avoiding bad workflow. This compares the lazy route (one heavy model does everything, twice) against the disciplined three-step route on the same task.
Workflow comparison
The lazy route
The disciplined route
Relative units, not currency — heavy models cost materially more per token than lighter ones, especially on long outputs. The ratio is what matters.
What does this prompt actually cost?
Tokens are the unit you're billed in — very roughly four characters of English each. Paste anything you'd normally send and see the size of the call before you make it.
Token & cost estimator
Rates are illustrative placeholders chosen to show the ratio between tiers, not any vendor's real pricing. Token counts are approximate — every model tokenises slightly differently.
Why one long chat costs more than ten short ones
This is the cost almost nobody sees. Every time you send a message, the whole conversation so far is sent again — so a chat doesn't cost a steady amount per turn, it compounds. Ten turns in one thread can cost several times ten turns across fresh chats.
Context accumulation
One long thread
0tokens processed in total, because each turn re-reads everything before it.
Fresh chat per task
0tokens processed — each task carries only its own context.
The difference
0×more work for the same number of questions.
Assumes roughly 600 tokens of new content per turn. The exact numbers vary; the shape — linear versus compounding — is the point. Caching and context trimming reduce this in some tools, but rarely eliminate it.
Agents change the arithmetic
Everything above assumes a person asking a question and reading an answer. Agentic tools break that assumption: you give one instruction, and the system decides for itself how many model calls to make. A single task can become dozens of calls, each carrying the accumulated context of the ones before it.
One instruction, many calls
Agentic workloads consume many times more tokens per task than a chat message — commonly cited as five to thirty times, and higher for long-running work. The person issuing the instruction has no visual cue that anything unusual is happening. It looks like one request.
The loop is the cost
Agents work by reading, acting, reading the result, and deciding again. Each cycle re-sends context. The context-accumulation problem from the previous section applies here too — except the agent is generating the turns, not you, and it may run for minutes without pausing.
Failure is expensive
When a chat answer is wrong you notice immediately and stop. When an agent is wrong it may keep going, correcting itself, retrying and burning capacity on a path that was never going to work. Set a scope and a stopping condition before you start.
Model choice matters more, not less
The ten rules do not stop applying because a tool is agentic — they compound. A premium model chosen for a routine agentic task multiplies across every call in the loop. Where a tool lets you set which model it uses for sub-tasks, that setting is worth more than any other single control.
Scope tightly, then widen
Give an agent the narrowest useful task first: one file, one folder, one defined outcome. Broad instructions on large inputs are where runaway costs come from. If the first pass works cheaply, widen it. If it doesn't, you've spent very little finding out.
Keep a person on the trigger
Anything that sends, pays, deletes, publishes or contacts a third party should need a human to approve it. This is a cost control and a risk control at once — it caps runaway spend and stops an agent acting on a misread instruction in a way that's hard to undo.
None of this is an argument against agents. They are the most useful thing to arrive in a while, and for the right task the economics are excellent — a task that would take a person an hour is cheap at almost any token price. The point is that the gap between a well-scoped agent task and a badly-scoped one is far wider than the equivalent gap in chat, so the habits matter more here than anywhere else on this page.
Cost is only half the question
Everything so far has been about spending less. That's the easy half. The harder and more important question is whether you're getting anything back — because a tool that costs twice as much and saves ten times the time is a bargain, and a cheap tool nobody uses is pure waste.
Payback estimator
Time released is only value if it gets used for something. Treat this as the size of the prize, not money in the bank — and be honest about the "minutes it takes now" figure, which must include checking the output.
What's worth measuring
Cost per successful task
Not cost per task. If one in three outputs is discarded, your real cost is three times the headline. This single measure catches quality problems that a spend dashboard will never show you, and it's the number most worth tracking over time.
Time on a named process
Pick two or three specific, repeated processes and measure those, rather than asking people to estimate time saved in general. Self-reported productivity gains are unreliable in both directions. A defined process with a before and after is evidence.
Quality, both directions
Track error and rework rates alongside speed. AI can improve consistency and it can introduce confident mistakes. If you only measure how fast work gets done, you'll miss the more expensive half of the picture.
Where the time went
Hours released only count if something happened with them. Did the team take on more work, reduce a backlog, drop an outsourced spend, or stop doing overtime? If the honest answer is "nothing visible," the saving is theoretical.
Adoption depth, not licence count
Seats issued tells you what you bought. Active users, and how often they use it for real work rather than trying it once, tells you what you got. A large gap between the two is the cheapest problem on this page to fix.
Be sceptical of your own numbers
Benefit cases assembled after a purchase tend to justify it. Set the measure before you roll something out, agree what failure would look like, and be willing to report that a tool didn't earn its keep. That's what makes the successes credible.
Energy and carbon, without the exaggeration
This gets discussed badly in both directions — dismissed entirely, or inflated into something it isn't. The measured position is more useful than either, and it happens to point at exactly the same behaviours as the cost argument.
One prompt is genuinely trivial
A peer-reviewed 2026 study in Joule put optimised frontier-scale inference at a median of 0.31 Wh per query — under a watt-hour, and comparable to a few seconds of a kettle. It also found that widely-quoted public figures overstate energy use by four to twenty times, because they assume non-production conditions. If you've seen the 2.9 Wh number, it's one of those.
Reasoning and agents are the exception
The same study found long reasoning and agentic queries raise consumption by more than an order of magnitude, through greater token generation and lower serving concurrency. Other researchers put reasoning mode at five to ten times the power draw of a standard response. The variation between a well-chosen and badly-chosen model dwarfs the baseline.
Scale is where it counts
Inference now accounts for 80–90% of AI compute load, and it runs continuously rather than once. The IEA's assessment is instructive: simple text queries replacing all conventional web search would use under 4 TWh a year — small. Video, reasoning and agentic tasks are what change the picture.
Energy & carbon estimator
Anchored to a median of 0.31 Wh for a standard single-turn query, scaled by tokens processed and model tier. The underlying range is wide (0.16–0.60 Wh) and providers publish little, so treat this as an order of magnitude. Excludes training, water and embodied hardware.
Nobody should stop using AI to save a watt-hour, and a guide that told you to would be silly — AI often replaces something more energy-intensive, like a journey or a stack of printed drafts. The useful observation is narrower: the efficiency argument and the environmental argument point at exactly the same behaviours. Choosing the right model, starting fresh threads and asking for the length you need reduces cost and compute together. You don't have to choose between them, which makes this one of the few places where the sustainable option is also the cheap one.
Survey your team
Twelve questions covering model discipline, token habits, data risk and verification. Each person answers and gets a short anonymous code. Whoever is coordinating pastes the codes in below to get a company-wide score. No accounts, no server, no personal data — the code carries only the twelve answers.
AI discipline survey
Answer all twelve to generate your code.
Codes work well up to a few dozen people. Beyond that, collect responses through your usual survey tool, or have a Cloudflare Worker receive them directly — the same twelve questions, stored and averaged automatically. Keep it anonymous either way: a survey that identifies individuals measures how people answer, not how they work.