Liftward · AI value & velocity intelligence

Your AI is working.
Now prove it.

Almost every AI business case is built the same way: seats, times hours saved, times a loaded rate. It produces a large number in about a minute, and it is wrong in four specific ways. This page builds that number from inputs you choose, then takes it apart one honest deduction at a time — and shows you what is left that you could put in front of an auditor.

No account Nothing leaves your browser Every figure recomputed, not asserted The method, not a measurement of you
Build the board-deck number ↓ liftward.ai ↗
Board-deck number—
After baseline & cost—
What you could evidence—
Read this before you move a single slider

This page is not connected to production. Liftward is in active development and in closed alpha access. This demonstration runs entirely in your browser: it reads nothing about you, connects to no system of yours, and reports on nobody's business. It makes no network request of any kind — you can watch that in your own devtools.

What it does instead is perform the method in front of you. The arithmetic is real, it runs in your browser on numbers you set, and every intermediate step is printed beside its result so you can check it rather than believe it. The starting values are round numbers chosen to make the shape of the argument visible — they are not benchmarks, not research findings, and not anybody's data. The point of the page is precisely that only your own measurement can supply the real ones, and that a figure produced from an assumption is that assumption with more decimal places on it.

Before anything else

What you are looking at, and what it is not

The other demonstrations in this suite earn their trust by recomputing rather than asserting. This one earns it the same way — and because it has less to show you, it has to be stricter about saying so.

what it is

Arithmetic, not assertion

Every figure on this page is computed from the inputs above it, in your browser, by the formula printed next to it. Move an input and the whole chain moves. Nothing is fetched, nothing is stored, nothing is sent — and the console in the corner logs each recomputation as it happens so you can watch the sums rather than take them.

what it is

The method, demonstrated

The deductions below are the questions an honest baseline has to answer before it credits a tool with a saving. You are invited to disagree with every one of them — the sliders are the argument, and their defaults have no authority. What you cannot do is skip the questions and still call the answer evidence.

what it is not

Not a measurement of your business

Nothing here reads your systems. Nothing here knows your adoption, your review burden or your baseline. You typed every number. The output is therefore a well-structured restatement of your own assumptions — useful for seeing the shape of an argument, worthless as a finding, and not to be pasted into a board pack.

what it is not

Not connected to production

Liftward is in active development and in closed alpha access, and this page is wired to none of it: no baselines read from your trackers, no evidence-linked deltas, no judge panel weighing what the data cannot say. Those are described in the roadmap at the bottom of this page, described as what they are, and demonstrated nowhere on it.

Why a page connected to nothing is still worth building Because the expensive mistake in enterprise AI is not a bad tool — it is a business case nobody subtracted anything from, defended by a number nobody can reproduce. That mistake is made with a spreadsheet, in an afternoon, and it is entirely visible in arithmetic. So the arithmetic is what this page shows, on the visitor's own figures, with the working left in.
Why this matters

Three independent measurements of one problem

These are the findings of other people, cited as they were published, and they are the only figures on this page we did not compute in front of you. Every other number here comes out of the sliders.

95% Can't show the money.

MIT's NANDA initiative (The GenAI Divide, 2025) found 95% of enterprise GenAI pilots delivered no measurable P&L impact — only 5% created significant value.

42% The purge has begun.

S&P Global found 42% of companies abandoned most of their AI initiatives in 2025.

30% And it is not finished.

Gartner expects 30% of the GenAI projects started in 2024 to be abandoned by the end of 2026.

Read together, those are three different instruments pointed at one thing: the money went out and the return could not be shown. None of them says the tools do not work. They say the work was never instrumented to answer the question that arrives afterwards — and the question always arrives.

What this page does

Seven things, each performed further down

Every row below is performed in your browser somewhere on this page, and the link takes you to the place it happens. If something is not on this list, this page does not do it — and the roadmap at the bottom says what is arriving next.

Act 1 · The number

The number you would put in a board deck

Four inputs. Seats, the hours people say they get back, what an hour of that person costs fully loaded, and how many of them actually use the thing in a given week. Set them to something you recognise from your own estate. Then enjoy the total for a moment, because it is the last time it goes up.

Act 1Seats × hours × rate
400
Everybody the licence covers, whether or not they open it.
55%
The share of those seats that touch it in a normal week. Licence counts and usage counts are rarely the same number.
3.2 h
What people report when you ask them. Act 2 has something to say about that.
€62
Salary, employer costs, and the overhead the finance team attaches to a working hour.
Annual saving, as the business case would state it — Over 46 working weeks — the fifty-two on the calendar less the leave, the public holidays and the training days that a finance team would not let you bill.
The working 
—
Everything wrong with this number is a subtraction Not an error — a subtraction. Each of the inputs above is defensible on its own; the figure they produce is not, because it credits the tool with every hour anybody associates with it. The four acts that follow take those hours back one category at a time, and none of the deductions is exotic. Any finance business partner would ask for all four in the first meeting.
Act 2 · The baseline

What would have happened anyway

A baseline is not a number you look up. It is a set of questions about the counterfactual — what the work would have cost if the tool had never arrived — and every one of them takes something off the top. The four below are the ones that always apply. Set each to what you believe, and watch the figure walk down.

Act 2Four honest deductions
18%
Measure the year before the rollout and the line is usually already moving: process changes, a new hire, a team that got good at something. A baseline that starts on the day the tool arrived credits it with the slope it inherited.
12%
The hours are not saved against zero. They are saved against whatever was doing the job before — a macro, a template library, a contractor, another licence you are still paying for.
22%
Self-reported time savings are recalled, not measured. Where an organisation has both, the remembered figure runs higher than the recorded one — which is an argument for instrumenting the work, not for distrusting the people.
65%
An hour returned to somebody with a queue behind them is value. An hour returned to somebody without one is slack — real, humane, and not a euro. This is the deduction finance asks for first and programmes model last.
DeductionRemovedRunning total
Where Act 1 left itSeats × hours × rate, over 46 weeks.——
The slope you already hadThe trend that predates the rollout.——
Displaced, not createdSaved against what was doing the job before.——
Remembered, not recordedThe gap between the recalled hour and the measured one.——
Hours that did not land on workReturned time with no queue behind it.——
Attributable gross valueBefore a single cost is counted.—
The working 
—
Why these compound rather than add Each deduction takes its share of what is left after the one before it, not of the original total. That matters: added together, four deductions of 18, 12, 22 and 35 per cent would remove 87% and the arithmetic would start producing negative hours at settings a real organisation could easily hold. Taken in sequence they cannot, because each is answering a question about a smaller and better-defined quantity than the last. The order is the argument, and the running total column is there so you can see the order doing its work.
Act 3 · The cost side

The cost nobody counts is the one that is made of people

Licence and inference are on an invoice, so they turn up in every business case. The third cost does not arrive on an invoice: it is the human time spent reading the output, correcting it, and running it again. It is charged at exactly the loaded rate you used to value the savings — which is why it is so often the largest number in this section.

Act 3What it costs to run
€360
Charged on every seat, including the ones that never open it.
€210
Tokens, calls, retrieval — the part of the bill that grows precisely when the tool is working.
9 min
Reading it, correcting it, re-prompting it, and the second pass when the first was wrong. Applied to every assisted hour, not only the attributable ones — you review the output regardless of who ends up getting the credit for it.
€95,000
Enablement, prompt libraries, the people who run the rollout, the governance the rollout needed.
CostBasisPer year
LicenceEvery seat, used or not.——
UsageActive people only.——
Checking the workHuman minutes, at the same loaded rate as the savings.——
Running the programmeEnablement, administration, governance.——
Total cost to runThe denominator every verdict below is measured against.—
Attributable value, less what it costs to run — —
The working 
—
One number in this act is doing something the others are not Checking the work is priced at the loaded rate, on the gross assisted hours, because that is what it is: people's time. Everything else here scales with seats or with usage, so it is predictable and it is on somebody's invoice. Review time scales with output, is on nobody's invoice, and is the one line where a small change in the input moves the verdict more than a hundred more seats would. Act 5 measures that rather than asserting it.
Act 4 · The evidence split

What can you actually point at?

Every euro of cost above is evidence-linked whether you like it or not — there is an invoice, a licence count and a usage record behind each one. The value side is not. Some of it traces to something a system wrote down; the rest traces to a number somebody typed. They spend identically and they defend completely differently, and the asymmetry is the reason programmes lose arguments they were arithmetically winning.

Act 4Evidence-linked, or asserted
42%
A ticket that closed, a document that reached a signature, a cycle time in a tracker, a call that did not come back. Not a survey response, not a workshop estimate, not a confident sentence in a steering pack.
Evidence-linked — Assumed —
The part you could defend: evidence-linked value, less the full cost — —
The working 
—
This is the bridge to the rest of the suite An evidence-linked delta is one you can defend to a sceptic, an internal auditor or a board that has read the research above. An assumption is one you can only assert, and asserting it harder is not a strategy. The instrumentation that turns the second into the first is usually already deployed — enterprises are putting gateways in front of AI traffic for governance and liability reasons anyway. The risk office paid for the pipes; the finance office is entitled to read them. That is the argument Liftward is built on, and it is the one thing on this page that is a product claim rather than an arithmetic one.
Act 5 · The decision

Scale, fix or kill — on the same workflow

The number exists to serve a decision, so here is the decision, taken three times on the inputs you set. Nothing about the workflow changes between these three panels. What changes is which number the room was working from. All three verdicts are arithmetically correct, which is exactly why the choice of number is not a technicality.

If the room believed the board-deck number — — — Counting only the costs that arrived on an invoice, against a saving nobody subtracted anything from. This is the business case that gets approved in the room and withdrawn eighteen months later.
If the room believed the baselined number — — — Attributable value after Act 2, against every cost including the human time in Act 3. This is the number a finance business partner would recognise, and the one worth arguing about.
If the room believed only what it could evidence — — — Only the part of the value with a record behind it, against the full cost — every euro of which has a record behind it. The gap between this panel and the one beside it is the size of your instrumentation problem.
What the third panel is and is not saying A negative defensible number is not a finding that the tool does not work. It is a finding that you cannot yet show whether it does — which is a different problem with a different fix, and a far cheaper one. The instruction it implies is not "kill"; it is instrument the part you cannot see, then decide. Killing a workflow you never measured is the same mistake as scaling one you never measured, made in the other direction.
And where the leverage actually is

The four levers, ranked by what they move

Each row re-runs the whole model with one input nudged and reports the difference. The order is computed, not chosen — reload with different inputs and it reorders itself.

More seats is a lever. It is very rarely the biggest one, and it is the only one on this list that a vendor is motivated to bring up.

Where it lands

Six industries, and what an honest number asks each of them

The arithmetic above is deliberately generic. The questions are not. Pick the business you recognise, and the panel works through one workflow: what it does, the naive figure somebody has already written into a deck for it, the questions that figure cannot answer, and the record that would be needed to answer them.

Read these as illustrations, not as findings The workflows are plausible ones, written by hand, chosen because the naive figure attached to each is one people genuinely publish. No customer, engagement or dataset stands behind any of them, and the naive figures are not claims about what tools achieve — they are examples of the shape of a claim. What is worth taking from a panel is the question column: those are the same four questions in six costumes, which is rather the point.
Reality & roadmap

What is running, what is next, and what we will never claim

This page is the thinnest thing in the estate and we would rather you heard that from us than worked it out. Each row says what is true on this page today and what the platform behind it is being built to do. Nothing here carries a date, because Liftward has not published one.

Where the line is

Everything you just used above is arithmetic running in your browser on numbers you typed. The rows below are about what surrounds it.

running now
The arithmetic on this page. Every figure recomputed from your inputs, with the working printed beside it and each recomputation logged in the console in the corner. There is nothing else: no server, no fetch, no stored state, no analytics. The claim that this page cannot leak anything is not a policy — it is a consequence of it holding nothing and calling nobody.
by design
It measures nothing about you. A demonstration that pretended to read your estate would be a better demo and a worse product introduction, because the first thing this product has to be trusted about is the difference between a record and an estimate. So the page draws that line on itself first: you typed every number above, and the page says so four times before it shows you a total.
in build
Baselines read from the work itself. The deductions in Act 2 are the product's actual job, and on this page they are four sliders you set by judgement. In the platform they are measured: each team and each workflow against its own history, read from the trackers, code platforms and service desks where delivery already lives, kept current as the work changes. Next · the honest before, without a survey A baseline that is measured and dated rather than recalled — which is the single input on this page most likely to be wrong, and the one nobody can supply from memory.
next
Evidence-linked deltas. Act 4 asks you to estimate what share of the value has a record behind it. That estimate is the least satisfying thing on this page, and deliberately so — in the platform it is not an estimate. A delta either traces to a captured event or it is filed as an assumption and named as one, in the same ledger, in front of the same people. Next · the split, computed rather than guessed The same tamper-evident records that already serve governance and audit, read a second time for what they say about value.
next
The panel for what the data cannot say. Plenty of value does not announce itself in a metric — a queue that stopped escalating, a decision that stopped being reopened. A panel of independent judges weighs that evidence, argues both sides and renders a verdict with its reasoning attached, including the dissent. Next · after the ledger it reasons over The panel proposes. Your people decide — and the reasoning travels with the verdict so they can disagree with it on the merits.
never
We will not tell you your AI paid off. Liftward measures value; it does not promise it. A verdict on this page is arithmetic on numbers you invented, and a verdict in the platform will be arithmetic on records you own. Neither is a forecast, neither is advice, and neither is a substitute for the judgement of the people who run the work. A product that told you the answer you wanted would be easier to sell and worth nothing in the room where it matters.