What should we record before starting?

How long the task takes now, how often it happens, who does it, and what currently goes wrong. Four facts, measured over a week rather than estimated, are enough to make a later comparison meaningful. That week is the single highest-value thing available in this subject and it has to happen before anything changes.

How do we count time saved credibly?

Say where it went. Time freed becomes money when a role is not backfilled, when work previously turned away is taken on, or when overtime stops. Where none of those applies, describe it as freed capacity and say what it was used for, which is honest and survives scrutiny better than a converted figure.

What period should we measure over?

A quarter is usually the shortest honest window, because a first month has setup paid for and enthusiasm running. Where the tool carries a renewal or a significant change of process, measuring across that is better. Anything claimed over a longer horizon than your evidence covers should be labelled as an expectation.

How do we handle benefits that are not time?

Name them without converting them. Fewer errors reaching a client, faster response, work that happens at all rather than being deferred: these are frequently the larger half and they resist a figure. Listing them as benefits is more defensible than inventing an amount that rests on an assumption nobody can check.

Who should produce the number?

Somebody close enough to the work to know what actually changed, with review from whoever holds the budget. Produced only by the team that championed the tool it tends to be generous; produced only by finance it tends to miss the benefits that are not cost. The disagreement between the two is usually where the real figure sits.

Should the calculation be able to say no?

Yes, and deciding the threshold in advance is what makes it a measurement rather than a ritual. A calculation that can only justify continuing tells you nothing and is recognised as such by anybody sceptical. Naming beforehand what result would make you stop is also what makes a positive result believable.

Glossary

AI ROIAI return on investment

AI return on investment is what an AI investment gives back measured against everything it costs, counted over a period long enough to include the effort of adopting and running it.

In plain terms

Somebody will ask what you got for the money, usually about nine months in and usually in a meeting. The honest answer needs two things almost nobody has: a record of how long the work took before, and an account of what the tool really cost including the effort around it. Without the first you have an anecdote, and without the second you have a licence fee compared against a feeling.

Why it matters

Because the question decides whether the work continues, and it is nearly always answered badly in both directions. Teams that cannot show a return lose funding for work that was genuinely paying off, and teams that produce a confident figure built on hours saved often find it does not survive contact with anybody who asks where the hours went. The difference between those outcomes is mostly whether somebody spent an afternoon recording a baseline before starting, which is the cheapest thing in this entire subject and the most commonly skipped.

How it works

Record the baseline before you start, because it cannot be reconstructed afterwards. How long the task takes now, how often it happens, who does it and what goes wrong: an afternoon of measurement before anything changes is worth more than any amount of estimation later. Teams that skip it are not left with a weaker answer, they are left with no answer, because there is nothing to compare against.

Count the whole cost, not the licence. Setup, wiring into your systems, training, the person who ends up owning it and the time spent reviewing output all belong on the cost side. A return calculated against the subscription alone is comparing a real benefit against a fraction of the real cost, which flatters the result in a way that tends to be discovered by somebody less friendly later.

Be specific about where the saved time goes, because that is where most claims fail. Twenty minutes saved on a task somebody does twice a day is real; twenty minutes saved across forty people who each absorb it into their day is not money unless something else changed. The credible version names what the freed capacity was used for, or it counts the benefit as quality rather than as cost.

Value that is not time is often the larger half and is harder to claim. Faster response to customers, fewer errors reaching a client, work that gets done at all rather than being deferred, a specialist freed from routine drafting: these are real and they resist a simple figure. Naming them as benefits without inventing a number for them is more defensible than converting them into pounds with an assumption nobody can check.

Choose a period that includes the awkward parts. A first month flatters the tool because setup has been paid and enthusiasm is high; a first quarter is usually the shortest honest window, and anything with a renewal in it is better. Returns that only appear over a horizon longer than your evidence should be described as expectations rather than results.

Attribute carefully, because several things usually changed at once. A team that adopted a tool also reorganised, hired, or simply got better at the work through practice. Claiming the whole improvement for the tool is the commonest overstatement, and the defence is to say plainly what else was going on rather than to hope nobody asks.

Decide in advance what result would make you stop. A return calculation that can only justify continuing is not a measurement, it is a ritual. Naming the threshold beforehand is what makes the exercise capable of telling you something, and it is also the thing that makes a positive result believable to somebody who was sceptical.

Two return claims

The difference between these is not rigour for its own sake, it is whether the claim can withstand one sceptical question. The left version fails on all three lines at once: nobody can verify ten hours because nothing was measured before, the comparison omits most of the cost, and the baseline is a recollection formed by people who now work differently. It will be believed by anybody already convinced and by nobody else, which makes it useless in exactly the meeting where it matters. The right version is less impressive and much harder to dismiss. It names what was measured, when, and against what; it puts the real cost on the other side; and crucially it says where the freed time went, which is the line most claims omit and the one a finance colleague asks about first. Notice that producing the right version costs an afternoon before the work starts and is impossible to produce afterwards at any price.

Seen in the wild

Time a routine drafting task by hand for a week before introducing an assistant, so the comparison has something to sit against.
ChatGPT
Count how often people currently fail to find an internal answer, before deploying search that promises to fix it.
Glean
Record the volume and handling time of the records an automation will process, since that is the number the return depends on.
n8n
Note what a specialist currently spends on routine summarising, which is the capacity a tool would free rather than the money it saves.
Notion AI

Common misconceptions

People assume

We can work the return out afterwards.

In fact

Almost never, because the baseline was never recorded and memory reconstructs it to fit the conclusion people already hold. What gets produced instead is an estimate of how long things used to take, made by people who now do it differently, which persuades nobody who was not already persuaded.

People assume

Hours saved multiplied by a salary rate is the answer.

In fact

It is the first line of the answer and it is where most claims fall over. Time saved becomes money only when something changes: a role not backfilled, work taken on that was previously turned away, overtime that stops. Absent one of those, the honest description is freed capacity rather than saved cost.

People assume

A negative result means the project failed.

In fact

It frequently means the use case was wrong rather than the technology, and finding that out cheaply is a good outcome. The failure worth avoiding is not a negative number; it is being unable to produce any number, which is what leaves a decision to whoever is most confident in the room.

Telling them apart

AI ROI vs Total cost of ownership

AI ROI

Both sides of the ledger: what came back set against what went in.

Total cost of ownership

One side of it, done properly: everything the tool costs across its life.

Cost of ownership is an input to the return, and doing it badly is the commonest way a return is overstated.

Questions

What should we record before starting?: How long the task takes now, how often it happens, who does it, and what currently goes wrong. Four facts, measured over a week rather than estimated, are enough to make a later comparison meaningful. That week is the single highest-value thing available in this subject and it has to happen before anything changes.
How do we count time saved credibly?: Say where it went. Time freed becomes money when a role is not backfilled, when work previously turned away is taken on, or when overtime stops. Where none of those applies, describe it as freed capacity and say what it was used for, which is honest and survives scrutiny better than a converted figure.
What period should we measure over?: A quarter is usually the shortest honest window, because a first month has setup paid for and enthusiasm running. Where the tool carries a renewal or a significant change of process, measuring across that is better. Anything claimed over a longer horizon than your evidence covers should be labelled as an expectation.
How do we handle benefits that are not time?: Name them without converting them. Fewer errors reaching a client, faster response, work that happens at all rather than being deferred: these are frequently the larger half and they resist a figure. Listing them as benefits is more defensible than inventing an amount that rests on an assumption nobody can check.
Who should produce the number?: Somebody close enough to the work to know what actually changed, with review from whoever holds the budget. Produced only by the team that championed the tool it tends to be generous; produced only by finance it tends to miss the benefits that are not cost. The disagreement between the two is usually where the real figure sits.
Should the calculation be able to say no?: Yes, and deciding the threshold in advance is what makes it a measurement rather than a ritual. A calculation that can only justify continuing tells you nothing and is recognised as such by anybody sceptical. Naming beforehand what result would make you stop is also what makes a positive result believable.

Key takeaways

Record the baseline before starting; it cannot be reconstructed afterwards.
Count the whole cost, not the licence, or the return is measured against a fraction.
Time saved becomes money only when something else changed; otherwise it is freed capacity.
Benefits that are not time are often larger and better named than converted.
Decide in advance what result would make you stop, or it is not a measurement.

Tools that use this

ChatGPT
Timing a routine drafting task by hand before the assistant arrives.
Glean
Counting failed internal searches before deploying something that promises to fix them.
n8n
Volume and handling time of records, which is what an automation's return rests on.