Marketing Automation ROI: How to Measure It Without the Vendor's Number

The marketing automation ROI figure on the first page of results for that query is not a prediction about your business. I have watched it travel for four years, re-dated, re-attributed and re-sold as a benchmark for accounts it never measured, and the pages repeating it copy each other rather than a source. The only figure that survives your finance review is the one you can read off your own workflows, and almost nobody shows you how to build it. Would a field that wanted you to measure anything keep its source in a footnote? Your number will be smaller than the borrowed one, and it will be the first number in this discussion you can actually defend.

The test takes four minutes. Search the query in a private window and read the return figure on each page in the top ten. Several print the same value: $5.44 returned for every $1 spent. Two name different sources for it. One dates it to 2024. Another presents a three-year average as a first-year outcome. None shows the arithmetic, and none was measuring your account. What your programme returns depends on your workflows, your invoices and your attribution rule, which is why the next section starts with the document everyone is quoting.

Where the Standard ROI Figure Comes From

The figure ends in one document: Nucleus Research, research note V61, published 26 March 2021 under the headline "Marketing automation returns $5.44 for every dollar spent".

The note describes its own method, and the method is the part that gets dropped in transit. It reviewed 16 ROI case studies published between 2016 and 2020 and averaged the outcome: $5.44 returned per dollar over the first three years after deployment, with payback under six months. The drivers it names are marketer productivity, campaign efficiency, cost savings from automation and simplified infrastructure, and better prospect engagement. Three years, sixteen deployments, written up because they worked.

Now watch what happens as that document travels. Reading the pages ranking for this query on 10 October 2026, the trail is public:

• One of the two current editorial results in this field states the 5.44 claim and links out to a statistics page rather than to the note itself.
• That statistics page, dated 15 May 2026, prints 5.44 as its first statistic, then attributes a newer $6.10 per $1 to a Forrester Research study of 1,200 mid-to-large enterprises, described there as a 12.1% increase on what it calls the 2025 benchmark of 5.44. The same page credits a 12.2% operational cost reduction to a Deloitte benchmark of 3,100 companies. Another page gives the identical 14.5% and 12.2% pair to Nucleus.
• One page dates the Nucleus note to 2024. Another, published February 2026, presents the three-year average as a first-year result.

One document, four attributions, two re-datings, and at least one page turning a three-year figure into a twelve-month one. That does not make the original note dishonest. It makes the field's citation practice unreadable, because a three-year average of sixteen self-selected deployments now reads as a general expectation for a business it never saw.

That is the surface your own number competes against. If you are choosing a platform rather than measuring one, the marketing automation tools cluster covers that decision. The arithmetic below starts with three inputs you can read without help.

The Three Inputs You Can Actually Measure

A platform's own ROI panel is not a neutral instrument, because the same product sells you the platform. It can count what happened inside it and nothing else: the email it sent, the form it captured, the deal your CRM later closed. Fine as an operational view, poor as a basis for a return claim. The three inputs below live outside it.

Time removed from a real workflow

This is the input teams overclaim most often, and the fix is a rule rather than a spreadsheet. Count hours only where the workflow existed before the automation went in, so you have a before and after for the same work.

Pick one sequence, say a new lead follow-up. Count touches per item per week before it was automated and after it settled, then multiply the difference by a loaded hourly cost you already know from payroll, not a market rate for a title you do not employ. Apply the honest test: if those hours came back as capacity you refilled with other work, you have a capacity input, not a saving. Do not book it as both.

Pipeline touched by automated sequences

Touched is not influenced, and influenced is not closed. Ask one question of your CRM over a fixed period: how many opportunities that closed in it had at least one automated touch before the close date? That count is real and you can defend it.

What you cannot defend is the leap from there to a return multiple, because those same opportunities also carried calls, a demo, a referral and whatever your buyer read on their own time. The defensible version is a sentence: automation touched this share of the pipeline that closed, under this attribution rule. The tooling for that count sits in your marketing analytics tools, and the outcomes you are counting are what marketing KPIs define.

Cost avoided, not revenue gained

The cost side is never the subscription line alone, and this is where ROI calculations quietly cheat most often. Published third-party estimates of the non-subscription cost run as follows, dated 2026 and not verified by us: setup and onboarding from $1,000 to $3,000 at the basic end, $3,000 to $10,000 once data migration is included, $10,000 to $50,000 at enterprise scope, with onboarding landing between 6 and 12 months of subscription fees. Advanced API connections run $500 to $2,000 each, custom middleware $5,000 to $20,000, and ongoing training, integrations and support add 15 to 25% of subscription cost every year.

Then there is the tier you land on. Our own final analysis of HubSpot, completed this year, shows Starter at around $20 per month against Marketing Hub Professional at roughly $800 to $900 and Enterprise at roughly $3,200 to $3,600, with per-seat pricing on the upper tiers and 3 to 6 months to full adoption. The mechanisms behind those gaps are the marketing automation pricing question and belong to that article. What matters here is simpler: your cost side is an invoice, an onboarding project, an integration and the internal hours to run the thing, plus the tools you now duplicate. That overlap is where integration cost hides in a martech stack.

Cost avoided means what you would otherwise have paid a person or a second tool for the same work. Read it from invoices and payroll, not from a pricing page.

Why Attribution Makes Most Automation ROI Claims Wrong

Before you quote any return number, declare the rule that decided who gets the credit. Almost no marketing team does, and that rule moves the answer by tens of percent with nothing changing in the account.

Here is what the tools in a typical stack will and will not do. Google Analytics 4 data-driven attribution cannot be inspected or customised, it sees Google-ecosystem touchpoints only, with no CRM records, no phone calls and no offline events, and it falls back to cross-channel last-click below 600 monthly conversions, which is the band most B2B teams sit in. Common setups cap the attribution lookback at 90 days, truncating any cycle longer than that, and multi-touch models usually count only touchpoints inside a 7, 14 or 30 day window. Every model also carries credit assumptions baked in before it sees your data: a linear model splits credit evenly across touchpoints, a position-based model hands most of it to the first and the last. Change the model, and the number moves while the business stands still.

What no model sees is a longer list: dark social shares, word of mouth, offline events, cross-device journeys and impressions inside walled gardens that never hand back a click. The measurement field is not confident in its own instruments either. The industry's own State of Attribution research reports that a majority of the organisations interviewed would not recommend their current attribution provider.

There is a usable gate before you trust any multi-touch figure. Ask two questions about your own data: can more than half of your conversions be tied to a deterministic user id, and do more than 60% of the interactions in a typical journey happen in trackable digital channels? If either answer is no, the model will over-credit the small slice it can see and under-credit everything else. That is the condition your report inherits, whether or not the report says so.

Building the Number, Step by Step

Six steps, in this order. Steps one and six are the ones that keep the result defensible.

1. Fix the period and the counterfactual first. One completed period, the same length as the comparison period, plus what you would have done without the automation. A number without a counterfactual measures activity.
2. Read the cost side off invoices. Subscription, onboarding amortised over a stated number of months, one-off integration work, and the internal hours to build and run it.
3. Count time removal only where a before exists. One named workflow, touch counts before and after, a loaded payroll rate, and the capacity-versus-saving distinction applied honestly.
4. Count pipeline touched, and separate it from pipeline influenced. One CRM query, one fixed period, one sentence naming which opportunities carried an automated touch.
5. Declare the attribution rule before you quote the number. Name the model, the lookback window, and whether you sit above or below the conversion threshold where the model degrades.
6. Subtract what the automation has not removed. A sequence that runs because it exists spends money and buys nothing. If two automated touches do the job one manual touch used to, the second one is cost.

Write the result as a sentence with its rule attached, not as a single multiple: automation touched this share of the pipeline that closed in the quarter, the cost side including onboarding was this much, and the attribution basis was a stated model over a stated window. Then divide the return by the cost side you read off invoices rather than by the subscription line. If the result sits below 1, the programme is not paying for itself yet, and the useful question is which of the six steps would change that.

This is the shape Allable.ai was built around, for a selfish reason. I run an agency, and our client reporting kept drifting away from what the platform dashboards showed, because the rule behind each figure lived in somebody's head. Keeping the measurement next to the work, in the same files as the automating reporting we already did, forced us to write the rule down before the number, which is the only thing that made the two comparable.

What a Good Result Looks Like at 90 Days

Everything in this section is an assumption, including the direction. I have not measured your account and nobody outside it can tell you what your first quarter returns. Each band names the inputs it assumes, so you can disagree with the band rather than with a benchmark.

• Days 0 to 30, assumption: expect the cost side to be fully visible and the return close to zero in your ledger. The inputs I would assume live by now are the subscription, the onboarding invoice and the first weeks of internal hours. I would not assume any pipeline movement, because the sequences have not been running long enough to earn a reply.
• Days 31 to 60, assumption: time removal is the first readable input, and only for workflows that existed before the automation. Assume a visible effect where you counted touches before and after, and assume nothing for workflows you invented at the same time you automated them.
• Days 61 to 90, assumption: contact-level pipeline effect starts to be readable if your sales cycle is shorter than the window. If your cycle runs longer, assume nothing yet and treat any early signal as partial. Even the figure this field quotes is a three-year average over sixteen deployments, so a first-quarter result below the borrowed multiple is the expected outcome, not a failure.

When Automation ROI Is Genuinely Negative

Negative is a real result and it is cheaper to find than a positive number you cannot defend. These are the shapes it takes.

Automating a broken process makes the mistake faster and more expensive: if the follow-up was not wanted, your sequence now sends it at scale and you pay for every send. Last-click attribution holds the number up while the business falls, because an automated email collects credit for deals that closed on a phone call, and you optimise toward a channel that only looks decisive. Contact and seat growth is the quiet failure, with engagement flat while the metres climb, and the renewal is where you find out. A tier upgrade arrives with per-seat pricing before your team has used the tier below, and that one is decided on your contract page rather than in your dashboard. The last shape is capacity that never became anything: time you reclaimed and immediately refilled with reporting work instead of selling or customer work.

None of these needs a dramatic intervention. Each needs the same six steps, run once.

Conclusion

The marketing automation ROI figure in the deck came from sixteen deployments that closed between 2016 and 2020, written up because they worked, averaged over three years, in a note published in March 2021. Is marketing automation worth it in your account? That is a different question, answered from one period of your own work with the invoices attached. Your number will be smaller.

So here is the question worth keeping: which of your automated sequences would survive being asked what you would have paid for that work by hand? The ones that survive are returning something. The ones that do not are the reason this field prefers to quote somebody else's average.

Observation date: 10 October 2026. The field claims above were read from live pages on that date, and the source note is the one named in the text. Third-party cost estimates carry their source and date in the sentence where they appear. Figures we could not measure on your behalf are labelled assumptions, deliberately and in every case.