AI ROI for UK Businesses: Measuring Impact Beyond the Pilot

Measuring AI's return on investment (ROI) in businesses has become the crucial question distinguishing those who merely tested a tool from those who transformed their operations. AI has moved beyond the pilot phase: teams use it, like it, and report working faster. Yet, in most cases, the bottom line hasn't seen the impact. This guide explains the source of this disconnect, how to calculate the cost before automating, and when it's not worth the investment.

TL;DR

• The figures cited at the AI Brasil Experience 2026 come from McKinsey's State of AI 2026 report: approximately 80% of respondents say AI has improved individual productivity, but only 37% attribute any positive contribution to EBIT.
• Personal productivity doesn't automatically translate into profit. Gains appear when AI is integrated into a process with clear ownership, metrics, and integration with existing systems.
• To measure AI ROI in your business, you need a baseline beforehand, a process metric, the total cost (not just the subscription), and a rule to convert time into value.
• Start with high-volume, clearly defined processes: customer service, repetitive document handling, and system integrations.
• If the process is infrequent, poorly defined, or lacks reliable data, do not automate yet.

The source of the 80% and 37% figures (we checked)

The press release for the AI Brasil Experience 2026, taking place on 28 and 29 October at Distrito Anhembi, São Paulo, references a "global McKinsey study". The event's text doesn't specify the title or sample size, so we investigated.

The source is The State of AI report, published by McKinsey on 25 August 2026, based on 1,719 participants from 97 countries. According to ITPro's coverage:

• approximately 80% say AI has improved individual productivity;
• 37% report some positive contribution from AI to EBIT (operating profit), virtually the same level as the previous year;
• around 20% say the operational cost of AI has already limited its use.

Two caveats before repeating these figures in a meeting. Firstly: it's respondents' perception, not audited accounts. Secondly: the 37% state there was an impact, not how much. It could be a small gain.

Local data confirms the pattern

Local data on AI ROI in Brazil confirms this pattern. A study by TOTVS with MIT Sloan Management Review Brazil, presented on 13 October 2026 with 583 participants, found 80% of organisations using AI in some form, but 88% not measuring the financial outcome and 86% not tracking AI costs (TI Inside). Only 8% reported extensive AI integration with management systems.

And at the top of the pyramid, it's no different: in PwC's Global CEO Survey 2026, with 4,454 CEOs from 95 countries, 56% said they hadn't seen significant financial benefit from AI so far. Only 12% saw gains in both cost and revenue simultaneously.

Why AI ROI doesn't appear in the bottom line

Over nearly two decades of developing systems, I've seen the same story with spreadsheets, ERPs, and now with AI: the tool improves the task, but no one changes the process. In practice, there are four reasons.

1. Time saved isn't money earned

If an analyst writes an email in 2 minutes instead of 10, they've saved 8 minutes. The company only gains something if those minutes translate into more customer interactions, fewer overtime hours, an avoided hire, or a shorter lead time for the client. Without this defined purpose, the time simply vanishes during the day.

2. Individual use, without a defined process

Each person uses AI their own way, on their own account, with their own prompt. There's no standard, no metrics, and often no authorisation. The TOTVS study shows 40% of companies have AI initiatives without cross-departmental coordination. An AI usage policy within the company addresses the risk aspect, but it doesn't create results; it merely tidies things up.

3. AI doesn't integrate with existing systems

An assistant that reads a document but doesn't save anything to the ERP, CRM, or Moodle simply pushes the work back to a human for copying and pasting. The bottleneck remains; it has just shifted. This is why integration often has a greater impact on results than the chosen model.

4. No one measured the "before"

Without a baseline, any subsequent figures are merely opinions, and AI ROI becomes a matter of faith. "Customer service got faster" doesn't justify a budget. "Average first response time dropped from 4 hours to 12 minutes" does.

How to measure AI ROI in 5 steps

The calculation doesn't require expensive consultancy. It requires discipline. This is the roadmap we follow before proposing any automation.

Step 1: Choose a process, not a tool

AI ROI begins with the question "which process is painful?", not "where can we use AI?". Good candidates have high volume, clear rules, and visible costs: answering repetitive queries, triaging documents, or entering data from one system into another.

Step 2: Record the baseline

For 2 to 4 weeks, measure the process as it is today. Monthly volume, time per item, error or rework rate, time until the customer receives a response. If the data isn't available in a system, a simple spreadsheet filled out by the team will suffice.

Step 3: Define process and financial metrics

Each process needs two interlinked metrics:

• Process metric: what AI directly changes (response time, items processed per hour, error rate).
• Financial metric: how this change translates into money (fewer paid hours, avoided hiring, sales that are no longer lost, avoided penalties).

If you can't articulate the second, the project isn't ready for investment yet.

Step 4: Calculate the total cost

The subscription is the smallest part. Also included are development and integration, API consumption (tokens and messages), hosting, monthly maintenance, human review time, and team training. We discuss this in detail below.

Step 5: Compare with the baseline and decide

The classic formula for calculating AI ROI works well:

ROI = (measurable gain over period − total cost over period) ÷ total cost over period

Run for 60 to 90 days, compare with the baseline, and decide: scale, adjust, or decommission. Decommissioning is also a result. It prevents a pilot from becoming an eternal fixed cost.

An example with figures (illustrative)

The figures below are hypothetical, purely to illustrate the mechanics:

• A company receives 1,200 WhatsApp enquiries per month, each averaging 6 minutes of human work.
• An AI assistant resolves 35% of these independently: 420 enquiries, or 42 hours per month.
• At £35 per hour in agent cost, this amounts to £1,470 per month.
• The monthly operational cost, including API, hosting, maintenance, and implementation amortised over 12 months, comes to £1,900.

Result: negative ROI. Even with a satisfied team. The calculation only makes sense if the 42 hours have a purpose (for example, avoiding hiring an additional agent during peak times, or responding outside business hours and recovering sales currently being lost). This is AI ROI seen up close: the same disconnect between the 80% and 37%, on a small business scale.

Metrics by process: what to track

This table summarises what to measure in the three processes where we see the most AI ROI in small and medium-sized enterprises (SMEs):

Process	Process Metric	Financial Metric	Data Source
Customer Service (WhatsApp, website)	First response time, % resolved without human intervention, enquiries per agent	Avoided service hours, out-of-hours lead conversion	Customer service platform, CRM
Documents (contracts, invoices, certificates, forms)	Time per document, error rate, rework	Back-office hours, avoided penalties and delays	Management system, protocol, control spreadsheet
System Integrations	Manual entries eliminated, system discrepancies	Data entry hours, avoided billing errors	Integration logs, ERP
EAD and Moodle	Student support tickets, grading and feedback time	Tutoring and administrative hours, student attrition	Moodle reports, helpdesk

Notice that no line measures "tool usage". Number of prompts and active users indicates adoption, not return.

Customer service, documents, and integrations: where automation becomes a process

Customer service: AI as the first tier, not the entire service

In WhatsApp customer service, ROI appears when AI handles repetitive queries (hours, prices, order status, duplicate invoices) and escalates to a human, with full history, anything outside the rules. This requires Meta's official API, an up-to-date knowledge base, and clear escalation rules. We detail the technical and policy limits in how to automate WhatsApp customer service with AI without getting banned.

A common pitfall: measuring only the "% resolved by the bot". If the customer gives up instead of being served, the number goes up, but sales fall. Track abandonment rate and satisfaction alongside this.

Documents: extraction and verification, with human review where risk exists

Reading PDFs, extracting fields, classifying, and verifying against a rule is work that AI does well, provided there's human sampling at the outset. In Moodle environments we manage, the volume of repetitive administrative tasks (e.g., verifying enrolment documents, answering common student queries, issuing and resending certificates) is the type of routine where automation typically pays off faster than in creative tasks.

Integrations: where overlooked savings reside

The most lucrative automation isn't always "AI" in the trendy sense. It's system A communicating with system B without manual data entry in between. When AI is involved, it handles the unstructured part (understanding an email, an audio message, or a free-text request) and delivers structured data for the integration to record. If you use an ERP and WhatsApp, it's worth reading our guide on how to integrate ERP with WhatsApp, including architecture and costs.

This type of project, combining AI, APIs, and legacy systems, is what we do in our bespoke system development service.

What it really costs: the total cost your spreadsheet forgets

An honest calculation of AI ROI considers all the items below, not just the tool's monthly fee:

1. Development and integration: connecting AI to your ERP, CRM, Moodle, or database. This is usually the largest initial cost.
2. API consumption: language models charge per token (piece of text processed). High volume, long context, and larger models multiply the cost.
3. Messaging costs: on Meta's official WhatsApp API, Meta charges per template message delivered, with varying rates by country and category. Rates change; always consult Meta's official pricing page.
4. Infrastructure: server, database, queues, backup, and monitoring. An automation that crashes at 11 pm unnoticed can be costly.
5. Maintenance: APIs change versions, models are discontinued, business rules evolve. Allocate monthly hours.
6. Human review: in the initial weeks, someone checks the outputs. This time needs to be factored into the cost.
7. Governance and UK GDPR: defining which data can be sent to which provider and documenting this.

McKinsey's report itself points out that around 20% of companies have already had their AI usage limited by operational cost. Those who don't estimate the cost beforehand will discover it on the bill.

When NOT to automate with AI

Knowing when to say no protects your AI ROI more than any model. We do not recommend automating with AI when:

• The process is infrequent. Ten occurrences per month won't justify the development or maintenance.
• The process is undefined. If each person does it differently, AI will automate the chaos. Standardise first.
• The error is costly and hard to detect. Legal opinions, expert reports, credit decisions without review: AI can support, but not decide independently.
• There's unreliable data. An outdated knowledge base generates confident but incorrect responses.
• A simple rule suffices. If an "if X, then Y" in the system solves it, there's no need for a language model. It's cheaper, faster, and more predictable.
• No one takes ownership. Unowned automation degrades silently.

Common pitfalls when measuring

These are the most common reasons AI ROI fails when a client arrives saying "AI hasn't delivered results":

• Measuring adoption instead of results. Active licences and prompts per day don't pay the bills.
• Comparing with memory, not a baseline. "It used to take ages" isn't a figure.
• Ignoring review costs. If a human reviews 100% of outputs indefinitely, the gain is minimal.
• Attributing everything to AI. If in the same quarter you hired, changed pricing, and launched a campaign, isolate the effect with a comparison group or a stable period.
• Scaling before proving. Expanding an unprofitable pilot to ten areas only multiplies the losses.
• Leaving the pilot running. Tools in testing become recurring subscriptions without an owner. Review quarterly.

For those who have moved past the testing phase and want to deploy agents, the roadmap for AI agents for businesses: from pilot to production complements this guide with governance and implementation details.

How Agathas Web solves this

Since 2008, Agathas Web has developed systems, integrations, and infrastructure for businesses and educational institutions. Today, we operate servers and managed hosting, Moodle environments, and integrations with Meta's official WhatsApp API. When it comes to AI, our work starts with the business case, not the tool.

What's included in the project:

• Process diagnosis: we map the current workflow, volume, and cost, and establish the baseline with you.
• AI ROI estimation before coding: the same calculation from this guide, using your figures. If it doesn't add up, we'll tell you before you invest.
• Development and integration: we connect AI to your existing systems (ERP, CRM, Moodle, WhatsApp via official API, databases, and third-party APIs).
• Metrics dashboard: process and financial metrics tracked from day one.
• Hosting, monitoring, and maintenance: the automation runs on infrastructure we manage, with backup and alerts.

How it works in practice: we start with a single process, with a closed scope and a 60 to 90-day period for measurement. If the figures stack up, we scale to the next. If they don't, we adjust or decommission, without a long contract tying you to an unprofitable pilot.

How to get started: speak to our team via the software development and integrations page, tell us which process consumes the most time today, and we'll put together an initial diagnosis.

Conclusion: from personal gain to bottom-line results

The gap between the 80% who feel AI helps and the 37% who see a profit impact isn't a lack of technology. It's a lack of process, integration, and prior calculation. AI ROI appears when you choose a painful process, measure the before, calculate the entire cost, and assign a purpose to the freed-up time.

Start small, with one process and one metric: AI ROI is built this way, one workflow at a time. If you want to do this calculation with those who develop and operate the system day-to-day, explore our bespoke development service and bring us your case.