AI Productivity vs Headcount: What More Capacity Means

AI productivity is real, but it arrives as hours, and hours only become value when a business decides what they are for. Studies of AI tools and agents show large gains on specific tasks, but those gains do not add up automatically across a business, and the time savings workers report across whole workforces are modest so far. The commercial decision is what the capacity AI creates should buy: lower cost, better service, new insight or growth.

In the first article in this series we looked at how AI is changing what companies offshore; this one turns to the commercial side. Our position: for most growing businesses, the best return on AI is more output per person, and growth will often still need more people to act on that output.

The AI question has changed from 'can we use it?' to 'what value will it create?'

Adoption is no longer the hard part; the harder question is what AI will be worth. Respondents to the Bank of England's Decision Maker Panel, reported in the Bank's July 2026 Monetary Policy Report, expected AI to raise their productivity by around 0.9% a year over the next three years and to reduce their employment by around 0.4% a year. Nearly 90% said AI had made no material difference to their employment over the past three years. Those expectations describe both outcomes at once: somewhat more output and somewhat fewer people.

For smaller businesses the question is also competitive. In PwC's 2026 Global AI Jobs Barometer (June 2026), built on more than a billion job adverts and company data, companies in the sectors most exposed to AI recorded 34% productivity growth between 2018 and 2025, against 24% for the least exposed, and their headcount grew faster too (52% against 36%). That is association, not proof of cause, but in our view the gap shows up in price, turnaround and service.

Three business outcomes: efficiency, insight and revenue

Efficiency is the same work in fewer hours, and it has the strongest evidence. In a field experiment with 758 Boston Consulting Group consultants, run in 2023 and published in Organization Science in March 2026, those randomly given GPT-4 completed 12.2% more of 18 realistic consulting tasks than colleagues without it, and finished them 25.1% more quickly. On a task chosen to sit outside what the tool did well, those using AI were 19 percentage points less likely to reach a correct answer. Agents extend the gains to multi-step work: in Microsoft's study of its early-2026 rollout of AI coding agents to tens of thousands of engineers (July 2026), adopters merged roughly 24% more pull requests than they otherwise would have, though a merged change measures output, not value.

Insight is better decisions from the same data: forecasts refreshed weekly, anomalies flagged before month-end. It is where AI can become a new service. As a thought experiment, an accountancy practice that already holds its clients' bookkeeping data could use AI to draft monthly cash-flow and margin commentary for each client and sell it as advisory work, with a management accountant checking it, adding the judgement and presenting it.

Revenue comes from more sales activity at the same cost, from enhanced or new services that command higher prices or win larger contracts, and from lower prices that win share. In EY's December 2025 survey of 500 US senior leaders, 29% of organisations seeing AI productivity gains had used them to cut prices to gain market share. The studies above measure time and quality, so test any revenue case in your own numbers.

Practical examples by function

These examples are thought experiments showing where AI is likely to add capacity and where people stay responsible.

Finance and accountancy: the VAT return. Since April 2022, VAT-registered businesses have had to keep digital records and file VAT returns through compatible software under Making Tax Digital for VAT, unless HMRC exempts them, and an authorised agent can use software to prepare and send a client's return. Picture a practice that hands the preparation to an AI agent, which reads each client's records, categorises transactions, flags anything unusual and drafts returns for the whole client list in the time an accounts assistant might spend on a handful. The accountant still reviews every return, and each client checks and confirms it before the practice, as the authorised agent, submits it, not least because, as HMRC has made clear, businesses remain responsible for meeting their VAT obligations when an agent acts for them. The practice could cut its cost per return, price more keenly and take on more clients, but every new client brings more flags to check, queries to answer and returns to approve. If you redesign a regulated process like this, take advice first.

In the BCG consultant study, results depended on whether a task suited the tool, which is why the recommendation and sign-off stay with people. For more, see offshore teams for professional services and consulting firms and the best offshore location for a sales and marketing team.

Capacity: what happens to the time and output AI creates?

A ten-person team working 37.5 hours a week each has 375 hours. Twenty per cent more capacity is 75 hours a week, roughly two more full-time people. Treat that 20% as a thought experiment: the time savings workers report so far are smaller. In the US, a nationally representative survey of adults aged 18 to 64 tracked by the Federal Reserve Bank of St. Louis puts self-reported time savings from generative AI at 2.2% of all work hours in the second quarter of 2026. At that average, our ten-person team would gain about eight hours a week, around a ninth of the thought experiment.

Saved time also does not become output by itself: it arrives in fragments across the team. The St. Louis Fed's economists caution that measured productivity rises by less if saved time goes on lower-value tasks, and a European Central Bank analysis (August 2026), based on a survey of about 20,000 people in 11 euro area countries, adds that the boost depends on whether the employer can put the extra capacity to use. More output can also mean more checking. In a study of 802 developers at a company that mandated AI coding tools (July 2026), output per developer reached 2.09 times its pre-mandate level by April 2026, while the load on each code reviewer roughly doubled and automated review overtook human review.

So decide where the hours go before they arrive; each of the four destinations needs something different. Absorbing growth needs demand to fill the capacity and a plan for the extra exceptions. Raising service needs a higher, measured standard. Building insight or a new service needs people with the judgement to interpret and present it. Reducing cost needs evidence that the work has gone rather than moved.

The Bank of England's Agents report firms in some industries slowing hiring or leaving vacancies unfilled so that output can grow without a proportional rise in headcount. In McKinsey's 2026 global survey (August 2026, 1,719 respondents), 14% of respondents from organisations using AI said it had contributed to a fall in their workforce over the past year, less than half the 32% who had expected reductions a year earlier, while 39% now expect declines in the coming year.

When additional people still make commercial sense

AI raises what each person can deliver, which in a growing business often creates more human work than the team can absorb. Five situations can make more people the right answer, provided the extra work is worth more than it costs.

• The output needs acting on. AI can research accounts and draft outreach at a volume that can outrun a team's capacity to call and qualify. Where that extra pipeline is qualified and worth working, it needs more people.
• Exceptions grow with volume. As a thought experiment, if automation clears 90% of 10,000 supplier invoices a month, 1,000 still need a person. Double the volume and that becomes 2,000 at the same automation rate.
• New services need judgement. An AI-assisted advisory line needs management accountants and FP&A specialists to interpret the numbers and stand behind them.
• Customers still want people. Complaints, disputes and complex queries go to a person, and growth brings more of them.
• Someone has to check the output. In the developer study above, review load per person roughly doubled as output rose, so measure oversight needs before you hire; in the BCG study, accuracy fell on work outside the tool's capability.

If AI absorbs your volume and your team can handle the exceptions and oversight, you may not need more people. Where growth creates more human work than your team can take on, an embedded offshore team is one way to add capacity. We build teams employed for your business in five hubs in four countries: Cape Town and Johannesburg in South Africa, Iași in Romania, Bengaluru in India and São Paulo in Brazil. We can build the roles in the table above in any of them, and the right hub depends on a discussion with you about your business.

Under the Embedded Offshore Team Model the offshore team works in your systems and report to your managers, while we handle their employment, the office and local HR support. Labour cost savings typically run at 30 to 50% against the UK in Cape Town and Johannesburg and 40 to 50% against Western European rates in Iași. In São Paulo, operating costs for equivalent roles typically run 30 to 50% below the UK, and in Bengaluru salaries often run 40 to 70% below the UK, especially for mid-to-senior roles.

RedFin shows what added capacity can do for a growing firm. It provides accounting, finance and HR outsourcing to creative, media and technology businesses, and we built its embedded finance team in Iași, starting with a single hire. Julian Davies, RedFin's CEO, sums up the result as "twice the team for the same budget". RedFin doubled its staffing capacity for the same employment cost and grew revenue by 30%, with its UK managers free to focus on strategy and client development while the Iași team handled delivery. See which finance work moves in offshore finance and accounting teams for UK companies, and how to choose the best offshore location for a finance and accounting team.

A simple framework for assessing AI impact

In McKinsey's 2026 survey, the 6% of respondents it classes as AI high performers were twice as likely as others to say their organisations had defined processes to measure the impact of their AI initiatives. These six steps build that measurement one process at a time.

1. Measure the baseline. Record one process's volume, hours, cost per unit, error rate and turnaround before anything changes.
2. Split the process into tasks. Mark which tasks AI can do, which it can help with and which stay with people, then test that boundary.
3. Pilot against a comparison group. Measure output, quality and the full cost, including licences, integration, data preparation, training and review, because self-reported time savings are a weak guide.
4. Decide where the capacity goes before it arrives. Growth, service, insight or cost, with a named owner and a number to hit. Then net it off over a set period: cash saved plus contribution from extra work, minus the full cost. Freed hours count only once they are redeployed.
5. Plan the human work that grows with output. Estimate exceptions, quality checks, customer contacts and sales follow-up at next year's volumes.
6. Choose how to add that capacity. Your existing team, a UK hire or an embedded offshore team, judged on cost, speed and control.

BCG's 10-20-70 rule, restated for agentic AI in September 2026, holds that 10% of the effort should go into the algorithms, 20% into technology and data and 70% into changes to people, organisation and processes. Most of this framework sits in that 70%, where in our view AI projects most often stall; we set out the common causes in why most AI implementation fails.

If your AI plans are creating more work than your team can absorb, book a discovery call to discuss combining AI-enabled productivity with the additional human capacity your growth still needs.

Frequently asked questions

Has AI actually improved productivity?
Yes, on specific tasks. In a randomised experiment run in 2023, consultants given GPT-4 completed 12.2% more tasks, 25.1% more quickly, on work that suited the tool (Organization Science, 2026), and Microsoft engineers who adopted AI coding agents in early 2026 merged roughly 24% more pull requests (July 2026). Task results do not establish business-wide gains, and the time savings people report across whole workforces are modest so far.

How much has productivity increased because of AI?
The evidence so far is mostly self-reported time savings and forecasts, and both are modest. In US survey data from the Federal Reserve Bank of St. Louis, self-reported time savings from generative AI equalled 2.2% of all work hours in the second quarter of 2026. In the Bank of England's July 2026 Monetary Policy Report, UK firms expected AI to raise their productivity by around 0.9% a year over the following three years.

How can a business use AI to increase productivity?
Start with high-volume, rule-based work such as invoice matching, reconciliations, drafting replies, ticket triage and account research. Measure the baseline, pilot against a comparison group and decide in advance where the freed time goes. Keep people responsible for exceptions, approvals and customer conversations.

Does AI reduce headcount?
Not for most UK firms so far. In the Bank of England's July 2026 Monetary Policy Report, nearly 90% of firms in its Decision Maker Panel reported no material impact of AI on their employment over the past three years. The same firms expected AI to reduce their employment by around 0.4% a year over the following three years.

What would 20% more capacity mean for a ten-person team?
As a thought experiment, a ten-person team working 37.5 hours a week has 375 hours, so 20% is 75 hours a week, roughly two full-time people. It becomes value only if the business decides where those hours go: more volume, better service, new analysis or lower cost.

What is the 10-20-70 rule for AI?
It is BCG's rule of thumb that 10% of the effort in AI should go into algorithms, 20% into technology and data and 70% into changes to people, organisation and processes. BCG restated it for AI agents in September 2026.

When should a business hire people even though it uses AI?
When AI output creates more human work than your team can absorb: more leads to follow up, more exceptions as volumes grow, new services that need judgement, customers who want a person and output that needs checking. If your team can absorb it, you may not need to hire.

Sources: Bank of England, Monetary Policy Report (July 2026); PwC, 2026 Global AI Jobs Barometer (June 2026); Dell'Acqua et al., Organization Science (March 2026; experiment run in 2023); Murphy-Hill et al., AI coding agents at Microsoft (arXiv, July 2026); He et al., enterprise AI coding mandate (arXiv, July 2026); St. Louis Fed, FRED Blog (August 2026); St. Louis Fed, generative AI adoption (November 2025); European Central Bank blog (August 2026); McKinsey, State of AI 2026 (August 2026); EY, US AI Pulse Survey (December 2025); HMRC, VAT Notice 700/22; BCG, agentic AI value (September 2026).