Most organisations have stopped asking whether to use AI. The harder question going into 2027 is narrower and more useful: which specific uses actually move output, and what has to be true for them to work?
The evidence is now good enough to answer that properly. Controlled trials, national statistics and large firm surveys all point the same way — AI produces real, measurable gains in particular tasks, under particular conditions, and produces very little when those conditions are missing. What follows is the ten applications with the strongest evidence behind them, each paired with the condition that makes it pay off and a way to measure whether it did.
The short answer
AI is improving business productivity in 2027 through: customer enquiry resolution, administrative load reduction, document drafting and review, decision-ready reporting, multi-step agentic workflows, software delivery, forecasting and planning, onboarding and skill-building, meeting-to-action capture, and the redeployment of freed capacity into work that was previously out of reach.
The last one is the hardest and the most valuable, and it is where most organisations lose the gain.
First, the number that explains everything else
In McKinsey’s 2026 Global Survey on the state of AI — 1,719 respondents across 97 nations, fielded between 4 May and 8 June 2026 — eight in ten said AI had improved their own productivity. In the same survey, 37% could attribute any EBIT impact at all to AI use, a share that did not move from the previous year. Roughly 6% qualified as high performers, attributing at least 5% of EBIT to AI.
Official statistics tell a similar story at national scale. The Office for National Statistics reported that 30% of UK businesses used at least one AI technology in September 2026, up from 9% when the question was first asked in September 2023. Among adopters with ten or more employees, only about 10% described their own use as extensive, and the average adopting business used 1.6 AI technologies. The ONS notes this implies relatively limited transformative impact for most adopting firms to date.
A survey of nearly 6,000 executives across the US, UK, Germany and Australia (NBER Working Paper 34836, Firm Data on AI) found executives reported AI had raised annual labour productivity growth by 0.2% over the preceding three years, and expected 0.6% a year over the next three.
So the technology works at the task level and frequently disappears at the company level. The gap is not mysterious — time saved in one step gets absorbed by queues, approvals and handoffs elsewhere, and licence costs land on the P&L immediately while small savings spread thinly across hundreds of tasks go unmeasured. Brysa’s analysis of why individual AI gains stall before they reach the P&L sets out the mechanics in more detail. Keep that gap in mind reading the list below; it determines which of these ten actually produce a number a finance director will recognise.
10 ways AI is improving business productivity in 2027
1. Resolving customer enquiries at first contact
This is the use case with the cleanest experimental evidence. Brynjolfsson, Li and Raymond studied the staggered rollout of a generative AI assistant across 5,179 customer support agents, published in the Quarterly Journal of Economics in 2025. Issues resolved per hour rose 14% on average, with a 34% improvement among novice and lower-skilled agents and minimal effect on the most experienced. Customer sentiment improved and staff retention rose.
Condition: the assistant has to be grounded in the organisation’s own resolved-case history, not generic knowledge. The gain came from circulating what the best agents already knew.
Measure: issues resolved per agent hour, first-contact resolution rate, and escalation rate — not deflection volume.
2. Clearing administrative load from around real work
The ONS asked UK businesses using AI which roles were most affected. In September 2026, administrative and clerical roles topped the list at 38%, up three percentage points from June and nine points since the question was introduced in December 2025. Creative and design roles came second at 24%.
That is where the time currently goes: data entry, record updates, scheduling, chasing, status reporting. These tasks are high-volume, low-judgement and well-suited to automation.
Condition: the administrative work has to be eliminated from the process, not just performed faster. Automating a report nobody reads returns nothing.
Measure: hours per week per team spent on process administration, tracked before and after.
3. Drafting and reviewing documents
Quotes, proposals, contract summaries, order confirmations, specification documents, tender responses. The productivity case here is strong because the drafting step is usually a bottleneck held by one or two senior people, and because a first draft is easy to check.
Condition: a human reviews before anything leaves the building. Microsoft’s 2026 Work Trend Index found 86% of AI users already treat output as a starting point rather than a final answer, which is the correct habit.
Measure: turnaround time from request to approved document, plus rework rate.
4. Turning scattered data into decision-ready answers
The common complaint is not a shortage of data but the time spent assembling it — pulling numbers from a CRM, a finance system and three spreadsheets to answer one question. AI shortens that assembly step considerably, and this is where mid-sized organisations often see the fastest visible return.
Condition: this is the use case most sensitive to data quality, and the one that fails most often for reasons that have nothing to do with the model. Siloed systems, inconsistent formats and no clear data ownership produce confident, wrong answers. Designing a data foundation that makes AI work covers the infrastructure, process and governance side of this in practical terms.
Measure: time from question asked to answer trusted, and how often the answer gets independently re-checked.
5. Agentic workflows that complete steps rather than suggest them
The shift that defines 2027 is from AI that drafts and AI that acts — systems that read or record, decide, call a tool, update the system and hand off. McKinsey found 40% of organisations with over $1 billion in revenue scaling agents in at least one function, up from 27% a year earlier. Microsoft reported 15 times year-on-year growth in active agents across Microsoft 365, rising to 18 times in large enterprises.
Condition: scope and permissions defined before deployment, not after. Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Forrester’s State of Agentic AI, 2026 found roughly three-quarters of enterprise leaders reporting adoption while only a small minority had reached meaningful production.
Measure: task completion rate without human intervention, and the rate at which agents hand back or fail.
6. Compressing software and internal tooling delivery
Coding agents have changed how quickly internal tools get built, and Gartner expects that by 2027 more than 65% of engineering teams using agentic coding will treat the IDE as optional, moving control and validation to automated platforms.
Condition: this is the use case with the clearest evidence of a reverse effect, and it deserves saying plainly. A randomised controlled trial by METR in 2025 had 16 experienced open-source developers complete 246 tasks in repositories they knew well. Tasks took 19% longer when AI was allowed. The same developers had forecast a 24% speed-up and believed afterwards they had gained 20%. The gains concentrate on unfamiliar codebases and less experienced developers; on mature code held by experts, they can invert.
Measure: cycle time on comparable tickets, with a holdout group. Self-reported speed-up is the one metric this trial proved untrustworthy.
7. Forecasting and planning with less manual consolidation
Demand forecasts, capacity plans, pipeline projections and cash forecasts all involve substantial manual consolidation before any analysis begins. Removing that step lets planning run more often, which matters more than making any single forecast marginally more accurate — a monthly forecast refreshed weekly is a different operational capability.
Condition: forecast inputs need a single agreed source. Two departments forecasting from different versions of the same number produces faster disagreement.
Measure: forecast cycle time and forecast accuracy against outturn.
8. Lifting the floor on skills and shortening onboarding
The customer support study found its largest effect among newer workers, with evidence that the AI assistant spread the practices of stronger performers and moved new starters down the experience curve faster. That makes AI a training mechanism as much as a productivity tool, which matters in roles with high turnover or seasonal hiring.
Condition: the ONS found only 11% of UK businesses reported that more than half their workforce had received AI-related training. Among firms citing a lack of AI expertise as a barrier, about 62% responded by training or retraining existing staff. The BIS and EIB study of over 12,000 EU and US firms found complementary investment in software, data and workforce training was essential to realising AI’s productivity potential.
Measure: time to full productivity for new starters, and output variance between new and experienced staff.
9. Making meetings and handoffs produce actions automatically
Decisions made in conversation routinely fail to reach the systems that act on them. Transcription with structured extraction closes that gap: commitments, owners and dates get written into the CRM or project tool rather than sitting in someone’s notebook.
Condition: output has to land in the system of record. A summary emailed round adds another document to read.
Measure: share of agreed actions that appear in the system of record with an owner and date.
10. Redeploying freed capacity into work that was previously out of reach
This is the one that separates measurable gains from invisible ones. Microsoft’s 2026 Work Trend Index surveyed 20,000 knowledge workers who use AI at work across ten markets between 18 February and 7 April 2026. Two-thirds said AI let them spend more time on high-value work, and 58% said they were producing work they could not have produced a year earlier — rising to 80% among the most advanced users Microsoft calls Frontier Professionals.
Condition: someone has to decide in advance where the freed hours go. Capacity that is saved but not directed gets reabsorbed, which is the single most common reason task-level gains never appear in financial results.
Measure: a named output the freed capacity produced — more customers served, faster quote turnaround, a service launched, a backlog cleared.
What separates the firms that get gains from the ones that don’t
Four things show up consistently in the research.
Workflow redesign rather than tool insertion. Dropping AI into an unchanged process speeds up one step and leaves the constraint where it was. McKinsey’s high performers, the 6% attributing 5% or more of EBIT to AI, are consistently the ones redesigning end-to-end workflows. Where those workflows already sit inside a CRM, the redesign has to cover the business case, the system integrations and the data migration together rather than in sequence — broadly the scope that specialist Salesforce consulting services are built around.
Data that the model can actually use. The BIS and EIB research found AI adoption raised labour productivity by 4% on average across European firms, driven by capital deepening rather than job cuts, and that gains were concentrated in medium and large firms with complementary investment in data and software.
Governance sized to the risk. Gartner predicted in May 2026 that by 2027, 40% of enterprises would demote or decommission autonomous AI agents because governance gaps surfaced only after production incidents. Uniform governance applied to every agent regardless of autonomy is, in Gartner’s analysis, itself a cause of failure.
Measurement of outcomes, not usage. Licence counts, seat adoption and token consumption measure activity. None of them tells a finance director whether anything changed.
What actually changes in 2027
Three shifts are worth planning around.
Agents move from assisting to executing, which changes the governance question from “is the output accurate?” to “what is this system permitted to do, and how do we know it did it?” Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, from a base of effectively zero in 2024.
Scrutiny of returns tightens. Two consecutive years of flat EBIT attribution in the McKinsey data, against rising spend, means 2027 budget conversations will ask for evidence rather than pilots.
Adoption depth, not adoption rate, becomes the differentiator. With 30% of UK businesses already using some AI but only around one in ten adopters using it extensively, the competitive gap is opening between firms that have integrated AI into core processes and firms that have bought licences.
How to measure AI productivity properly
A workable approach, in order:
- Pick one process with a number attached to it — quote turnaround, support resolution time, invoice cycle, time to submit.
- Record the baseline for at least four weeks before changing anything.
- Deploy to part of the team and keep a comparison group. The METR trial is the clearest demonstration of why perceived gains cannot substitute for measured ones.
- Decide in advance where saved capacity goes, and write it down.
- Report the business metric and the full cost of running the AI, including licences and usage.
- Review at 90 days and be willing to stop.
Where to start
Pick the one process where you already know the number you want to change, baseline it honestly, and resist expanding scope until that number moves. The organisations seeing financial returns from AI in 2027 are not the ones running the most pilots — they are the ones that redesigned a workflow, fixed the data underneath it, and measured the outcome rather than the usage.
