by Claude Opus 5.5

Which business processes tend to yield the fastest, most reliable returns from AI adoption, and what prerequisites typically determine success?

The fastest reliable returns come from high-volume, text-heavy processes that are already digital, where a person can check the output quickly and mistakes can be reversed. Examples include contact-centre agent assistance, invoice and claims intake, clinical and meeting notes, internal knowledge search and drafting, and software maintenance. Whether a project succeeds depends less on the model than on four things: a measured baseline, clean access to the right data, a review step built into the system staff already use, and a manager who decides in advance what the freed time is for.

Where the evidence is strongest

The strongest studies share a pattern: AI helps most on bounded, repeated tasks with a clear standard of “good”.

  • Customer support. In a study of 5,179 support agents, Brynjolfsson, Li and Raymond found that an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and lower-skilled agents, with little effect on the most experienced. That is the clearest field evidence that AI spreads best practice to newer staff.

  • Knowledge work inside the frontier. In the Harvard/BCG experiment with 758 consultants, those using AI on tasks within its capability completed 12.2% more tasks, 25.1% faster and at more than 40% higher quality. On a task outside its capability, they were 19 percentage points less likely to get it right.

  • Clinical documentation. The government estimates that ambient voice technology could save about 90 seconds per GP appointment and free the equivalent of more than 2,000 full-time GPs. That is an estimate, not a measured trial result. A Midlands procurement now covers 70,000 clinicians (both reported figures).

UK evidence also shows the limits. In the cross-government Microsoft 365 Copilot experiment (20,000 civil servants, late 2024), users reported saving 26 minutes a day, mostly on drafting and searching for information. The figures were self-reported. A separate Department for Business and Trade evaluation “did not find evidence that time savings have led to improved productivity”. The METR study of experienced open-source developers found they were 19% slower with AI tools on their own mature codebases, while believing they had been sped up by 20%.

The lesson is not that AI fails. Returns arrive reliably only when the process converts saved minutes into more output, better quality or lower cost.

A ranking of processes by speed and reliability of return

  • Contact-centre agent assist and after-call summaries. Why it pays back: Huge volume, existing metrics (handle time, resolution), many new starters. Key prerequisite: Knowledge base that is current and owned. Typical trap: Cutting staff before call-volume effects are proven.

  • Document intake: invoices, claims, forms, onboarding packs. Why it pays back: Extraction and classification are checkable field by field. Key prerequisite: Confidence thresholds and an exception queue. Typical trap: Automating a messy upstream process instead of fixing it.

  • Clinical, case and meeting notes. Why it pays back: Saves professional time; the person in the room checks the note. Key prerequisite: Consent and records-management rules agreed. Typical trap: Notes drift longer, so the time is spent reading instead of writing.

  • Internal knowledge search and drafting (policies, bids, HR queries). Why it pays back: Cuts search time; drafts are reviewed anyway. Key prerequisite: Permissions tidied so the tool does not surface what staff should not see. Typical trap: Stale documents producing confident wrong answers.

  • Software maintenance, tests, migrations, IT service desk. Why it pays back: Expensive staff; output can be tested automatically. Key prerequisite: Test coverage and code review. Typical trap: Gains vanish on complex legacy code without tests.

  • Marketing variants and sales research. Why it pays back: Cheap experimentation. Key prerequisite: Brand and claims sign-off. Typical trap: Volume up, conversion flat.

Two categories tend to disappoint early. The first is decisions about people. Hiring, performance management and credit carry Data (Use and Access) Act safeguards on automated decision-making (in force since 5 February 2026) and Equality Act exposure. The ICO’s March 2026 “Recruitment rewired” review found that recruitment automation often had “no meaningful human involvement”. The second is fully autonomous customer-facing agents. Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear value or weak risk controls. In August 2025 Commonwealth Bank of Australia reversed 45 redundancies made on the strength of a voice bot, admitting it had not “adequately consider[ed] all relevant business considerations”.

The prerequisites that decide success

1. A baseline taken before launch. You need cost per unit, cycle time, error and rework rates, and case mix, measured for a normal period beforehand. Without them every pilot “works” and none can be defended to a finance director. See 3.7.

2. A stable process with defined exceptions. If staff cannot say which cases must go to a human, an AI system cannot either. Map the exception routes first. Often this step alone yields a good share of the gain.

3. Clean, permissioned data. Retrieval tools inherit your SharePoint. Old policy versions, duplicate templates and over-broad access rights become wrong answers and data leaks. Finance teams have a structural tailwind: mandatory e-invoicing for VAT invoices from 2029 will turn invoice capture into structured data, so extraction projects should be designed to transition to it.

4. Embedding in the system of record. Assistance that lives inside the CRM, ERP, case-management system or code editor gets used. A separate chat window gets used less, and use drifts towards personal tools. Deloitte’s 2026 UK survey of 25,000 workers found that 31% of generative AI users use it without their employer’s knowledge, and 46% use free tools.

5. Training on the actual work. ONS reports that about 62% of firms citing a lack of AI expertise are training or retraining staff, but only 11% of businesses with 10 or more employees have trained more than half their workforce. Short, task-specific sessions run by team leads beat generic courses on “prompting”.

6. A decision about freed capacity. The DBT finding is the warning. Before launch, decide whether saved time goes to clearing backlog, absorbing growth without hiring, improving quality or reducing overtime and agency spend. Then measure that outcome.

7. Someone who owns the result. That should be a named business owner with a target, not “the AI team”. A central team should provide the platform, guardrails and evaluation methods.

A quick screen for candidate processes

Score each candidate yes or no:

  1. Does it run more than a few hundred times a month?

  2. Are its inputs and outputs already digital?

  3. Can a competent person check an output in under a fifth of the time it takes to produce one?

  4. Is a typical error reversible before it reaches a customer, a regulator or a payment?

  5. Is there a current KPI with at least three months of history?

  6. Can the tool sit inside the existing system?

  7. Is the data lawful and safe to use, with a DPIA done where personal data is involved?

Five or more yeses means a fast-return candidate. Three or fewer means fix the process or the data first.

Bottom line

The quickest wins in 2026 are unglamorous: summarising, extracting, drafting and searching inside existing systems, with a human checking. Agentic automation can follow once those foundations, and the measurement, are in place. ONS reports that 29% of UK businesses now use AI and 49% of large firms do. What separates the firms that see returns is mostly discipline about baselines and redeploying freed time, not access to better tools.

Sources

From AI and Jobs: UK, October 2026