Before any AI work, fix five things: write down how the process runs today, get its data into one place, check your systems can be connected, name an owner, and measure today's time and cost. Then pick one high-volume, low-risk use case, such as email triage or invoice data extraction, and give it a 90-day pilot: two weeks to baseline and choose, six weeks to build, four weeks to measure and decide.
This guide is for the COO or operations lead at a UK firm of 50 to 500 staff who wants a result this quarter. It covers what to modernise first, eight use cases that suit firms your size, and a week-by-week plan.
Where UK firms are with AI right now
Use of AI is growing, but most of it is still light. The Office for National Statistics reported on 8 October 2026 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 businesses with 250 or more employees the figure was 51%. Among businesses with 10 or more employees, the most common uses were text generation with large language models (21%) and visual content creation (19%).
The government's AI adoption research 2026, commissioned by DSIT and published on 29 October 2026, tells a similar story. It found 32% of businesses with five or more employees were using AI, double the 16% of a year earlier, and 47% of medium businesses (50 to 249 staff). But it also found most businesses were "still in an early or exploratory stage", and that a lack of a clear business case or proven return on investment was cited by 65% of non-adopters. Limited AI skills affected 67% of adopters.
So the gap for a mid-sized firm is rarely access to tools. Staff already have chatbots. The gap is turning one real process into something measurably faster or cheaper, with evidence a finance director will accept. That starts before any model is chosen.
Part 1: what to modernise before AI
AI speeds up a process that already works. It does not fix one that nobody can describe. Run each candidate process through this checklist. If you can't tick at least four of the five, spend the first weeks fixing the gaps rather than building.
| Check | What good looks like | Warning sign |
|---|---|---|
| The process is documented | A one-page description of the steps, decisions, exceptions and who does what. Someone new could follow it. | "Ask Sarah, she knows how it works." Every team does it differently. |
| The data is in one place | The emails, documents or records the process uses sit in a known system or shared store you have permission to use. | Inputs are spread across personal inboxes, desktops and paper. |
| Systems can be connected | The systems at each end (CRM, finance, case management, document store) have an API, a reliable export or a supported connector. | A locked legacy system with no API, where the only way in is someone retyping data. |
| There is a named owner | One person who runs the process, can change it, and will decide whether the pilot goes live. | Ownership is shared between three departments, or sits with IT only. |
| Today's baseline is measured | You know volume per month, minutes per item, error or rework rate and cost per item, from real data over a few weeks. | Estimates like "it takes ages" with no numbers behind them. |
The locked-system point deserves a moment. If the destination system has no API and no export, the AI can still read and prepare the data, but someone or something has to get it in. It is better to find this out in week one than in week seven.
This checklist is a single-process view. For the wider picture across strategy, data, technology, people, governance and operations, see our guide to the six dimensions of enterprise AI readiness.
Part 2: eight AI use cases a mid-sized firm can pilot in 90 days
These eight share three traits. They are high volume, a person can check the output quickly, and the result can be counted. The pilot timeframes below are typical for a tightly scoped pilot with data access already agreed. They are not guarantees. Security reviews, data access and supplier lead times are the usual causes of delay.
| Use case | What must be in place first | Typical time to a working pilot | How to measure success | Main risk |
|---|---|---|---|---|
| Inbound email and enquiry triage | A shared inbox or form, a written list of categories and routing rules, a few hundred past emails labelled by outcome | 3 to 6 weeks | Minutes from arrival to correct owner; share routed correctly first time; staff hours saved | Urgent or sensitive emails misrouted; needs a clear fallback to a person |
| Invoice and document data extraction | Sample documents of each type, a defined list of fields, access to the finance or document system | 4 to 8 weeks | Field-level accuracy; minutes per document; exceptions per 100 documents | Poor scans and unusual layouts; wrong figures reaching the ledger if checks are skipped |
| Contract and clause review, first pass | Your standard positions or playbook written down, a set of past contracts, a lawyer to own the review | 4 to 8 weeks | Time to first review; clauses flagged that a lawyer agrees with; issues missed | Over-reliance on the AI; it is a first pass, a qualified person still signs off |
| Customer support reply drafting with human approval | Help-desk system with API access, an up-to-date knowledge base, tone and policy guidance | 4 to 6 weeks | Handle time per ticket; share of drafts sent with little or no edit; customer satisfaction held or improved | Confident but wrong answers if the knowledge base is out of date |
| Sales call notes to CRM | Call recording with consent, CRM with API access, an agreed list of fields to fill | 3 to 6 weeks | Admin minutes per call; CRM field completion rate; rep adoption | Recording consent and data protection; errors in deal values or next steps |
| Internal knowledge Q&A over policies | Current, owned policy documents in one store, access permissions mapped | 4 to 8 weeks | Questions answered with a correct cited source; fewer repeat questions to HR, finance or IT | Out-of-date policies giving wrong answers; staff seeing documents they should not |
| Report and board-pack drafting | Stable data sources, a template, past packs as examples, a named editor | 4 to 8 weeks | Hours to first draft; number of factual corrections; on-time delivery | Numbers misstated in commentary; every figure must be checked against source |
| Onboarding and KYC document checks | Written checklist of required documents and rules, secure document intake, compliance owner | 6 to 10 weeks | Minutes per file; share of files complete first time; false passes found in sampling | Regulatory exposure; the AI prepares, a trained person decides |
Three of these, invoice extraction, contract review and KYC checks, are forms of document processing. If that is your area, our explainer on what intelligent document processing is covers how extraction, classification and validation fit together.
To choose between them, score each candidate against the Part 1 checklist and against volume. A process that handles 2,000 items a month at eight minutes each is a far stronger first pilot than one that handles 50 items at an hour each, even if the second feels more important.
Part 3: a 90-day plan
| Weeks | Goal | Main tasks | Output at the end |
|---|---|---|---|
| 1 to 2 | Baseline and pick one | Shortlist two or three processes; run the readiness checklist; measure volume, time and error rate; confirm data access and security needs; agree the owner | One chosen process, a measured baseline, and a written pass mark for the pilot |
| 3 to 8 | Build the pilot | Connect to the data and systems; build the first version; test on past items; run it alongside the current process with a person checking every output; fix the failure cases weekly | A working pilot used by the real team on real volume, with a human approving outputs |
| 9 to 12 | Measure and decide | Track the same measures as the baseline; sample outputs for accuracy; gather team feedback; cost the move to production | A go, change or stop decision backed by numbers, and a costed plan if it is go |
Weeks 1 to 2: baseline and pick one
Pick one process, not three. Running several pilots at once splits the owner's attention and makes it harder to learn. Measure the current state from real records where you can, such as help-desk timestamps or finance system logs, rather than asking people to guess. Write the pass mark down now. For example: "cut average handling time per invoice from 6 minutes to under 3, with field accuracy of at least 97% on a 300-invoice sample." The exact numbers are yours to set; the point is that they are agreed before anyone sees results.
Weeks 3 to 8: build the pilot
Build the smallest version that does the job end to end. Run it in shadow mode first, where the AI produces an output but the team still does the work the old way, so you can compare. Then switch to assisted mode, where the AI does the first pass and a person approves. Review failures every week with the process owner.
Sort out data protection early. If the process touches personal data, check where it will be processed and stored, and whether you need a data protection impact assessment. For options on keeping client data in the UK or in your own environment, see our guide to Claude and UK GDPR for client data.
Weeks 9 to 12: measure and decide
Measure the same things you measured in week one, the same way. Sample outputs by hand to check accuracy, because a dashboard of averages can hide a small number of serious errors. Then make one of three calls. Go: the pass mark was met and you have a costed production plan. Change: it nearly worked and one clear fix is needed. Stop: it did not work, and you have learnt something useful for the next choice. A stop is a fair result. What you want to avoid is a pilot that never ends.
Moving from a successful pilot to daily use across the firm is its own piece of work: monitoring, support, training and ownership all need to be in place. Our article on how to scale AI pilots into enterprise adoption covers that step.
Common reasons 90-day pilots slip
The AI is rarely the slow part. The usual causes are waiting for system access or API keys, security or procurement review starting late, the process owner being pulled onto other work, and scope creeping from one document type to five. Each of these can be handled in weeks 1 to 2 if you ask about them up front.
How Kelriva can help
We are a two-person London AI consultancy that works on fixed fees. Our AI Readiness Assessment starts from £3,500 and takes 9 to 14 days, and a pilot or proof of concept starts from £15,000 over 4 to 8 weeks. If you want a quick, free first view of where your firm stands before talking to anyone, take the 2-minute AI Readiness Score. It asks six questions and shows where to focus first.
Sources
All links checked on 5 November 2026. Office for National Statistics, Business insights and impact on the UK economy: 8 October 2026. GOV.UK, AI adoption survey report 2026: executive summary, published 29 October 2026. GOV.UK, Artificial intelligence adoption research 2026 (full report, PDF).