How AI Automation Saved a UK Recruitment Agency 40 Hours a Week

The Agency Before Automation

A 12-person recruitment agency in the West Midlands placing temporary and contract workers across logistics and manufacturing. Three consultants, two administrators, and a compliance officer spending the bulk of their week on tasks that had nothing to do with placing candidates.

The agency processed approximately 200 timesheets per week, mostly handwritten or photographed and sent via WhatsApp. Each one needed manual data entry into their CRM (Bullhorn), cross-referencing against shift patterns, and then exporting to Xero for invoicing. That single workflow consumed 15 hours per week across two admin staff. Candidate communication ate another 12 hours: sending interview confirmations, collecting availability, chasing references, and following up on document submissions. Compliance document handling, primarily Right to Work checks, DBS certificates, and health and safety sign-offs, took the compliance officer 10 to 12 hours per week during peak onboarding periods.

Total administrative overhead: roughly 38 to 40 hours per week across the team. That is one full-time equivalent salary spent entirely on data entry, chasing, and copying information between systems. For context, our full guide to AI automation for recruitment agencies covers the broader opportunity, and the AI automation solutions built specifically for recruitment businesses page outlines the service offering. This post covers what happened when the agency decided to act on it.

Workflow 1 — Timesheet OCR and Data Entry

• Handwritten timesheets photographed on a phone are one of the hardest inputs to automate because handwriting quality, lighting, and image angles vary with every submission.
• The solution combined AWS Textract for OCR, GPT-4.1 Mini for field validation, and n8n for orchestration.
• Build time: 3 weeks. The workflow processes 200+ timesheets per week with 94% first-pass accuracy.

The previous process: a temp worker photographs their handwritten timesheet and sends it via WhatsApp to the agency. An admin opens the image, reads the handwriting, types the hours into Bullhorn against the correct placement record, checks the shift pattern matches the client’s requirements, and exports the data to Xero for payroll and invoicing. Average processing time: 4 to 5 minutes per timesheet. At 200 timesheets per week, that is 13 to 17 hours of pure data entry.

The automated workflow runs on n8n with four stages. First, WhatsApp Business API receives the image and forwards it to a webhook. Second, the image passes to AWS Textract via an HTTP node. Textract’s Detect Document Text API extracts raw text from the handwritten image at a cost of $0.0015 per page. Third, the extracted text feeds into a GPT-4.1 Mini prompt that parses the raw OCR output into structured fields: worker name, date, start time, end time, break duration, and total hours. The prompt includes validation rules: flag any entry where total hours exceed 12 or break duration is zero for shifts over 6 hours. Fourth, validated entries push into Bullhorn via its API, and the data simultaneously exports to Xero for invoicing.

The 94% first-pass accuracy means roughly 12 out of 200 timesheets per week require manual review. These are flagged in a Slack channel with the original image and the AI’s extracted data side by side. An admin corrects them in under a minute each. Total human time per week on timesheet processing dropped from 15 hours to approximately 45 minutes of exception handling.

Monthly cost of the automation: £35 for AWS Textract (200 pages/week at $0.0015), approximately £4 for GPT-4.1 Mini API calls, and zero for n8n (self-hosted). Total: under £40 per month replacing 15 hours of weekly admin time.

Workflow 2 — Candidate Communication Sequences

The second workflow automated the repetitive messaging that consumed 12 hours per week across three consultants. Every new candidate submission triggered a chain of manual messages: interview confirmation, pre-interview prep notes, post-interview follow-up, reference request, and availability check for upcoming shifts.

The consultants were writing similar messages hundreds of times per month. The phrasing varied slightly, but the content was the same. This is the definition of a task where AI adds value without replacing human judgment: the consultant still decides who to submit for which role, but the follow-up communication runs on autopilot.

The n8n workflow triggers when a candidate record in Bullhorn moves to a specific pipeline stage. An LLM generates the appropriate message using the candidate’s name, the role they have been submitted for, the client name, and the interview date. The prompt constrains the output to match the agency’s tone: direct, friendly, no corporate fluff. Messages route through the WhatsApp Business API for candidates who prefer WhatsApp or through email for those who provided only an email address.

The AI-powered CV screening and candidate filtering guide covers the earlier stage of this pipeline. This workflow picks up where screening ends: once a candidate is shortlisted, communication automation handles everything up to placement confirmation.

A key design decision: every automated message includes a “sent by [consultant name]” signature, and replies route back to that consultant’s direct inbox. Candidates never interact with a bot. They interact with their consultant, backed by a system that ensures no message gets forgotten.

Result: consultant communication time dropped from 12 hours per week to 3 hours. The remaining 3 hours cover personalised conversations that genuinely need human input, responding to candidate questions, handling scheduling conflicts, and managing client expectations.

Workflow 3 — Compliance Document Collection and Chasing

• UK recruitment agencies placing temporary workers must verify Right to Work status, collect DBS certificates where required, and maintain health and safety documentation. Non-compliance carries fines up to £20,000 per illegal worker under the Immigration, Asylum and Nationality Act 2006.
• The compliance workflow automated document requests, reminders, expiry tracking, and status reporting.
• The compliance officer’s weekly workload dropped from 10-12 hours to 2-3 hours.

Compliance is the most painful admin task in recruitment because it is high-stakes, repetitive, and never-ending. Documents expire. New starters need collecting. Audits happen without warning. The agency’s compliance officer was spending most of her week sending the same document request emails, chasing candidates who had not responded, and manually checking expiry dates across a spreadsheet.

The automated workflow tracks every candidate’s compliance status in Bullhorn using custom fields. When a candidate enters the pipeline, n8n generates a personalised document request via WhatsApp or email listing exactly which documents are needed based on the role type and sector. If the candidate has not responded within 48 hours, the first reminder fires. A second reminder follows at 96 hours. After 7 days, the candidate is flagged for manual follow-up by the compliance officer.

Documents submitted by candidates are received via a webhook (photo or PDF upload), stored in Google Drive with standardised naming conventions, and the Bullhorn record updates automatically. For Right to Work documents, the system performs a basic validity check: is the document type recognised, does it contain a name matching the candidate record, and is the expiry date in the future. Documents that fail any check route to the compliance officer for manual verification.

Expiry tracking runs on a daily schedule. Any document expiring within 30 days triggers an automated renewal request to the candidate and a notification to the compliance officer. This replaced a manual spreadsheet check that previously took 2 hours every Monday morning.

Result: compliance admin dropped from 10 to 12 hours per week to 2 to 3 hours. The compliance officer now spends her time on document verification and audit preparation rather than chasing and data entry.

Total Time Saved and ROI Breakdown

Combined results across all three workflows after 8 weeks of production use:

The 32 hours per week saved translates to approximately £640 per week in admin labour costs, assuming a blended rate of £20 per hour for administrative staff. That is £2,560 per month in recovered capacity for a total automation running cost of £75 per month.

The total build cost across all three workflows was £11,500, covering discovery, build, testing, and a 30-day support period. Payback period: under 5 months. After payback, the agency saves approximately £2,485 per month net.

For ROI benchmarks from AI automation projects across five common workflow types, this project sits at the higher end of the range because recruitment agencies have unusually high volumes of repetitive, document-heavy admin work.

The agency did not reduce headcount. The two administrators and compliance officer redirected their time to higher-value tasks: candidate care, client relationship management, and business development support. The agency’s placement rate increased by 18% in the quarter following automation, which the managing director attributed partly to consultants having more time for actual recruitment. These are workflow automation builds designed for recruitment and staffing agencies where the ROI is measurable within the first quarter.

What Did Not Work and What We Changed

Not everything worked on the first attempt. Three issues required iteration after the initial build.

Timesheet OCR accuracy on poor-quality images. The initial build achieved 87% first-pass accuracy, below the 93% target. The problem was not the OCR engine but the image quality: photos taken in dim warehouses, creased paper, and ballpoint pen on carbon copy paper. The fix was a pre-processing step that adjusts image contrast and brightness before sending to Textract. This raised accuracy to 94%. For the remaining 6%, the agency introduced a simple rule: if Textract’s confidence score falls below 70% on any field, the entire timesheet routes to manual review rather than attempting a partial extraction.

Candidate message tone mismatch. The first iteration of the communication prompts produced messages that were too formal for the agency’s casual, first-name culture. Candidates noticed. The fix was adding three example messages from real consultants to the prompt as few-shot examples, which brought the AI output in line with the agency’s voice.

Compliance document type recognition. The initial build used a simple filename-based classification that failed when candidates sent documents with generic names like “IMG_4521.jpg.” Switching to an LLM-based classification step that reads the document content and identifies the document type (passport, driving licence, DBS certificate, health and safety card) solved the problem. This added roughly £8 per month in API costs but eliminated the false classifications entirely.

Honest reporting on what went wrong is part of a comparison of AI automation approaches for recruitment agencies because every project has iteration cycles. The difference between a good automation project and a failed one is not whether problems arise. It is whether the team expects them and has a process for resolving them.