£580k/yr in partner time recovered. Tax-season throughput tripled.

Whitfield Tax is a 12-partner accounting practice in the South East serving owner-managed businesses, SMEs, and private clients. Every January-March, the firm hit the same wall: partners could review roughly 240 returns total across the season, even with juniors preparing first drafts.

Each SA100 or CT600 needed 4 hours of partner attention — review the source documents, check the computation, flag anomalies, sign off. Demand outstripped capacity by January 10th every year. New clients were turned away. Existing clients waited weeks for a callback.

Junior staff drafted returns in Iris and CCH, then queued them for partner review. The bottleneck was structural: partners read every line because the cost of a missed adjustment was real, and juniors couldn't fully relieve them on technical work.

The firm had already digitised intake. Clients uploaded documents via a portal. But the documents arrived as a mix of scans, photos, PDFs, and spreadsheets — and the structured extraction was still mostly manual.

Every uploaded document hits an OCR + Claude extraction step that pulls the line items relevant to the return type — P60, P11D, dividend vouchers, rental statements, bank interest, capital disposals, Companies House filings, control accounts. Each extraction is cross-checked against the firm's expected schema and flagged if anything is missing or out of range.

Once a client's documents are complete, the tax computation engine runs against UK 2025/26 statutory rates — personal allowance, dividend allowance, CGT bands, corporation tax bands and marginal relief, pension annual allowance taper, child benefit clawback. The draft return is generated in the firm's house format and queued for partner review with anomaly flags surfaced at the top.

Partners review a structured diff against the prior year, not a blank screen. Adjustments are inline. Sign-off triggers Companies House filing (for CT600s where applicable) and HMRC e-filing. The audit log captures every change with timestamp, user, and reason — sized for ICAEW and HMRC enquiry response.

The system integrates with Iris and CCH so partner sign-off doesn't require leaving the firm's existing workflow. Nothing about how partners interact with their tools changed; the prep work that used to take 3.5 hours of junior + partner time now happens automatically.

Tax-season throughput went from 240 returns to 850 in the first full season after rollout. The constraint stopped being partner review hours and started being client document collection — a much easier problem to manage.

Partner time per return dropped from 4 hours to 38 minutes — partners now review high-confidence drafts with anomalies pre-surfaced. The remaining 38 minutes is genuine judgement: discretionary disclosures, planning recommendations, client relationship moments. The 3.5-hour mechanical review evaporated.

AI-generated draft accuracy held at 99.7% against partner-corrected final returns over the first 850 submissions. The 0.3% adjustment rate is concentrated in edge cases (split-year residence, lifetime allowance protections, employee share scheme reporting) where partner judgement was always required — and those edits typically take 10 minutes, not 4 hours.

Annualised, £180k of partner time was recovered. The firm took on 220 new clients in the year following deployment instead of turning them away.

Week 1–2: Discovery, document corpus audit, tax computation schema. Week 3–4: OCR + extraction pipeline, computation engine. Week 5: Partner review interface, anomaly detection, IRIS/CCH integration. Week 6: HMRC + Companies House connectors, audit log. Week 7: Pilot run on 30 returns, partner QA, production deploy.