AI Reception Costs: Build a Defensible UK Case

There is no honest universal percentage for what an AI receptionist “saves”. The result depends on demand, call length, existing staffing, customer needs, integration quality, supplier charging, failure rates and what happens to the work the system cannot complete. A defensible calculator uses the organisation’s own records and reports a range, not a marketing promise.

This model is current to 31 July 2026. Payroll references apply to the relevant UK rules and 2026–27 tax year where stated; employment contracts, pension schemes and sector duties still need local review. Healthcare, finance, legal services and emergency-sensitive operations require additional safety and regulatory analysis.

The comparison should normally be current operation versus proposed AI-assisted operation, not “one employee versus one bot”. Reception is a service made from people, telephony, systems, supervision and recovery. Automation usually changes that mix rather than removing it.

Define the service before entering costs

Set the same scope on both sides of the comparison:

• channels included: calls, web chat, SMS, email or messaging;
• operating hours and seasonal peaks;
• languages and accessibility routes;
• enquiry types, actions and explicit exclusions;
• target answer, abandonment and resolution standards;
• required identity checks and records;
• human escalation hours and response target;
• business-continuity requirement.

If the proposed service covers nights but the baseline covers office hours, show the extra coverage as a separate benefit or option. Do not quietly treat it as a like-for-like saving. Likewise, “answered” is not “resolved”: a call that creates a manual callback belongs in residual work.

Collect at least a representative period of switchboard and workflow data. Use distributions, not only averages. Record contact volume by half-hour, duration, abandonment, transfers, repeat contacts, after-call work, corrections, complaints, bookings completed, urgent escalations and unavailable-service periods. Sample calls to establish why people contact the business and which journeys need judgement or reassurance.

Build the current-state cost

Start from actual payroll and ledger data. A useful annual structure is:

Current service cost = direct labour + employer on-costs + cover + supervision + technology + facilities + recruitment/training + failure cost

Direct labour and on-costs

Use the salary or hourly rate actually paid for reception work, including contractual overtime and shift premiums. Check current wage compliance against the government’s National Minimum Wage rates from April 2026, but do not substitute the statutory floor for the real cost.

Calculate employer National Insurance with the correct employee category, pay frequency and thresholds. HMRC publishes employer rates and thresholds for 2026–27 and explains the current National Insurance categories and employer rates. Use payroll output where possible because allowances, categories and pay patterns can change the result.

Add the employer’s actual pension contribution. The Pensions Regulator explains that employers must assess automatic-enrolment duties and contribute for eligible staff; its current earnings-threshold page provides 2026–27 inputs. A scheme may require more than the statutory minimum, so use scheme and payroll records.

Workers are generally entitled to 5.6 weeks’ paid holiday under the government’s holiday entitlement guidance. Avoid double counting. Salary already includes paid leave; the incremental service cost is the temporary cover, overtime or reduced capacity needed while people are absent, not another 5.6 weeks of salary automatically. Treat sickness, family leave and training in the same evidence-based way.

Operating and failure cost

Add team-lead or manager time attributable to reception, telephony, CRM licences, headsets, equipment, premises allocation where genuinely avoidable, recruitment, onboarding and quality checks. Separate fixed costs that remain after automation from costs that can actually change.

Estimate failure cost from observed events:

• abandoned demand that creates a repeat contact or lost booking;
• incorrect bookings or messages requiring correction;
• avoidable transfers;
• complaints and service recovery;
• overtime caused by peaks;
• downtime and manual backlog.

Use a conservative method. A missed call is not automatically a lost customer, and a recovered hour is not automatically cash. Present unmonetised service measures alongside cost.

Build the AI-assisted cost

Use the supplier’s written pricing and a measured demand model. The annual structure is:

AI-assisted cost = implementation + platform/usage + telephony + integrations + assurance + monitoring + residual people + failure/continuity + exit

One-off and recurring technology

Include configuration, workflow design, number routing, integrations, knowledge preparation, data migration, testing, security review, legal review, staff training and project management. Amortise one-off costs over a stated period and also show the first-year cash requirement.

Recurring items may include a platform fee, per-minute or per-interaction use, speech or model charges, telephone numbers, call recording/storage, CRM licences, premium support and analytics. Model peak concurrency from the contract; do not assume “unlimited” means no capacity, throttling or fair-use constraint.

Add a sensitivity case for higher contact volume and longer calls. Include VAT treatment appropriate to the organisation rather than assuming it is recoverable.

Residual human work

This is the most commonly omitted line. Measure or estimate:

• enquiries deliberately transferred to staff;
• callbacks when the system is uncertain or closed;
• identity and safeguarding exceptions;
• complaint handling;
• daily knowledge and queue review;
• sample quality assurance;
• incident response and supplier management;
• correcting failed bookings or CRM updates;
• maintaining an alternative for people who cannot use the automated channel.

Assign each category a volume, minutes and loaded labour rate. If the pilot has not yet measured it, use low, central and high cases. A system that resolves simple demand may leave a smaller but more complex workload, so the old average handling time is often unsuitable.

Compliance, security and continuity

Budget for data mapping, contracts, a DPIA decision, retention configuration, access reviews, penetration or supplier-assurance review where proportionate, audit-log retention and periodic testing. The ICO’s DPIA guidance requires high-risk processing to be assessed before it begins.

Apply the NCSC secure AI system guidelines to the configured service: document components and data flows, minimise privilege, protect secrets, monitor changes and plan secure decommissioning. Price the fallback route, diversion number, status monitoring and exercises. Continuity is a service cost, not a footnote.

Include export and termination assistance, number portability, data deletion evidence and the cost of operating manually during migration. A cheaper service that traps records or telephone numbers can have a high exit cost.

Calculate outcomes without inventing value

Use three views:

• Cash view: costs that enter or leave the budget during the period.
• Capacity view: verified staff hours released and the approved use for them.
• Service/risk view: access, quality, complaints, safety and resilience.

The central annual difference is:

Net cash change = current avoidable cash cost − AI-assisted cash cost

For a first-year view:

First-year net cash change = annual net cash change − unamortised implementation cash

For capacity:

Net hours released = removed handling and wrap time − review, exception, maintenance and recovery time

State how released hours create value. If staff remain and use the time for customer care or revenue work, report the hours and resulting measured outcome. Do not label their salary a saving.

Present at least three scenarios:

Do not hide a negative scenario. It tells management which assumption controls the decision.

Quality and legal gates come before payback

The CMA’s AI-agent consumer-law guidance is clear that a business remains responsible when an agent is supplied by a third party. The system needs accurate information about price, rights and refunds, meaningful human oversight and prompt correction. It should identify itself appropriately and make escalation workable.

Set non-financial gates before approving release:

• no emergency or safeguarding journey can be trapped in automation;
• no material policy, price or appointment error remains unresolved;
• identity checks match the action’s risk;
• consent, transparency, retention and recording arrangements are approved;
• human transfer works at peak and during degraded service;
• customers can use an accessible alternative;
• every write action is attributable and reversible;
• supplier and internal incident routes have been exercised.

For healthcare, the dedicated [clinic AI reception guide](/blog/ai-reception-healthcare-clinics-complete-guide-2025) explains why administrative booking and clinical triage must be separated. For a broader selection process, use the 90-day SME AI guide.

Failure modes the spreadsheet must represent

A financial model should include operational reality:

• the voice service misunderstands names, dates, accents or noisy calls;
• a knowledge answer is plausible but out of date;
• a caller discloses special-category or confidential information unexpectedly;
• the CRM connector creates a duplicate or writes to the wrong record;
• an attacker attempts prompt injection or account enumeration;
• the model, price or supplier subprocessor changes;
• the primary service or telephony carrier fails;
• staff skills decay because the manual route is rarely used.

Link each failure to a control, owner, test frequency and cost. If a high-consequence failure has no workable control, narrow the use case rather than assigning it a convenient probability.

A 90-day evidence plan

Days 1–30: measure and model

Define equal service boundaries, sample demand and reconcile payroll, telephony and failure costs to records. Obtain written supplier pricing and data-flow information. Build the low, central and stress cases. Complete privacy, security and sector-assurance decisions.

Gate 1: no pilot with live customers until scope, prohibited actions, human fallback, retention, permissions and incident ownership are approved; no unresolved critical risk is accepted merely because the central spreadsheet is positive.

Days 31–60: shadow and test

Run the system without completing live actions, or with every output held for staff approval. Measure correct intent, correct resolution, transfer, repeat contact, correction minutes and customer-access issues by journey. Test outage, overload, malicious prompts and connector errors.

Gate 2: no autonomous booking, messaging or record update until all high-consequence tests have a safe result, audit records are complete and transfer meets the defined target under peak conditions.

Days 61–90: limited live operation

Release a restricted set of low-risk journeys. Reconcile supplier invoices and internal labour weekly. Compare the same service and demand window with baseline, and review complaints, accessibility, material errors, staff load and downtime.

Gate 3: approve expansion only if the conservative case is affordable, the central cash and capacity claims reconcile to evidence, service guardrails hold, severe failures are closed and the manual service can be restored. Otherwise change the workflow, renegotiate or stop.

The board-ready output

The decision pack should show assumptions and provenance, not a single ROI percentage. Include the service definition, baseline period, payroll sources, supplier quote date, volume distribution, first-year cash, steady-state range, residual-work measurement, sensitivity table, quality results, risk register, fallback test, implementation owner and review date.

Re-run the calculator when wages, tax thresholds, pension arrangements, demand, supplier prices, integrations or scope change. An AI-assisted reception service can be worthwhile, but only the organisation’s observed demand and controlled pilot can establish that. The calculator’s job is to make uncertainty visible before a contract turns it into cost.