What does AI really cost in a fiduciary firm? Total cost and payback by mandate size

What this is about: the four cost blocks

Introducing AI in a fiduciary firm rarely means paying a single price. Total cost consists of four blocks that are billed differently. First, usage-based costs (tokens for API models, or per-seat licence/subscription fees). Second, onboarding costs (setup, integration with your accounting and document systems, data preparation). Third, training (staff time until a tool is used productively). Fourth, ongoing operations (maintenance, updates, quality control, possibly hosting).

These blocks behave inversely. A ready-made subscription such as an AI assistant inside an office suite has low setup costs but fixed per-head licence fees. A custom solution with a provider API and your own knowledge base (RAG) has higher onboarding costs but variable usage costs that scale with actual volume.

This article is a method, not a price tag. It shows how to calculate for your firm – specific tariffs change constantly and must be checked with the vendors.

Why the calculation matters especially in fiduciary work

In a fiduciary firm the key figure is not the AI price but the billable hourly rate. Swiss hourly rates for fiduciary and accounting work depend heavily on qualification: simple data-entry and assistant tasks sit at roughly CHF 90 to 120, certified mandate leaders (federal diploma) at around CHF 140 to 200, and demanding tax or expert work above that. For the payback calculation the higher specialist rate is often the relevant one, because AI is meant above all to free up qualified staff time. Every hour saved this way has a clearly quantifiable counter-value.

At the same time, usage-based AI costs per transaction are often small – but they are invoiced in US dollars and are not flatly 'a few francs'. Providers bill per million tokens, and output tokens are markedly more expensive than input tokens. For short tasks with cheaper models a single document or e-mail task costs fractions of a franc; with long outputs or high-end models it can be a few rappen to a few francs. What matters is not the nominal price but the ratio between these token costs and the billable hourly rate.

The honest question is therefore not 'Is AI expensive?' but 'At what transaction volume and time saving does the rollout pay for itself?'. Skipping this calculation risks two errors: an expensive solution for rare tasks – or missing a saving that would pay off within a few months.

The calculation method: TCO and break-even

Total cost of ownership (TCO) over a period (e.g. twelve months) equals: one-off onboarding costs (setup + integration + initial training) plus running costs (licence/tokens + operations + quality control × number of months). Estimate the benefit as: hours saved per month × billable hourly rate × number of months.

Break-even is reached when cumulative benefit exceeds cumulative TCO. Methodically: estimate time savings conservatively and per clearly defined task type (e.g. receipt capture, standard correspondence, research in regulations). Multiply only by hours you would otherwise actually have billed or used for higher-value work – otherwise you overstate the effect.

Order of magnitude on the usage side (as of 2026, for method not as an offer): API models are billed per million tokens in US dollars, with input and output tokens priced separately. Input tokens on cheaper models range from fractions of a dollar to a few dollars per million tokens; output tokens are typically three to five times higher and, on high-end models, reach roughly USD 15 to 25 per million tokens. In typical document tasks output tokens often dominate, so they drive the unit cost. Office-bound AI assistants instead cost a fixed amount per user per month (sources: OpenAI, Anthropic, Microsoft). Which model is cheaper depends solely on your usage profile; for a CHF figure, convert at the daily rate.

Scaling rule of thumb: at high, even volume across many staff, fixed per-head licences favour cost control; with few power users or fluctuating volume, usage-based API models are often cheaper because you pay only for actual volume.

Your own cost calculation in six steps

1. 01List task types: which recurring activities qualify (receipt capture, correspondence, research)? Estimate monthly volume per type.
2. 02Measure current time: how long does a task take today? Time it for real over a few days rather than guessing.
3. 03Set time savings conservatively: assume a realistic percentage per task type and multiply by the appropriate billable hourly rate (qualification tier).
4. 04Build TCO: sum onboarding (setup, integration, training) plus running costs (licence/tokens, operations, quality control) over the chosen period; convert token prices to CHF at the daily rate.
5. 05Determine break-even: compare cumulative benefit against cumulative TCO and read off the month from which it pays off.
6. 06Pilot instead of full rollout: limit to one task type and a small group, measure real figures, then scale.

When the investment typically pays off

The rollout pays off most readily where there are recurring, uniform tasks with a measurable time share: receipt capture and pre-coding, standard client correspondence, summarising lengthy documents, or locating passages in regulations and guidelines. Here the time saving per task is easy to estimate and the volume is high.

The calculation is also favourable when the time saved can be redirected into billable or capacity-critical work. If a tool saves peak-season hours that would otherwise mean overtime or external temps, the counter-value exceeds the plain hourly rate.

Also sensible: a scoped pilot on one task type and a small user group. This lets you measure real time savings and real costs before scaling to the whole firm – the most reliable basis for any TCO calculation.

When the investment does not (yet) pay off

For rare or highly individual tasks without a recurring pattern, the time saving per case is too small to cover onboarding and operating costs. Fixed costs dominate, and rolling out 'just in case' ties up funds without clear value.

Be cautious too with highly variable quality without control: if every AI output must be checked fully by hand, net time savings shrink. In regulated fiduciary processes a final review by a qualified person is mandatory anyway – the TCO calculation must include this review effort as a cost item, not assume it away.

Finally, the calculation is unreliable while data protection and data residency questions are unresolved. When client data is processed, data processing agreements, data location and confidentiality must be part of the assessment. A seemingly cheap solution that later has to be replaced for compliance reasons is more expensive than one that is clean from the start. This is not legal advice; involve qualified parties for contracts and data protection.

FAQ

Related topics

Sources

1. OpenAI – API Pricing (offizielle Preisseite) · 2026-06
2. Anthropic – Claude API Pricing (Entwickler-Preisseite, nicht claude.com/pricing) · 2026-06
3. Microsoft – Microsoft 365 Copilot Plans and Pricing · 2026-06
4. TREUHAND|SUISSE – Branchenverband der Schweizer Treuhänder (Berufsbild und Leistungen) · 2026-06