AI Project Quotes in Switzerland: A 14-Line Cost Normalizer for SMEs
A short AI consulting quote is not comparable to a longer one until both describe the same work, operating period, and assumptions. The practical answer is to request one shared scope, separate one-time and recurring charges, include internal effort, and make unknowns visible instead of treating them as zero.
The IAPME guide to AI consulting for Swiss SMEs discusses partner choice and engagement models. This article adds a narrower buyer tool: a reusable way to normalize offers that use different labels, billing units, and exclusions. IAPME publishes consulting content and also provides consulting services; treat its guide as a useful starting point, not as independent evidence about suppliers.
The decision this worksheet supports
Use the method when two or more providers propose help with the same AI-enabled workflow. Before approval, the company should be able to answer four questions:
- What is included in the first delivery?
- What recurring obligations begin after launch?
- What work will the company itself have to supply or pay for?
- What does the offer leave unknown, and what happens if that assumption is wrong?
The output is not “Vendor A is cheaper.” It is “under a common scope and stated assumptions, this offer has a first-year cost range; these items are confirmed, these are estimates, and these remain unpriced.” If offers do not cover equivalent results, do not force them into one total.
A fixed implementation fee, monthly service, consumption charge, and internal staff time are different cost types. A quote may call a recurring license included while limiting the number of users or requests. Another may show the license separately but include support. Both can be reasonable. The buyer’s job is to make the boundary legible before comparing totals.
What the available evidence does and does not say
I checked a small first-party Search Console slice for the IAPME property. For Switzerland, country code CHE, from 13 October through 10 November 2026, the query “ai consulting switzerland” returned 103 impressions and zero clicks for the English partner-comparison page, at an average position of 10.60. The query “ai consultants switzerland” returned 41 impressions and zero clicks for the same page, at an average position of 8.34. These are two query-page rows from one property and one reporting window. They are not a market-size estimate, survey of buyers, proof users want a pricing calculator, or evidence of ranking improvement. They are a narrow signal that the page is being shown for partner-selection language.
The existing IAPME guide explains partner categories, selection criteria, recurring costs, and questions to ask. Its advice says not to choose a consultant from an emailed quote alone. The gap addressed here is more operational: if a buyer already has two offers, what fields make their totals comparable? This worksheet is an editorial tool for that task, not a validated industry standard.
A Swiss Economic Institute at ETH Zurich, KOF, discussion of its 2020–2022 innovation survey reports that AI use differed by company size in that period and that costs and skilled-worker shortages were obstacles to innovation. That evidence is useful context for why a small company should budget internal time and operating capability, but it is historical data, not a current estimate of 2026 adoption or consulting prices. Do not silently replace the period with the publication date. KOF’s article describes the period and survey context.
A 2024 Boston Consulting Group report says organizations it classifies as more advanced focus on core business processes and on people and processes, not only technology. It is consulting-firm global research, not a representative study of Swiss SMEs, so it is a directional reminder rather than a price benchmark or proof a particular partner will succeed. BCG’s report sets out that view.
A recorded Swiss SME webinar hosted by MTF lists data protection, confidentiality, contracts, AI Act developments, tool configuration, anonymization, and internal rules among its topics, and says audience questions were included. The event page does not publish a transcript or the exact questions. I use it only as evidence that these subjects were put on a Swiss SME agenda, not as a count of buyer concerns or as legal authority. MTF’s webinar page.
For a specific EU-facing feature, Article 50 of the EU AI Act describes transparency obligations for certain providers and deployers. It is not a blanket rule that every Swiss SME or every AI use case has the same duty. Whether it applies depends on the system, role, use, and territorial scope; have qualified counsel assess a real deployment. The European Commission’s AI Act Service Desk provides the Article 50 text. The worksheet asks buyers to price required review and evidence work; it does not classify a system or give legal advice.
The normalization rule
Set one shared comparison period, such as the first 12 months from the date the pilot enters production. Write down the start event and date range. Then calculate:
Comparable first-year cost = one-time external fees + recurring external fees + usage-based fees + expected change work + internal effort + required third-party work + transition/exit allowance
Use the same tax basis for every offer. If VAT is excluded, exclude it consistently and ask the finance owner to confirm treatment. Do not convert currencies using an unstated rate; write the rate and date if a quote is in EUR or USD. Keep uncertainty beside each line rather than burying it in a footnote.
This is a management estimate, not an accounting valuation. It does not estimate economic benefit, probability of success, or the value of avoided risk. Those belong in separate columns. A low price does not prove a good result, and a higher price does not prove quality.
For each line, record four fields: quoted amount, normalization assumption, evidence status, and owner. The evidence status should be confirmed, estimated, or unknown. The owner should be vendor, your team, or a third party. Unknown is not zero. If a vendor cannot price a line yet, leave it unknown and ask what information would make it priceable. A precise-looking total with many unknown lines is not precise.
The 14-line quote worksheet
| Line | What to ask both providers | Evidence to request | Common comparison trap |
|---|---|---|---|
| 1. Discovery and workflow mapping | Which process, users, locations, and exceptions are in scope? | Named workshop outputs and process map | One offer includes interviews; the other assumes the buyer supplies a complete map |
| 2. Baseline measurement | How will current time, volume, error, and rework be measured? | Baseline definition, data source, sampling period | Comparing a measured baseline with an unsupported savings assumption |
| 3. Data preparation | Who cleans, labels, masks, deduplicates, and validates input data? | Data assumptions and acceptance checklist | “Data readiness” means different things |
| 4. Design and configuration | What workflows, prompts, rules, and user roles are configured? | Versioned configuration list and revision rounds | Treating a workshop prototype as production configuration |
| 5. Software and model access | Which subscriptions, seats, API plans, or model services are required? | Product, tier, seat count, limits, renewal terms | “Included” means access for a small test, not production |
| 6. Usage and overages | What happens when requests, tokens, storage, or processing exceed the allowance? | Unit, cap, rate, alert threshold, owner | Assuming usage is unlimited because the pilot is small |
| 7. Integration | Which systems are connected, by which interface, and who owns credentials? | Interface list, field mapping, test environment | Comparing a manual export with a live integration |
| 8. Security and privacy work | What review, documentation, access control, and contract work is included? | Deliverable list, reviewer, exclusions | Treating “compliance included” as a legal conclusion |
| 9. Testing and correction | Which test cases, defect classes, and correction cycles are included? | Test plan, defect log format, retest allowance | One quote covers a demo; the other a bounded correction cycle |
| 10. Training and handover | Who is trained, for how long, and what can they operate afterwards? | Materials, recipient, runbook, session scope | Counting a presentation as a handover |
| 11. Internal staff time | What hours must the buyer provide for interviews, review, tests, and approvals? | Role-by-role estimate and calendar assumptions | External fees are compared while employee time is free |
| 12. Support and monitoring | What response window, monitoring, maintenance, and incident handling exist? | Service hours, escalation route, included volume | Comparing a retainer with a one-off project that has no support |
| 13. Change requests and renewal | What counts as a change, what is included, and how are rates revised? | Rate card, approval process, renewal notice | Low initial fee moves work to unbounded hourly changes |
| 14. Exit and portability | How are data, configurations, accounts, logs, and credentials returned or removed? | Export format, assistance hours, deletion process | Exit cost is omitted because it occurs after the quote period |
Ask each provider to mark every line included, excluded, allowance, or unknown. Included should name a quantity or boundary. “One connector, one test environment, and two correction rounds” is more useful than “integration included.” If a provider refuses the shared template, keep the offers separate and record the refusal as a procurement fact; do not infer that the more detailed provider is automatically better.
How to use the lines without turning them into a vendor score
These are cost categories, not quality scores. A provider may reasonably price discovery separately because the problem is not understood. Another may offer a fixed first phase. The comparison is useful only if it preserves the offer’s actual structure. Do not rewrite a conditional estimate into a fixed commitment.
Use a status vocabulary that can be audited:
- Confirmed: contract or quote names amount and scope.
- Estimated: amount depends on a stated quantity, rate, or assumption.
- Unknown: provider has not supplied a defensible amount or boundary.
- Buyer-owned: work is not in the vendor fee and your team plans to do it.
- Third-party: another supplier or professional must be engaged.
For every estimate, keep both a low and high scenario if the relevant quantity is uncertain. A monthly fee can be exact while actual usage is uncertain. A fixed project fee can be exact while the number of change requests is not. These are different uncertainties; do not compress them into one contingency percentage unless the buyer can explain its basis.
The buyer’s one-page request
Send both providers the same short input pack: workflow, users, systems, target operating date, definition of success, and exclusions. Ask them to return the 14-line table with one-time, monthly, usage-based, and buyer-supplied work separated. Include currency, tax basis, quote validity, renewal date, and assumptions that would change the estimate.
Keep the scope pack brief enough to answer. Do not send live personal, customer, financial, or confidential data merely to get a preliminary estimate. Use synthetic or redacted examples until the parties agree on data handling and contract route. The MTF webinar agenda discusses confidentiality, contracts, and data hygiene as SME topics; its page is a vendor event description, not independent legal review.
Ask the provider to identify who will perform each material task, not just the account lead who presents the offer. If subcontractors are expected, request their role and the deliverables they touch. Outsourcing may be acceptable, but its effort and interfaces should not disappear from the model.
A worked example with synthetic quotes
The numbers below are invented solely to demonstrate arithmetic. They are not observed Swiss rates, a benchmark, a real quote, or a recommendation. The assumed scope is one internal document workflow, one system integration, a 12-month comparison period, and a buyer-defined internal hourly cost of CHF 90. The model excludes taxes, financing, benefits, and costs that cannot be estimated without discovery.
| Cost component | Hypothetical Offer A | Hypothetical Offer B |
|---|---|---|
| One-time implementation | CHF 14,400 | CHF 11,500 |
| Recurring platform and monitoring | CHF 850 × 12 = CHF 10,200 | CHF 1,250 × 12 = CHF 15,000 |
| Buyer staff effort | 32 h × CHF 90 = CHF 2,880 | 18 h × CHF 90 = CHF 1,620 |
| Change-work scenario | 30 h × CHF 160 = CHF 4,800 | 20 h × CHF 160 = CHF 3,200 |
| Exit assistance scenario | 20 h × CHF 160 = CHF 3,200 | 10 h × CHF 160 = CHF 1,600 |
| Synthetic first-year comparison | CHF 35,480 | CHF 32,920 |
The table makes Offer B appear CHF 2,560 lower under these assumptions. That arithmetic does not establish that B is cheaper in practice. Recurring and labour assumptions are fictional, change and exit hours are scenarios, and none of the figures has been verified with a supplier. If the buyer needs a security review, additional connector, more seats, or weekend support, the comparison changes. If the workflow cannot be measured or is a poor fit for AI, both offers may be poor investments.
Use the example to ask better questions, not to anchor a negotiation. Replace every number with an actual quote, clearly labelled estimate, or unknown. Then calculate a range with the same low, base, and high assumptions for both offers. If a line is unknown for one provider but priced for another, show it as an open decision rather than silently giving the unknown offer a cost advantage.
A second useful view is cost per accepted unit of work. Define accepted before calculating it: for example, a record counts only when required fields are correct, a reviewer approves it, and the result reaches the destination system without a duplicate. Then use:
Cost per accepted unit = comparable period cost / accepted units in that period
This ratio is meaningful only if both offers use the same denominator and quality threshold. A system that processes more items but creates more rework is not necessarily less expensive. Include human review and correction work consistently. Do not label a test set as a representative operational sample.
A small executable worksheet
The following standard-library Python program calculates the illustrative first-year total from a CSV. It rejects blank, negative, duplicate, or malformed rows. It deliberately does not fill missing prices with zero; unknown items should remain visible in the buyer’s separate register.
Create quote.csv:
offer,category,amount_chf,period,certainty
A,implementation,14400,once,synthetic
A,platform_and_monitoring,850,monthly,synthetic
A,buyer_staff_hours,2880,once,synthetic
A,change_work,4800,once,synthetic
A,exit_assistance,3200,once,synthetic
B,implementation,11500,once,synthetic
B,platform_and_monitoring,1250,monthly,synthetic
B,buyer_staff_hours,1620,once,synthetic
B,change_work,3200,once,synthetic
B,exit_assistance,1600,once,synthetic
Save this as normalize_quotes.py:
import csv
from collections import defaultdict
from decimal import Decimal, InvalidOperation
MONTHS_IN_COMPARISON = Decimal("12")
allowed_periods = {"once", "monthly"}
totals = defaultdict(Decimal)
seen = set()
with open("quote.csv", newline="", encoding="utf-8") as handle:
for row_number, row in enumerate(csv.DictReader(handle), start=2):
offer = row["offer"].strip()
category = row["category"].strip()
period = row["period"].strip().lower()
if not offer or not category:
raise ValueError(f"row {row_number}: offer and category are required")
if period not in allowed_periods:
raise ValueError(f"row {row_number}: period must be once or monthly")
key = (offer, category)
if key in seen:
raise ValueError(f"row {row_number}: duplicate offer/category {key}")
seen.add(key)
try:
amount = Decimal(row["amount_chf"].strip())
except (InvalidOperation, AttributeError):
raise ValueError(f"row {row_number}: amount_chf must be numeric")
if amount < 0:
raise ValueError(f"row {row_number}: negative amounts need explicit treatment")
first_year = amount * MONTHS_IN_COMPARISON if period == "monthly" else amount
totals[offer] += first_year
for offer in sorted(totals):
print(f"{offer}: CHF {totals[offer]:,.2f} first-year synthetic total")
if not totals:
raise SystemExit("No quote rows found")
Run python normalize_quotes.py. The program’s output should match the synthetic table. It is intentionally simple: it does not decide whether a line is in scope, whether a recurring charge is optional, whether usage is capped, or whether an entered estimate is credible. A spreadsheet can calculate arithmetic; a human must validate contract meaning and assumptions.
For a real procurement, add columns for source document, quote version, date, tax treatment, owner, confidence, low/base/high values, and evidence link. A new revision should create a new version of the comparison rather than overwrite the old quote. That makes it possible to explain why the approved budget changed.
Make the scope equal before comparing price
An apples-to-apples comparison begins with a common outcome statement. Avoid a broad phrase such as “automate document processing.” Write down incoming document types, volumes, languages, required fields, destination system, exception path, reviewer, and what the workflow must not do. State whether the system only prepares a draft or can write to a live record. Those are materially different scopes.
Define the pilot boundary too. Is it a sandbox, a production test with a restricted user group, or a live operational service? Does the quote include test accounts, permissions, rollback, monitoring, and a way to disable automation? If one proposal includes a guarded launch and the other stops at a demo, their totals are not directly comparable even if both use the word pilot.
Ask for assumptions in checkable form: buyer supplies a clean export with these fields, provider configures one connector, or the estimate excludes historical data cleanup. Avoid accepting “normal data quality” without a definition. If the input is not ready, price a discovery or data-assessment phase separately rather than pretending implementation can be estimated precisely.
Tie each included deliverable to recipient-run evidence: test results, configuration inventory, operating instructions, or a handover session. This does not turn cost comparison into a generic acceptance test; it confirms what the fee purchases. Keep acceptance criteria separate from the cost model. A deliverable can be included and still fail quality review; a higher fee does not guarantee acceptance.
Contract and compliance questions to price, not improvise
The buyer should not ask a software vendor to certify a legal position through a sales quote. Instead, list the review work that may be needed, the professional who owns it, and whether it is included or separately priced. This may include privacy review, security assessment, contract review, or evaluation of obligations for a use case that affects people or customers. Scope depends on the activity and data, so a generic “compliant AI” line is not a substitute for qualified advice.
If a feature is used in a context covered by the EU AI Act, Article 50 addresses transparency duties for specified providers and deployers, including some interactions and generated content. The exact duty depends on role and system. A Swiss company should not assume the Act applies to every internal tool simply because it uses AI, nor assume it is irrelevant solely because the company is Swiss. Ask counsel to assess the service, territory, and user relationship against the current text. The official Article 50 text is a starting point, not the conclusion.
A Swiss SME webinar published by MTF in October 2025 paired a lawyer and technology executive and listed data protection, professional secrecy, contract drafting, EU AI Act developments, configuration, anonymization, and internal rules as topics. The page says audience questions were included but does not reproduce them; it supports the relevance of those topics to a Swiss SME discussion, not a claim about what most buyers ask. The event description and recording are vendor-hosted, so keep that limitation visible.
On privacy, a cost sheet can identify review work and assign an owner; it cannot decide lawfulness. Write down which party prepares the data map, identifies relevant subprocessors, documents data flow, confirms retention and deletion behavior, and obtains advice where needed. Avoid describing data-center location as a complete privacy or compliance solution. Location, access, contract terms, security, subprocessors, and actual data use are separate questions.
For security, ask who creates and controls production accounts, how access is granted and removed, who can view logs, how credentials are transferred, and how incidents are handled. If those activities are not priced, record them as buyer-owned or unknown. Do not give a provider live credentials just to reduce quote friction. A discovery can use synthetic records while contract, identity, and access controls are agreed.
Why internal time belongs in the total
In a small company, one employee may act as process owner, data steward, reviewer, IT contact, and approver. That effort is real even when no invoice arrives. KOF’s analysis of Swiss innovation survey data for 2020–2022 noted a gap between smaller and larger firms in AI and big-data use, and cited cost and skilled-worker constraints. It does not say how many hours a current project needs. It reinforces the practical point that an offer assuming unlimited access to scarce staff may be cheaper only on paper.
Estimate hours by role rather than using one generic client-contribution line. For example: operations lead for workflow mapping; subject-matter expert for examples and exceptions; IT for access and integration; privacy or legal reviewer for data-flow and contract review; frontline user for testing; manager for approval. Record the estimate and update it when the provider’s actual requests become clear.
Do not inflate the estimate with arbitrary salaries. If the business has an approved internal loaded hourly rate, use it and record its source. If not, show hours separately and leave the money value blank. An explicit 24-hour estimate is more honest than currency built on an invented wage.
The KOF report also describes firms adapting or developing AI software for their own needs, while many firms did not. The correct conclusion is not that every SME should build in-house. The quote should explain which future operating tasks stay with the buyer and which require paid support. A solution with modest technical demands and good documentation may be maintainable internally; a connected workflow with changing data sources may need a service contract.
Check recurring costs under more than one usage pattern
A small pilot can make consumption look negligible. Ask the provider to calculate at least three scenarios using your estimated volume: expected use, a plausible busy period, and an upper bound that triggers an alert or approval. The model input should include the unit that generates the charge: request, token, minute, document, seat, storage, or connector. Do not mix requests and accepted work as if they were the same thing.
For each variable cost, capture its control mechanism. Can the buyer set a budget cap? Can the system pause or downgrade gracefully? Who receives the alert? What happens if the cap is reached outside business hours? These are not merely technical details; they determine whether cost is predictable and who must respond.
Separate price risk from outcome risk. A contract can limit spending while the workflow still produces errors. Conversely, high usage charges may accompany a useful system. The comparison should show both financial exposure and the evidence used to judge whether the workflow is useful.
Do not assume monthly support includes model upgrades, connector changes, bug fixes, security reviews, or staff training. Ask which events are included, which are charged separately, and what response times are promised. A service-level promise is meaningful only when its hours, severity definitions, and exclusions are written down.
Make the decision without inventing a winner
Once the offers are normalized, use a three-column decision sheet:
- Comparable cost: first-year low/base/high scenarios under shared assumptions.
- Scope and evidence: what each provider will actually deliver and how the buyer can verify it.
- Residual unknowns: unresolved dependencies, unpriced work, legal review, or operational ownership.
Do not average those columns into a single score unless the decision team has explicit weights and can defend them. A weighted score can create false precision when offers differ in value or risk. A simple log is more useful: “Offer B has a lower synthetic base-case estimate, but the support scope and exit allowance remain unknown; request clarification before approval.”
A provider may propose a paid discovery phase before giving a fixed implementation quote. That is not automatically a red flag. A real workflow can contain unknown systems, poor data, or exceptions that make a confident fixed quote impossible. The buyer can ask the provider to cap discovery, specify deliverables, and identify what decision the phase will enable. If the provider refuses to define the output or how it changes the estimate, uncertainty remains.
Likewise, time-and-materials is not inherently inferior to a fixed fee. It can be reasonable when scope is uncertain, but it needs a budget ceiling, named approval path, rate card, and reporting interval. A fixed-price offer can shift risk into exclusions, change requests, or narrow boundaries. Compare the whole commercial mechanism, not the label.
When neither provider can state a credible recurring cost, request a bounded measurement phase. Define what will be measured, for how long, what data is used, how the buyer can stop the work, and what result permits a larger commitment. Do not treat a demo’s usage or a week of manual testing as a full-year operating forecast.
Failure modes and what to do instead
A single total hides exclusions. Return to the 14 lines and ask the provider to classify each one. Unknown entries stay visible.
The quotes describe different outcomes. Rewrite the scope in buyer language, send the same brief to both suppliers, and ask each to identify what it cannot cover.
Recurring fees are treated as future IT. Add subscriptions, hosting, usage, support, renewals, monitoring, and maintenance to the first-year model.
The buyer’s time is zeroed out. Estimate hours by role. If rates are unavailable, show hours without inventing CHF totals.
A percentage contingency hides its basis. Replace it with named scenarios such as extra data-cleaning hours or a second connector. Keep a percentage only if its basis is explicit.
The cheapest line wins although scope is smaller. Compare deliverables and constraints before totals. If they differ, show separate options rather than a false head-to-head.
A sales claim becomes the benefit forecast. Measure a workflow baseline and define accepted work. Treat savings as a hypothesis until measured.
A free trial uses sensitive production data. Stop and agree data handling, access, contract, and review first. Use synthetic or redacted examples for initial demonstrations.
No one owns recurring operations. Name an internal owner for access, exception handling, monitoring, and provider escalation before launch.
Exit is omitted. Ask for data export formats, configuration transfer, account control, deletion confirmation, and assistance fees. Do not assume a spreadsheet is enough to recreate a workflow.
The worksheet is treated as legal advice. Use it to expose review work; have counsel assess applicable obligations and contract language.
Synthetic values become a budget promise. Remove them from the approval pack or label them placeholders until replaced by actual quotes and internal assumptions.
A simple decision flow
flowchart TD
A[Define one workflow and period] --> B[Send the same scope pack to each provider]
B --> C[Map each offer to the 14 cost lines]
C --> D{Are outcome and assumptions comparable?}
D -- No --> E[Request clarification or keep offers separate]
D -- Yes --> F[Calculate one-time, recurring, usage, buyer, and exit costs]
F --> G{Are material lines confirmed?}
G -- No --> H[Label unknowns and price a bounded discovery step]
G -- Yes --> I[Review scope, evidence, risk, and operating owner]
H --> I
I --> J[Approve, renegotiate, defer, or decline with a written reason]
This flow makes stop conditions explicit. If the workflow is not defined, do not compare totals. If material assumptions remain unknown, buy information only through a bounded phase. If no operating owner exists, the project is not ready for production regardless of which offer looks cheaper.
Reusable approval checklist
- [ ] Both providers priced the same workflow, users, period, and target outcome.
- [ ] One-time, monthly, and usage-based charges are separate.
- [ ] Currency, tax basis, quote date, validity, renewal, and rate revisions are recorded.
- [ ] Each of the 14 lines is marked included, excluded, allowance, or unknown.
- [ ] Included names a quantity, boundary, or deliverable.
- [ ] Usage assumptions include expected and high scenarios with a cap or alert.
- [ ] Internal hours are shown by role, even if the company cannot price them yet.
- [ ] Data, security, legal, and contract review have named owners.
- [ ] The buyer knows which production accounts and credentials will be needed.
- [ ] Support hours, escalation, monitoring, maintenance, and changes are defined.
- [ ] Exit and portability are priced or explicitly unknown.
- [ ] Evidence for judging the workflow is defined before the project begins.
- [ ] Regulatory questions go to qualified counsel rather than being inferred from a vendor label.
- [ ] Synthetic examples are not mistaken for market data or a supplier quote.
- [ ] The approval note preserves assumptions and decision date.
Method, exclusions, and limits
The Search Console observation was retrieved read-only from the IAPME Suisse property for Switzerland, with query and page dimensions, country filter CHE, and dates 13 October–10 November 2026. It contains two query-page rows cited here. Search Console reports impressions, clicks, and average position for the selected property and period; average position is not a ranking guarantee, and these rows do not establish the size or intent of the whole Swiss market.
The 14-line worksheet and Python example are original editorial artifacts created for this article. The example contains two fictional offers and invented amounts solely to demonstrate arithmetic; it includes no real vendor, buyer, project result, market price, or measured dataset. The code verifies arithmetic and duplicate rows only; it cannot validate a quote, scope, compliance position, or supplier claim.
Source selection was limited to the IAPME partner guide, KOF’s discussion of Swiss innovation-survey findings, BCG’s global consulting research, the MTF webinar page, and the European Commission’s Article 50 text. The MTF page describes an event agenda and does not provide the audience-question transcript. BCG’s report is consulting research rather than an independent Swiss SME survey. KOF’s figures describe 2020–2022 and must not be presented as 2026 values. I did not use inaccessible government pages or pages without a verified self-canonical as factual authorities.
This method fails when the work is an off-the-shelf subscription with no supplier proposal, when the use case is too risky to scope without specialist advice, or when two providers propose fundamentally different outcomes. In those cases, compare the relevant products or risks separately, and commission bounded discovery only if it produces a concrete decision artifact.
References
- IAPME Suisse: AI Consulting for Swiss SMEs—complete partner guide — primary source page for the related comparison guide; group-owned editorial and consulting content.
- KOF Swiss Economic Institute, ETH Zurich: Digital technology and small firms — Swiss firm data for 2020–2022 and stated limitations.
- BCG: Where’s the Value in AI? — consulting firm’s global research and people/process observations.
- MTF Solutions: AI and compliance webinar recording — Swiss SME event page; agenda and recording, not independent legal authority.
- European Commission AI Act Service Desk: Article 50 — official text for specified transparency duties; application depends on role and scope.
Top comments (0)