Optimize Desk Blog

AI readiness assessment: what you should get for the money

Wed, 21 Oct 2026 08:00:00 +0000

A fixed-price AI readiness assessment costs USD 2,500 to 10,000 and takes three days to eight weeks; a consultancy programme is quoted at USD 50,000 to 500,000 over three to six months. Whatever you pay, it should give you five things, and the most useful of them is a verdict per role with the human share written next to it.

What do the three price levels deliver?

Price level	Example and published price	Time	What you receive
About USD 2,500	Acropolium, USD 2,500 to 5,000 fixed (package page)	3 to 5 business days	A workflow audit, the top three AI opportunities ranked by return, a risk and compliance check and a roadmap; the fee is credited in full against a build
About USD 10,000	Main and Machine, USD 3,500 to 8,500 for firms of 5 to 100 people (service page); Vantage Point, USD 7,500 (package page); Mitori SMB audit from USD 10,000 (pricing)	2 to 8 weeks	A map of the workflows examined, ranked opportunities with effort and payoff ranges, up to 6 interviews and 10 scored use cases, an ROI model and a phased plan; two of the three credit the fee against later work
USD 150,000 and up	Mitori describes the consultancy alternative as USD 50,000 to 500,000 over 3 to 6 months (pricing); UK G-Cloud 14 ceiling day rates run GBP 1,425 to 2,450 at Deloitte and GBP 1,280 to 2,750 at PwC (Consulting Demand)	3 to 6 months	A maturity benchmark, a transformation programme and a team on site; two consultants for eight weeks at the lowest Deloitte rate is GBP 114,000, about USD 152,000 at the ECB reference rates of GBP 0.837 and USD 1.119 per euro (ECB)

Two notes on the table. Enterprise workforce tools sit outside it: Draup, for example, offers a free 14-day "board-ready AI transformation roadmap" and scores tasks as automate, augment or human, but its worked examples are large organisations (Draup). In Germany the BAFA programme pays small and medium-sized companies 50 percent of up to EUR 3,500 of consulting cost, for up to two consultations a year and five in total, until 31 December 2026 (BAFA; IHK Nürnberg), which halves the bill at the first price level.

What must any assessment contain?

A task list per role, with hours. Not a list of use cases. The public occupational database holds 18,838 task statements under an open licence (O*NET); an assessment that starts from your job descriptions and maps them to it can be checked line by line.
A split of every task into three classes: AI runs it, AI drafts and a person finishes, stays with people. The method should be public. The exposure rubric from Eloundou and colleagues, which asks whether a language model can halve the time for a task at the same quality, is published (arXiv); the observed-usage data in Anthropic's Economic Index (Hugging Face) and the Microsoft applicability scores for 785 occupations (GitHub) are free downloads. If the assessor cannot name the method, you are buying an opinion.
The cost of each role from official statistics, with employer costs. In the United States, median pay comes from the Bureau of Labor Statistics and employer costs add about 43 percent on top of wages (BLS ECEC). An assessment that uses your payroll instead is fine; one that uses neither is guessing.
The tool that could take the automatable share, and its published price at your volume. Without this line the saving cannot be computed.
A verdict per role: go or no-go, the human share that remains, a confidence level and the legal flag for your country. The roles that do not qualify belong in the report too.

How do you read a per-role verdict?

Read the confidence level first. A high verdict means your own volumes confirmed the split and at least two public signals agree; a medium one rests on public data for the role type; a low one rests on an interview. Then read the published range as a ceiling, never as a promise: 76 percent of support conversations resolved is a vendor average across more than 12,000 customers (Fin), not your number until a pilot measures it. Then read the human share. Every role that qualifies keeps a person for exceptions, sign-off or the desk itself; if the share is written as zero, the assessment is selling, not assessing. Finally, check that the verdict is about the role and its tasks, not about the person who holds it. That is what keeps the decision with you and, in Europe, outside the high-risk rules of the AI Act.

Optimize Desk's role scan contains these five things for every role in a 20 to 200 person company: Find out which roles AI can take. Then let it.

Before you cut a role for AI: seven numbers to check first

Fri, 16 Oct 2026 08:00:00 +0000

Of the business leaders who made people redundant because of AI, 55 percent now say the decision was wrong, and 32 percent of US employers who cut a role after bringing in AI have rehired for it. Seven numbers, measured before anyone leaves, separate the cuts that hold from those that come back.

How often do AI cuts get reversed?

Orgvue surveyed 1,000 senior decision makers in April 2025: 39 percent had made employees redundant because of AI, and 55 percent of those admit wrong decisions (Orgvue).
Forrester's Predictions 2026 expects half of AI-attributed layoffs to be reversed, with much of the work rehired offshore or at lower pay (The Register).
Robert Half reports 32 percent of US employers who eliminated positions after implementing AI have since added them back (Robert Half); rehiring was highest in finance at 44 percent, HR at 35 percent and technology at 32 percent (TechSpot).
Gartner predicts that by 2027 half of organisations that planned to cut customer service staff significantly will abandon the plan; 95 percent of 163 service leaders polled plan to keep human agents (CMSWire).
Commonwealth Bank of Australia reversed 45 redundancies tied to a voice bot in August 2025, saying its assessment "did not adequately consider all relevant business considerations and this error meant the roles were not redundant" (ACS Information Age).
The base rate is low. In the New York Fed's September 2026 survey, 4 percent of service firms had laid anyone off because of AI (New York Fed).

What happened at Klarna, in order?

Date	What Klarna said	Source
27 February 2024	The AI assistant handled 2.3 million conversations in its first month, two-thirds of customer service chats, "the equivalent work of 700 full-time agents", and cut resolution time from 11 minutes to under 2	Klarna
9 May 2025	The chief executive said cost "seems to have been a too predominant evaluation factor" and "what you end up having is lower quality"; Klarna started hiring remote human agents again	Fortune
14 May 2026	Revenue per employee reached nearly USD 1.4 million, four times the 2022 level	Klarna

The assistant was never switched off. The chat volume stayed with the AI; Klarna rebuilt the human path for the conversations the AI handled badly. Every reversal above follows this pattern: the volume moved, the human share was cut to zero, and it came back at a price.

Which seven numbers should you have before you decide?

#	Number	How you get it	Benchmark
1	Baseline volume	Count 30 days of calls, tickets, invoices or entries for the role, by type	Your own data; no benchmark replaces it
2	Containment	The share of items the AI finishes with no person involved, measured in a pilot	Published ceilings: 76% of support conversations (Fin), 71% of documents straight through (Rossum), 85 to 90% of bookkeeping entries categorised (Puzzle)
3	Error rate	Sample the AI's finished items; set the bar before the pilot starts	McDonald's ended its voice ordering test at low-to-mid 80s percent accuracy against a 95% bar (NBC News); Air Canada was ordered to pay CAD 812 after its chatbot gave a wrong fare rule (CBC)
4	Exception hours	The hours a person spends each month on what the AI hands back	A hybrid phone service charges from USD 2.50 a call against USD 1 to 2 for AI alone (Smith.ai)
5	Loaded cost of the role	Median pay for the occupation plus employer costs	A US customer service representative earns USD 44,770 a year at the median (BLS); employer costs make the total about 1.43 times pay (BLS ECEC), USD 5,335 a month
6	Vendor cost at your volume	Price per outcome or per minute times your baseline	USD 0.99 per resolution (Fin pricing); USD 1.50 to 2.00 per automated resolution (Zendesk); USD 149 for 1,000 phone minutes (Rosie)
7	The human share that remains	100 minus containment, turned into hours and a named owner	20 to 35% of support conversations in the published figures above

What is the decision rule?

A worked example, with the assumptions in the open. A support role handles 1,000 conversations a month (number 1). In a 60-day pilot the AI finishes 76 percent of them (number 2) at an error rate inside your bar (number 3). The 240 conversations it hands back take a person 10 minutes each, 40 hours a month (number 4). The role costs USD 5,335 a month loaded (number 5). The vendor charges 760 outcomes at USD 0.99, USD 752 (number 6). The 40 hours cost about USD 1,232 at the same loaded rate of USD 30.79 an hour. Net saving: about USD 3,350 a month, only if a named person owns those 40 hours and the escalations (number 7).

The rule has three parts. Cut nothing on a vendor's average, which is a ceiling from thousands of mature customers. Cut only after 60 days of measured containment and error rate in your company. If the human share has no owner, you have not removed a role; you have removed the person who caught the mistakes.

Optimize Desk measures these seven numbers in a 90-day pilot before any role changes, and you decide about people, always: Find out which roles AI can take. Then let it.

The EU AI Act for a company of 50 people: what applies to AI in your office, and from when

Thu, 08 Oct 2026 08:00:00 +0000

If your company has 50 people and uses AI in the office, three rules of the EU AI Act already bind you: no AI that infers the emotions of your staff (since 2 February 2025), a duty to support AI literacy among staff who use AI (same date), and any AI that talks to customers or staff must make clear it is an AI (since 2 August 2026). The heavy rules for high-risk AI, including AI used in employment decisions, apply from 2 December 2027, after the 2026 omnibus regulation moved that date back from 2 August 2026.

There is no exemption for small companies. The Act reaches anyone who uses an AI system under their authority in the course of business, which it calls a deployer (Article 3). Only the fine changes with size: small and medium-sized companies face the lower of the two ceilings (Article 99).

Which rules already apply to AI in your office?

Rule	Where	Since	What a 50-person company does
No AI that infers the emotions of people at work, except for medical or safety reasons	Article 5(1)(f)	2 February 2025	Do not buy tools that score mood or stress in calls, emails or on camera, including inside ticket analytics you already own
Support AI literacy among staff who use AI	Article 4	2 February 2025	Short training and written rules for the people who use AI tools; the amended text says no specific level has to be guaranteed for any individual
AI that interacts with people must make clear that it is an AI	Article 50(1)	2 August 2026	Phone, chat and email agents say so unless it is obvious; check the setting on every vendor tool
Fines	Article 99	In force	Up to EUR 35 million or 7% of turnover for prohibited practices, EUR 15 million or 3% for other breaches; an SME pays the lower of the two

The dates are the European Commission's official timeline (Commission). The law is Regulation (EU) 2024/1689 (EUR-Lex); the article links go to the consolidated text with the 2026 amendments marked.

Which uses of AI at work count as high-risk?

Annex III, point 4, lists the employment uses. Point 4(a) covers recruitment and selection: targeted job advertisements, filtering applications, evaluating candidates. Point 4(b) covers AI "intended to be used to make decisions affecting terms of work-related relationships", promotion or termination, allocating tasks "based on individual behaviour or personal traits or characteristics", and monitoring or evaluating the performance and behaviour of people at work (Annex III).

Article 6(3) takes some tools out: a system that performs a narrow procedural task, or a preparatory task for such an assessment, is not high-risk. The same paragraph closes one door: an Annex III system "shall always be considered to be high-risk where the AI system performs profiling of natural persons" (Article 6). A tool that scores tasks and roles is preparatory; a tool that scores named employees is high-risk, whatever it is called.

From 2 December 2027 a company that uses a high-risk system must use it as the provider's instructions say, assign human oversight to trained people, monitor it, keep its logs for at least six months and, before it goes into service at the workplace, inform workers' representatives and the affected workers (Article 26).

What changed with the 2026 omnibus?

Regulation (EU) 2026/1744 of 8 July 2026 entered into force on 27 July 2026 (AI Act Explorer; Commission). It moved the application date for Annex III high-risk systems from 2 August 2026 to 2 December 2027, and for AI built into regulated products to 2 August 2028 (Article 113). It added new prohibitions that apply from 2 December 2026 and rewrote the literacy duty so that companies "take measures to support the development of AI literacy" (Article 4). The transparency rule and the existing prohibitions were not delayed.

What must an employer do when AI changes a role?

Four things. None is new law for a European employer, and all are easy to forget when a vendor promises savings.

Decide about people yourself. Under GDPR Article 22 a person has the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects (GDPR). A dismissal or a change of role decided by software is such a decision. Let the tool assess tasks and roles; a manager decides about the person.
Inform and consult early. In Germany the works council must be informed in good time about the planning of work processes "einschließlich des Einsatzes von Künstlicher Intelligenz", including the use of artificial intelligence, and consulted on the effects (§ 90 BetrVG). In a business with more than 20 voting employees, fundamentally new working methods are a Betriebsänderung that triggers information and consultation (§ 111 BetrVG).
Make the AI say it is an AI on every channel where it meets a customer or a colleague (Article 50).
Keep people's data out of the analysis. Role-level task lists and volumes answer the question. Per-person performance data turns an assessment into monitoring and, from December 2027, into a high-risk use.

Optimize Desk assesses tasks and roles, never people, and every AI role it runs says that it is an AI: Find out which roles AI can take. Then let it.

Which office roles can AI take today? The evidence by role

Tue, 29 Sep 2026 08:00:00 +0000

On published evidence, AI can take most of the volume in three office roles today: tier-1 support by chat and email, document processing and data entry, and bookkeeping preparation. A person still handles about 10 to 35 percent of the work in each, phone work has thinner evidence, and for outbound sales, software, legal and analysis work the measured results say assist, not replace.

Which roles have published automation rates?

The table lists only numbers a vendor, customer or research group has put in writing. They come from mature deployments, so read them as ceilings, not as a first month.

Role type	Published rate	What stays with people
Tier-1 support by chat and email	76% of conversations resolved, averaged across 12,000+ customers (Intercom Fin); 70% at Chime and an 80% deflection rate at Duolingo (Decagon); two-thirds of chats in the first month at Klarna (Klarna, February 2024)	20 to 35% of conversations: complaints, money, anything that needs a policy judgement
Document processing and data entry	71% straight-through processing at Morton Salt and 90 to 93% accuracy on new document types (Rossum); 67% of data entry keyer tasks show observed AI use (Anthropic, March 2026)	A 7 to 30% exception queue and every new document type
Bookkeeping preparation	85 to 90% of transactions categorised automatically, and nothing posts until a person approves it (Puzzle)	The close, judgement calls, tax and audit sign-off
Phone support, tier 1	50% and more deflection on voice at Valon, 70% across chat and voice at Chime (Decagon); McDonald's ended its drive-through test at low-to-mid 80s percent accuracy against a 95% bar (NBC News)	Upset callers, complex cases, regulated topics

A caution on the first row: in May 2025 Klarna's chief executive said cost had been "a too predominant evaluation factor" and the result was "lower quality", and Klarna hired human agents again (Fortune). The volume stayed with the AI; the human path was rebuilt.

Which roles are assist only on the evidence?

Role type	Measured result	Source
Outbound sales	The vendor's own expectation is 1 to 3 qualified meetings per 100 leads	AiSDR
Software development	16 experienced developers took 19% longer with AI tools on 246 real issues while believing they were 20% faster; the best agent finished 30% of simulated company tasks on its own	METR; TheAgentCompany
Legal research	Incorrect information more than 17% of the time (Lexis+ AI) and more than 34% (Westlaw AI-Assisted Research) in a test of over 200 queries	Stanford HAI
Analysis inside business systems	About 58% success on single-turn tasks, about 35% on multi-turn tasks	CRMArena-Pro

These are drafting and research aids. Nobody publishes a rate at which they finish the job alone; the error rates above are the reason.

Which roles have no evidence yet?

Front desk and phone intake is the biggest gap. AI receptionist vendors publish prices, answer rates and cost per call, but not the share of calls they finish without a transfer, and Smith.ai says most firms start on its hybrid plan with people behind the AI (Smith.ai). Content drafting has no containment figure because an editor approves every piece. Recruiting screening tools make result claims, but recruitment and selection sit on the EU AI Act's high-risk list for employment, so a person decides about a candidate whatever the tool can do (Annex III). For these roles the honest number is the one your pilot measures.

How should you read the task-exposure studies?

Three kinds of numbers get mixed up. Exposure studies measure what could be done. The 2023 paper by Eloundou, Manning, Mishkin and Rock found about 80 percent of US workers could have at least 10 percent of their tasks affected by language models, about 19 percent at least half, and that 15 percent of all tasks could be done significantly faster by a model alone, rising to 47 to 56 percent with software built on it (arXiv).

Usage studies measure what people hand over. Anthropic's March 2026 analysis found observed coverage of 75 percent of tasks for computer programmers and 67 percent for data entry keyers, zero coverage for 30 percent of workers, and no systematic rise in unemployment among exposed workers since late 2022 (Anthropic). The Microsoft team that scored 785 occupations from 200,000 Copilot conversations says its metrics "should not be misconstrued or misrepresented as measuring the ability of AI to replace jobs" (GitHub).

Outcome studies measure what happened. In the New York Fed's September 2026 survey, 61 percent of service firms used AI, 4 percent had laid anyone off because of it, about 15 percent had hired fewer people and about 13 percent more (New York Fed). The Stanford payroll study found employment of 22 to 25 year olds in the most exposed jobs down about 11 percent from November 2022 to June 2026 while the least exposed grew about 10 percent, mostly through fewer hires, not dismissals (Stanford Digital Economy Lab).

Put the usage number, measured in your own company, in front of a payroll decision. The exposure number tells you where to look, not what to cut.

Optimize Desk applies this evidence role by role in companies of 20 to 200 people and runs the roles that qualify with a person on every exception: Find out which roles AI can take. Then let it.

AI receptionist or a human receptionist: the real monthly cost in 2026

Mon, 21 Sep 2026 08:00:00 +0000

A full-time receptionist in the United States costs about USD 4,530 a month with employer costs added to median pay, and in the United Kingdom about GBP 2,465. An AI phone agent costs USD 49 to 349 a month on the plans published in 2026, but it takes the phone, not the front desk.

What does a human receptionist cost per month?

The US Bureau of Labor Statistics puts the 2025 median pay for receptionists at USD 38,010 a year, or USD 3,168 a month (BLS). In its June 2026 employer cost survey, wages were 70.0 percent of what private employers pay per hour worked and benefits 30.0 percent (BLS ECEC). Divide pay by 0.70 for a loaded cost of about USD 4,530 a month.

In the UK, the provisional 2026 survey of the Office for National Statistics gives full-time receptionists a median of GBP 25,865 a year (ONS ASHE 2026, table 14.7a), or GBP 2,155 a month. Employers pay National Insurance at 15 percent above GBP 96 a week (gov.uk) and at least 3 percent into a workplace pension on earnings between GBP 6,240 and GBP 50,270 (gov.uk). That adds about GBP 310 a month, for a loaded cost of about GBP 2,465.

Option	Monthly cost	What you get
Receptionist, US, median pay plus employer costs	USD 4,530	About 173 hours on site: phones, visitors, calendars, post, data entry
Receptionist, UK, median pay plus employer costs	GBP 2,465	The same
Human-staffed answering service	USD 6 to 10 a call, or USD 1.50 to 1.90 a minute	Phones only, billed per call or per minute
AI phone agent, minute plans	USD 49 for 250 minutes to USD 349 for 1,000 minutes	Phones only, around the clock, messages, bookings, transfers
AI phone agent, call plans	USD 79 for 100 callers to USD 299 for 2,000; USD 150 for 75 calls	The same, priced per caller or per call
Hybrid, AI first with people behind it	From USD 2.50 a call	The AI answers, a person takes what it cannot

Sources: the BLS, ONS and gov.uk pages above and the vendor pricing pages linked below, at the time of writing.

What do AI phone agents cost?

Plans come in two shapes. Minute plans: Rosie charges USD 49 a month for 250 minutes, USD 149 for 1,000 and USD 299 for 2,000 (Rosie); Dialzara runs from USD 29 for 60 minutes to USD 349 for 1,000 (Dialzara); My AI Front Desk is USD 99 a month for 200 minutes and USD 0.25 for each extra minute (My AI Front Desk). Call plans: Goodcall charges by unique caller, USD 79 for 100, USD 129 for 500 and USD 299 for 2,000 (Goodcall); Smith.ai's AI receptionist starts at USD 150 a month for 75 calls, USD 2.00 a call, falling to USD 1.67 at 300 calls, with extra calls at USD 2.17 to 2.50 (Smith.ai).

For human answering, Smith.ai quotes about USD 6 to 10 a call, and from USD 2.50 a call on its hybrid plan, which it says most firms start on (Smith.ai). Goodcall puts live answering at about USD 1.50 to 1.90 a minute (Goodcall). One thousand minutes a month would cost USD 1,500 to 1,900 with people and USD 149 to 349 with an AI agent.

What does the AI handle, and what still needs a person?

The BLS lists eight typical duties for a receptionist. Four happen on the phone or in software: answer the telephone and take messages or forward calls, schedule and confirm appointments, enter customer information into the database, handle correspondence. The AI plans cover these. Rosie's plans take messages, text appointment links, book into the calendar from the middle tier up, and transfer a call to a person with a briefing, around the clock, in English and Spanish (Rosie).

The other four need a body in the building: greet visitors, check them in and escort them, tell colleagues who has arrived, copy and file paper documents. No phone agent does these, and no vendor claims to. Vendors also do not publish how many calls their agents finish without a person, and Smith.ai's advice to start on hybrid tells you to expect transfers.

When does each win?

An AI agent wins when the job is the phone: after hours, overflow, missed calls, a company with no front desk, or a second line for bookings. At USD 149 to 349 a month it costs a few percent of a receptionist.
A person wins when visitors walk in, payments are taken at the desk, post arrives, or a caller is upset or in a regulated conversation.
Most offices with a real front desk end in the middle: the person keeps the desk and the daytime phone, and the AI takes the calls that went to voicemail. Measure the share of calls the AI finishes alone before you count any saving.

Optimize Desk scans every role in a 20 to 200 person company, including the front desk, and runs the ones that qualify with a person on every exception: Find out which roles AI can take. Then let it.