// What AI can do
The full range, from asking questions to handing over whole jobs.
Most people picture either a better Google or a machine that knows everything. The real picture sits in between, and it moved a long way in the past year.
// What changed
If you last tried AI seriously a year ago, you used a different tool. Two public tests show how far it has moved.
Single, well-defined tasksShare of real tasks from 44 jobs where AI matched or beat industry experts, graded blind (GDPval)
Whole real projectsShare of real freelance projects AI finished to a standard a client would accept (Remote Labor Index)
Each bar is the best AI model on that test at that date, from different AI companies. Most people’s sense of AI formed at the “then” mark. The sources are in the next section.
// The numbers
What the numbers say. On single, well-defined tasks, today’s best AI matches or beats industry experts most of the time (GDPval). Given a whole real project and left alone, it delivers work a client would accept about one time in five (Remote Labor Index).
How we read it. The model can do the pieces. What it can’t yet do is catch its own mistakes, finish what it starts and judge what good looks like. That is a person’s job. Someone who knows the work and knows how to steer AI turns one-in-five into finished work, across many jobs at once. That’s where the large gains come from, and it’s the skill we build in your team.
“Just under half” is Claude Opus 4.1, the best model in OpenAI’s September 2025 GDPval study. The 84.9% is OpenAI’s own figure for GPT-5.5. The 2.5% was the best score when the Remote Labor Index launched, and 21.25% is Claude Opus 5.5 on Scale AI’s leaderboard today. The two tests measure different things, so linking them is our interpretation.
// The range
AI use runs along four levels. Each one takes more setup and gives back more. Most people stop at the first.
Ask AI a question, or have it write a draft or explain something.
“Explain this clause in plain words.”
Save the instructions for work you do the same way more than once.
The Friday client update, written the same way every week.
Let it read your email, calendar and shared files. You choose which ones. The paid business versions of ChatGPT, Claude and Copilot do not train their models on your files by default.
Tomorrow’s meeting brief, built from the last six months of email.
Let it carry a set job through several steps, then stop for your review.
A proposal drafted from the request, every claim checked against your files, waiting for your sign-off.
The aim is the fourth level. Most teams start at the first two. Getting a team to the fourth takes months of setup and practice, so treat it as where you are heading, over time.
// Our projects
Sales, proposals, follow-ups and admin: the work that fills a normal week. Seven run at Sidekick, and our Inside Sidekick page shows how each one runs, with real outputs where we can share them. The rest are client projects we built or coached. Each is work we can set up in a firm like yours. Figures are each client’s own estimate.
About 40 hours to about 4. A Metro Vancouver home builder now drafts proposals with Claude. Client’s own estimate. See the project
Runs at Sidekick
When it’s time to prospect, a ranked list of up to 25 businesses that fit, why each one fits and a suggested first message. A one-page brief before each call, and the customer database (CRM) updated after.
You decide: Who to contact and what to send.See it running at Sidekick
Runs at Sidekick
Finds post ideas in lessons from finished client work. No client is named or can be identified without their permission. Drafts a batch in your voice and lines it up for you to approve. Finds other people’s posts worth a reply and drafts the comment.
Now being set up for a client, an HR consulting firm.
You decide: What gets published.See it running at Sidekick
Runs at Sidekick
Checks each page of the website to see whether Google and AI tools like ChatGPT can find it and quote it as an answer. Lists what is missing, most important first, drafts the fixes and keeps the business name, address and hours the same on listing sites such as Google Business Profile and Clutch.
You decide: Anything that goes live.See it running at Sidekick
Runs at Sidekick
Builds a list of local firms that fit, researches each leader, drafts the letter, LinkedIn note and call script, and records each contact in the CRM.
You decide: Who to contact, and every send.See it running at Sidekick
Runs at Sidekick
A chat assistant answers visitors’ questions using only content we have approved. The AI Opportunity Map gives each visitor a short written report, and we get their answers with the lead.
You decide: What the assistant may say, and who follows up.See it running at Sidekick
Nathan Parkhouse Advisory
Before each call, an AI assistant (Claude) builds a brief from Nathan’s files, calendar and recent emails.
Preparing properly for a call went from about 45 minutes to about 5. His own estimate.
Whitney Shaw, executive coach
After full-day sessions, her notes become recaps and follow-ups drafted in her voice, ready for her to review.
“I hate sending follow-up emails.”
Whitney Shaw
Runs at Sidekick
When a client call ends, the recap, the follow-up email and the CRM note are drafted from the call notes and email, ready to review.
You decide: What gets sent and saved.See it running at Sidekick
Built for a Metro Vancouver home builder
Reads the request, pulls past proposals, pricing rules and statements the builder has already approved, drafts in the builder’s structure and checks every claim against the builder’s own records before anyone reads it.
About 40 hours per proposal down to about 4. Client’s own estimate.
You decide: Scope, price and final sign-off.
Built for a client
Standard proposals, large bid responses and marketing content now start as AI drafts that the team reviews.
Standard proposals take 75 to 80% less time. First drafts of large bids went from days to hours. Client’s own estimates.
Built for a client
Turns colours, fonts, logo and tone into templates, starts each new document from them and checks layout and brand before anyone sees the file.
Delivered as a brand kit and deck templates, so every new deck or proposal starts on-brand instead of from a blank page.
You decide: The brand rules, once. Then each finished file.
Whitney Shaw’s own project, after coaching
A 56-question team questionnaire that was always scored by hand now scores itself in a connected spreadsheet.
“That 56-question team effectiveness questionnaire that was always hand-tabulated is now a Google Form with automated scoring in a linked sheet.”
Whitney Shaw
Runs at Sidekick
Each week, everything waiting on the founder arrives as yes or no questions, plus a three-line update on this quarter’s goals.
You decide: Each yes or no.See it running at Sidekick
Aman Grewal, managing partner
Claude drafts his meeting prep, presentations, investment memos and financial models from his files. He reviews each one.
“It feels like you suddenly have an admin, an analyst, and a small team working behind the scenes.”
Aman Grewal
Kevin Chaput, Bloom Benefits
Sidekick’s founder, Chase, coached Kevin until he could build his own AI tools.
“I built an interactive dashboard with my company’s branding in two minutes. Commission structures, stages, all of it. Then I built an accounting folder that basically works like a bookkeeper.”
Kevin Chaput
Jennifer Principalli, G2 Virtual
AI prepares client spreadsheets, schedule changes and audits for two or three clients at once, ready for her to finish.
“I can load it with two or three clients’ work while I’m on someone else’s. It’s already prepping the work, and I jump in and take it from there. I almost have my own support system instead of needing someone else to support me.”
Jennifer Principalli
// Your firm
Start with the free AI Opportunity Map. It takes about 5 minutes, and you get a short written report on which of these fits your work first.
// Elsewhere
These are bigger than anything a firm your size needs on day one. They show where the ceiling is today. Each links to its source, and most figures are the organization’s own.
Nearly 40%more sales from service calls
Verizon. Gave its 28,000 customer service reps a Google AI assistant that finds answers and suggests offers during calls. Sales through that team rose nearly 40% after the rollout.
Reuters, April 9, 2025 Reported by: Verizon’s figure, before and after
220,000leads worked in one quarter
Salesforce. AI agents followed up on leads its own sales team never had time to call. In one quarter they worked 220,000 leads and added US$42 million of new sales pipeline.
Salesforce Q1 FY27 earnings call, May 27, 2026 Reported by: Salesforce, which sells the product
20%higher win rates
Microsoft. Gave its own sellers an AI assistant for call notes, research and follow-ups. In the first few months, individual win rates rose 20% and revenue per seller rose 9.4%.
Microsoft Inside Track, August 28, 2025 Reported by: Microsoft, which sells the product
16.3%the largest sales gain across seven tests
A large online retailer. Tested AI in seven parts of the business, from customer service to advertising, against its usual methods with real customers. Sales rose in most of them.
Columbia researchers, arXiv, June 29, 2026 Reported by: Independent researchers, controlled tests
1,794doctor working days saved in a year
Kaiser Permanente. 7,260 doctors used AI scribes that draft the visit note while they talk with the patient, across more than 2.5 million visits. 47% of patients said their doctor spent less time looking at the computer.
The Permanente Medical Group, April 7, 2025 Reported by: Its own analysis, published in NEJM Catalyst
76% vs 72%customer satisfaction, AI vs people
Octopus Energy. In a three-month trial, an AI email agent answered about 8,000 customer emails a week. Customers rated its replies higher than comparable replies from human advisors.
Octopus Energy press release, July 7, 2026 Reported by: Octopus Energy’s own trial
3 billionclient interactions since 2018
Bank of America. Its in-app assistant, Erica, has passed 3 billion client interactions and now handles about 58 million a month. The bank says more than 98% of users find what they need, which cuts calls to its service centres.
Bank of America newsroom, August 20, 2025 Reported by: Bank of America
2 minto resolve a request, down from 11
Klarna. Its AI assistant handled 2.3 million conversations, two-thirds of its customer service chats, in its first month. In 2025 its CEO said cutting customer service costs with AI had gone too far, and customers can now always reach a real person (Bloomberg, May 8, 2025).
Klarna press release, February 27, 2024 Reported by: Klarna, then Bloomberg on the change of course
Up to 117.9hours saved per lawyer per year
CMS (law firm). Gave more than 7,000 lawyers and staff in 50+ countries a legal AI tool. 93% of users reported productivity gains.
CMS newsroom, December 10, 2025 Reported by: Users’ own reports; 117.9 is the top of the range
67%more productive proposal writing
Fujitsu. Built an AI agent that drafts sales proposals for its sales teams from the company’s own knowledge.
Microsoft customer story, March 1, 2025 Reported by: Microsoft, which sells the platform
0.7%of Google’s worldwide computing recovered
Google DeepMind. Google’s AI coding agent found ways to run Google’s own data centres more efficiently. At Google’s size, 0.7% of its computing power is a very large amount of hardware.
Google DeepMind blog, May 14, 2025 Reported by: Google
£2for a payment reminder from a law firm
Garfield.Law. The first AI-only law firm approved by the solicitors’ regulator for England and Wales. It helps businesses recover unpaid debts, charging £2 for a reminder and £7.50 for a formal demand letter.
Legal Futures, May 6, 2025 Reported by: Trade press; approval confirmed by the regulator
40%better results, as rated by graders
Boston Consulting Group. 758 consultants did realistic consulting tasks with and without AI. With it they finished 12.2% more tasks, 25.1% faster, and graders rated the results about 40% better. The consultants who started weakest gained the most.
BCG, with Harvard Business School, September 21, 2023 Reported by: Controlled experiment
34%more issues solved by newer staff
A Fortune 500 support centre. About 5,000 support agents got an AI assistant. Issues resolved per hour rose 14% on average, and 34% for new and less-skilled agents.
Brynjolfsson, Li and Raymond, NBER, April 2025 Reported by: Measured on live support work; Quarterly Journal of Economics
1 personwith AI matched a two-person team without it
Procter & Gamble. 776 professionals worked on real product problems, alone or in pairs, with and without AI. One person working with AI matched the performance of a two-person team without it.
Harvard and Wharton researchers, NBER, April 2025 Reported by: Controlled experiment
98%of advisor teams use the firm’s AI assistant
Morgan Stanley. Gave its financial advisors a chatbot that answers questions from the firm’s own know-how, then a meeting assistant that, with client consent, takes notes and drafts the follow-up email for the advisor to edit and send.
Morgan Stanley press release, June 26, 2024 Reported by: Morgan Stanley
// What needs people
In every Sidekick project above, a person signs off at the points that matter.
How we work with a firm: leaders first, then their teams, with results measured against goals you set. See how a company rollout runs
Source for the 19%: Harvard Business School working paper 24-013, September 2023.
Go further
The AI Opportunity Map is free and takes about 5 minutes. You get a short written report on where to start. Your answers are scored by fixed rules Chase wrote and never go into an AI tool.
Have a question first? Ask it here, text Chase at 778-999-0985, email, or connect on LinkedIn.