# Research

Sourced reference material I keep on hand and maintain properly - measurement ranges, bounds and citations, collected for work I actually do - and the frameworks built on top of it, for anyone who wants a model to work with.

## For accounting firms

- [Accounting firm software: what 21 vendors charge at your size, and where they say client data goes](/research/accounting-firm-software)

  A directory of practice management, client payment and portal, hosting and AI tools sold to accounting firms. Each entry was read on the vendor's own site on 2026-10-18, with public pricing or a note that there is none. Built from a link-graph scan of three vendors' inbound links.

- [Will AI replace bookkeepers, accountants and assistants? What the 2025-35 projections and the benchmarks say](/research/ai-and-bookkeeping-jobs)

  The U.S. Bureau of Labor Statistics projects bookkeeping clerks down 6 percent by 2035, accountants up 5 and executive assistants flat. Accounting benchmarks from 2025 and 2026 show agents far short of a full close. Read on 2026-11-05.

- [Dext and Hubdoc: what bookkeepers pay for receipt and invoice capture, where the documents are stored and what AI is inside](/research/dext-and-hubdoc)

  Dext charges bookkeeping firms from 22.07 Canadian dollars per client a month and stores data in AWS Ireland and Frankfurt, with no other region on offer. Hubdoc comes with Xero business plans. Read on each vendor's own pages, 2026-11-06.

## For anyone who wants to be found

- [How content gets found: a working model of distribution, and what it did for a $500k business](/research/content-distribution)

  Search catches demand that exists; social creates it. Each piece serves one stage of a buyer's readiness, volume matters with diminishing returns, and one clear topic per account wins. The model, the evidence for each part, what it did for a business stuck at $500k a year, and a diagnostic to run on your own accounts.


- [What Common Crawl covers and what it misses, and how far it stands in for Google](/research/common-crawl)

  Common Crawl is a free copy of part of the web that our scans use. Its FAQ calls it "a randomly selected subset", and a Mozilla study says it is not representative. What its web graph holds, what it leaves out, and what happened when we looked for our own site in it.

## For business owners

- [AI weekly reports and prompt templates for a small office: what an n8n weekly summary does, and what Anthropic and OpenAI say a prompt should contain](/research/ai-weekly-reports-and-prompt-templates)

  An n8n template turns Google Sheets rows into a weekly emailed summary with 10 minutes of setup and no accuracy figure. OpenAI is shutting down reusable prompt objects on 30 November 2026.

- [Is n8n used for accounting work? What its QuickBooks and Xero nodes can do, what the invoice templates claim, and what the licence means for a bookkeeper](/research/n8n-for-accounting)

  n8n's QuickBooks node covers ten resources, four with a Delete operation, and can be used as an AI tool; its Xero node covers contacts and invoices. A bill-processing template claims to run "without human intervention" and gives no accuracy figure.

- [Getting a PDF invoice into the books: what Excel, QuickBooks Online, Xero and Azure's invoice model each read, and what each leaves to you](/research/pdf-invoices-into-accounting-software)

  QuickBooks Online's emailed-bill route lists dates, amounts, vendor and card digits; Xero fills six header fields; Excel's connector reads tables only; Azure's model also extracts line items.

- [Does OpenAI train on your data, and who answers for what a chatbot says? OpenAI's own data terms and the Air Canada tribunal decision](/research/ai-privacy-and-chatbot-liability)

  OpenAI says API data has not been used for training since 1 March 2023, but keeps abuse logs up to 30 days; individual ChatGPT plans may be trained on. A BC tribunal ordered Air Canada to pay $812.02 for its chatbot's wrong advice.

- [How to automate follow-up emails with AI: what Gmail, HubSpot, a cold-email vendor and Canada's anti-spam regulator say](/research/ai-follow-up-emails)

  Gmail's Gemini drafts a follow-up when you ask and you send it; HubSpot sequences send on a timer and stop when the contact replies, on Professional plans and up. One vendor's data puts 42 percent of cold-email replies after the first email. CASL sets consent rules for each message.

- [Can AI do data entry? What a vendor test, a trial-report study and OpenAI's own pages show, set against human typing error rates](/research/ai-data-entry)

  People key data with about 0.29 errors in 100 fields; in an October 2026 vendor test, ChatGPT and Gemini silently returned nothing on a 20-page document. BLS projects data entry keyers down 26 percent by 2035.

- [How long does it take to build an AI agent? What four agencies quote, what one set of 47 enterprise rollouts measured, and what the vendors' quickstarts claim](/research/how-long-to-build-an-ai-agent)

  Agencies that sell agent builds quote days to 3 weeks for a simple agent and 3 to 6 months for a large one. A 47-rollout enterprise sample puts the median at 248 days from contract to production. No source measures a small-business build.

- [Do AI agents actually work? What four benchmarks and one analyst forecast measured between 2024 and 2026](/research/do-ai-agents-work)

  Benchmarks from 2024 to 2026 show agents finishing some tasks and few of them every time. Gartner expects over 40 percent of agentic AI projects to be canceled by the end of 2027. Read on 2026-11-06.

- [AI agent or chatbot? Where IBM and Anthropic draw the line, and why one product can be both](/research/ai-agent-vs-chatbot)

  IBM says assistants and chatbots react while agents plan and act toward a goal. Anthropic says agents choose their own steps where workflows follow fixed code. Gartner says many products are relabelled chatbots. Read on 2026-11-06.

- [AI pilot purgatory: where the 88, 95 and 42 percent figures come from, and what each one counts](/research/ai-pilot-purgatory)

  The figures for AI pilots that never reach production come from four different measures: IDC's 88 percent, MIT's 95, S&P Global's 42 and Gartner's forecast of 30. Each counts something else. Read on 2026-11-06.

- [Automation opportunity assessment: what three published frameworks ask, how they score, and where they differ](/research/automation-opportunity-assessment)

  Three published frameworks for choosing which process to automate share volume, rules, exceptions and ownership, but one scores readiness only while another adds value, safety gates and weights. None cites evidence for its thresholds. Read on 2026-11-06.

## For people building with AI

- [Which AI agent harnesses lead right now? Claude Code, Codex, Cursor, Antigravity, Hermes Agent, pi and OpenCode, ranked by one day of OpenRouter tokens and by GitHub stars, and what each page says it is](/research/top-ai-agent-harnesses-survey)

  On one day's OpenRouter token count, Hermes Agent (1.73 trillion) and Claude Code (1.55 trillion) lead. By GitHub stars, Hermes Agent (261k) and OpenCode (225k) lead. OpenCode and Antigravity are missing from the token list, so neither ranking covers all seven.

- [How do the top AI agent harnesses handle heartbeats? Which of Claude Code, Codex, Cursor, Hermes Agent, OpenCode, pi and Antigravity document a way to wake an agent on a timer with no new message from a person, and how each one does it](/research/how-top-ai-agent-harnesses-handle-heartbeats)

  Only Hermes Agent documents a feature called a heartbeat: one recurring prompt per session that fires when the session is idle. Claude Code, Codex, Cursor and Antigravity document timers or schedules under other names; the docs read for OpenCode and pi describe none.

- [How do the top AI agent harnesses handle hooks? What Claude Code, Codex, Cursor, Hermes Agent, OpenCode, pi and Antigravity document for the points a hook can attach to, whether it can block a call, and what happens when a hook fails](/research/how-top-ai-agent-harnesses-handle-hooks)

  Six of the seven document a hook that runs before a tool call and can block it. In Claude Code, Codex and Cursor a hook that crashes or times out lets the action through unless it is set to fail closed.

- [How do the top AI agent harnesses handle tool calls? What Claude Code, Codex, Cursor, Hermes Agent, OpenCode, pi and Antigravity document for the tools they ship, who approves a call, and what limits it](/research/how-top-ai-agent-harnesses-handle-tool-calls)

  The seven ship the same core tools and differ on who approves a call. OpenCode and pi run tools without asking by default, Antigravity's default asks only outside a sandbox, Codex separates sandbox from approvals, and Claude Code layers allow, ask and deny rules.

- [How do the top AI agent harnesses handle memory? What Claude Code, Codex, Cursor, Hermes Agent, OpenCode, pi and Antigravity document for instruction files, notes the agent writes for itself and search of past sessions](/research/how-top-ai-agent-harnesses-handle-memory)

  All seven read a hand-written instruction file such as AGENTS.md or CLAUDE.md. Only Claude Code, Codex and Hermes Agent document notes the agent writes for itself: a 200-line index, background-built files and a 2,200-character file.

- [Where can you learn to build AI agents for free? The Hugging Face course, UiPath Academy, Google's Kaggle course and OpenAI's PDF guide, and what a search for an agents book PDF turns up](/research/learn-ai-agents-free)

  Hugging Face's course assumes Python and teaches three libraries; UiPath's is for business users and shows no price; Google's is a self-paced five-day program; OpenAI's is a 34-page PDF. Manning sells the agents book and links free content.

- [How do you stop an AI agent? What n8n and LangGraph give you for human approval, what LangChain, CrewAI and the OpenAI SDK give you for loop limits, and what I found on kill switches](/research/stopping-an-ai-agent-approval-and-limits)

  n8n and LangGraph can pause an agent until a person approves a tool call; LangChain's loop cap defaults to 15 steps and CrewAI's to 20. A kill switch came up only on vendor blogs.

- [How to test an AI agent, and whether retrieval cuts hallucinations: Anthropic's eval guide, a study of legal AI tools, and Vectara's leaderboard](/research/testing-ai-agents-and-hallucinations)

  Anthropic separates a task passed once from one passed every time: 75 percent per trial is 42 percent over three. Retrieval cut hallucinations in legal AI tools but left 17 to 33 percent.

- [Connecting an AI agent to QuickBooks, HubSpot and Google Workspace: which connectors are official, what they can read and write, and where each is limited](/research/ai-agent-connectors-quickbooks-hubspot-google)

  Intuit's official QuickBooks connector for Claude is US-only and can delete invoices. HubSpot's connectors mark Delete for no object. Google's Workspace MCP servers are a Developer Preview.

- [Which AI agent builders are free? A survey of eight: the licence, the free tier, and what each leaves out](/research/free-ai-agent-builders)

  Eight builders have a free route. Flowise's repository was archived on 13 August 2026. Dify's Community edition is described as non-commercial while its licence allows commercial use with conditions. Langflow is MIT. Lindy gives a 7-day trial, and OpenAI's drag-and-drop Agent Builder shuts down on 30 November 2026.

- [Is agentic AI paying off or just hype? What Gartner, McKinsey and a vendor-run survey each measured about return](/research/agentic-ai-roi-and-hype)

  A vendor-run survey says 96 percent of organizations with agentic AI in production met or beat their ROI expectations. McKinsey finds 37 percent of respondents attribute any profit effect to AI, and Gartner forecasts over 40 percent of agentic projects canceled by 2027.

- [Are Zapier and Make AI agents? What their help pages say an agent is, how each one counts the bill, and what Zapier is retiring](/research/zapier-and-make-ai-agents)

  Zapier is migrating its standalone Agents product into an AI step inside the Zap editor, billed at 1x, 3x or 5x tasks by model tier. Make's agent app is in open beta and bills a credit per operation plus tokens. Neither page reports how often its agent finishes a job correctly.

- [Can you stop an AI agent being tricked? What OWASP, OpenAI, Meta and the attack papers say guardrails catch and miss](/research/ai-agent-security-and-guardrails)

  OpenAI says prompt injection is "unlikely to ever be fully solved". In one study, adaptive attacks got past 12 published defences, most at over 90 percent. OWASP and Meta advise limiting what the agent can do; none of the sources measures whether that works.

- [CrewAI: what Crews and Flows are, the three ways to put a human check in, what the tutorial needs, and what the licence and platform say](/research/crewai-human-input-and-pricing)

  CrewAI is MIT-licensed Python with three documented ways to pause for a person, one of them labelled Enterprise. Its tutorial needs an LLM key and a Serper key, and its pricing page shows a free tier of 50 executions a month.

- [How to watch an AI agent and keep an audit trail: what OpenTelemetry records, and the licence, price and retention of four tools](/research/ai-agent-monitoring-audit-trails)

  OpenTelemetry's agent conventions are still marked Development and keep message content opt-in. Langfuse, Phoenix, Helicone and LangSmith keep traces from 14 days to 3 years.

- [n8n AI agents: what the AI Agent node is, what three of its templates need, what a run costs and what the licence allows](/research/n8n-ai-agents)

  n8n's AI Agent node needs a chat model and at least one tool, and its AI category holds 9,723 community templates. n8n counts a run as one execution however many steps it has, and says it is not open source.

- [Vibe manufacturing: where the money goes when AI designs physical things](/research/vibe-manufacturing)

  AI makes drafting a physical product cheap. Checking it is still scarce, making small parts is cheap but priced all over the map, and the last time a hardware cost collapsed, the losses landed at making and delivery. The evidence for each, the routes ranked by where the losses landed, and a check to run on your own idea. Every figure attributed.

- [OpenAI Agents SDK: how to start, how it connects to MCP servers, and what its default tracing sends to OpenAI](/research/openai-agents-sdk)

  What OpenAI's own docs say about the Agents SDK, read on 2026-11-06: the four steps to a first agent, the four ways to connect an MCP server, and that tracing is on by default and includes model inputs and outputs.

- [AI document processing for invoices and receipts: what Amazon, Microsoft and Google charge per page, where your documents go, and what one study found](/research/ai-document-processing-invoices-receipts)

  Amazon Textract and Azure charge $0.01 a page for invoices and receipts; Google charges $0.10 for up to 10 pages, so a one-page receipt costs ten times as much. Read on each vendor's own pages, 2026-11-06.

- [What an AI automation engineer does and earns: the job has no official category, so here are the closest official numbers and one job description](/research/ai-automation-engineer-jobs)

  The U.S. job handbook has no AI automation engineer entry. Its nearest one, software developers, has a median wage of $135,980 and 10 percent projected growth to 2035. Read on 2026-11-06.

- [How AI agents fail: the 14 failure modes found in 1,642 traces of multi-agent systems, and the security failure OWASP names](/research/ai-agent-failure-modes)

  A study of 1,642 runs of multi-agent systems found 14 ways they fail, in three groups: design, agents not coordinating, and weak checking of the result. OWASP names a separate failure: an agent allowed to do too much. Read on 2026-11-06.

- [No-code AI agent and automation builders: what n8n, Zapier and Make charge, and what each one counts](/research/no-code-ai-agent-builders)

  n8n charges per workflow run, Zapier per step and Make per action, so the same workflow costs different amounts. Entry plans read on the vendors' own pages on 2026-11-06: n8n 20 euros, Zapier 19.99 dollars, Make free to 1,000 credits.

- [Building an AI agent with LangChain: what its docs say an agent is, how its four products divide the work, and what LangSmith costs](/research/langchain-agents)

  LangChain's docs define an agent as a model calling tools in a loop and offer create_agent, LangGraph and Deep Agents. The library is MIT-licensed; its tracing service, LangSmith, is free for one seat to 5,000 traces a month, then 39 dollars a seat. Read on 2026-11-06.

- [Microsoft Copilot Studio: what Microsoft charges for building agents, and how a Copilot Credit is counted](/research/copilot-studio)

  Copilot Studio bills in Copilot Credits at 1 US cent each on pay-as-you-go. A generative answer uses 2 credits, an agent action 5 and a grounded answer 10. Microsoft lists quotas and limits for agents, including a five-minute response timeout. Read on Microsoft's own pages on 2026-11-06.

## For character artists

- [Facial parameters for character work: the dials that carry a face, and how far they travel](/research/facial-character-parameters)

  Which measurements make a face read male, female, older or from another population, ranked by how reliably they separate faces, with measured ranges by sex, a slider mapping that spends its codes where faces differ, and the data as JSON. Every figure attributed.

- [Cartilage for character work: the dials bone and fat don't give you](/research/cartilaginous-tissues)

  The lower nose and the whole ear except the lobe are cartilage. It's thinner than it looks, where it sits decides how a nostril reads, the ear stands off the head, and it ages on its own clock. The measured range for each, as dials for a character creator. Every figure attributed.

## For designers

- [How should body text be set on a screen? What a typography book recommends for size, line length and line spacing, and what a study of long lines on screen found](/research/setting-body-text-for-the-screen)

  Butterick's Practical Typography recommends 45 to 90 characters a line, 15 to 25 pixels of body text and line spacing of 120 to 145 percent. A 1998 study found long lines were read faster with no change in comprehension; a review by its author found them least preferred.

---

Published 2026-10-04. Web page: https://www.mackenziebowes.com/research
