Latch

• Downloading each month's bank and card statements, naming them to the firm's convention and filing them in the right Google Drive folders
• Reconciling a business checking account by matching the bank statement to ledger cash activity in Xero and writing the result to a Google Sheets workpaper
• Pulling actuals from Xero into an existing financial model in Google Sheets on a weekly schedule
• Turning a monthly order report and pricing sheet into grouped, priced QuickBooks invoices, with near-miss customer names sent back for review
• Reviewing bills, receipts and card charges for service-period cutoff and proposing accruals for approval
• Rolling forward the month-end close checklist and asking task owners in Slack for missing inputs
• Asking Claude or ChatGPT, through Latch's MCP server, to turn a recorded process into an SOP a new hire could follow

Latch takes a sensible approach to a real problem in small finance teams: the hard part of automating close work is rarely the clicking, it is the unwritten rules, and recording someone who explains those rules as they work captures more of them than a prompt or a flowchart would. Keeping each sidekick to fixed steps, a narrow set of permissions and a pause on anything unexpected also matches how accountants want automation to behave, and the MCP server means the same recordings can feed SOPs in Claude or ChatGPT. What a buyer cannot yet see is the evidence. Documentation stops at installation and MCP setup, there is no list of which apps run through APIs rather than the screen and no fixed price per step, and, most importantly for client work, Latch's own documents pull in different directions: the getting-started guide promises sessions never train AI while the privacy notice allows it, and the customer terms tell you not to send bank account numbers through a product whose flagship example is a bank reconciliation. The free plan makes it easy to try on an internal task. For client books, get the data questions answered in writing first.

About Latch

Latch replaces the prompt with a demonstration. In the desktop app you start a learning session, pick a screen and microphone, and do one full cycle of the task the way you normally would while explaining the decisions and exceptions out loud; Latch follows the screen, listens to the voiceover and asks clarifying questions at natural pauses. Afterwards it builds a sidekick from the session, lists every step and the reason for each permission it wants, and lets you test it and preview what it would change before the first real run. You authorize only the apps that sidekick needs, and each sidekick runs in its own sandbox on your machine. The company is emphatic that a sidekick is not a free-roaming AI agent: AI is used to learn the process and explain problems, but the sidekick runs the same steps in the same order within the same scope every time, and when it meets something it was not taught (an account in one system but not the other, more than one possible match, a multi-factor prompt at a bank portal) it pauses, explains the problem in plain language, suggests what to do and waits for you. Decisions can be remembered so the same question is not asked twice, and every run leaves a plain-language, step-by-step log. Latch's examples are month-end work: its use-case pages sketch how sidekicks could collect and file statements to a naming rule, run an eight-step bank reconciliation across a bank portal, Xero, Google Sheets and Google Drive, review AP cutoff and accruals, prepare payroll journals and update deferred revenue schedules, and a case study describes a few recorded invoices becoming a sidekick that builds about 150 QuickBooks invoices a month. Recordings also feed a Latch library that AI assistants can query: an MCP server included on every plan connects Claude, ChatGPT, Codex and GitHub Copilot, so you can ask for an SOP or a list of exceptions drawn from what you recorded. Pricing is credit-based: teaching and sidekick generation are free, credits are spent when sidekicks run (the run price depends on its number of steps, and failed runs are not billed), and plans go from a free single-user tier to pooled team plans with multi-recording synthesis, role-based access and audit log export. The tradeoffs come from the same design. Latch is new (it launched publicly on Product Hunt on November 10, 2026), its documentation is limited to an install guide and an MCP guide, and the site does not say which apps are driven through APIs and which through the screen. Recordings capture whatever is on the recorded screen and said aloud and are stored in Latch's cloud, and Latch's documents do not agree on how that content may be used: the getting-started guide says sessions are never used to train AI, while the privacy notice allows training Latch's and third-party providers' models. The customer terms also tell customers not to transmit financial account information such as bank account numbers, which sits awkwardly with bank reconciliation as a flagship example. Firms with client-data obligations should read the terms, privacy notice and DPA, and get the training question answered in writing, before recording live client work.