How to set up Claude Cowork in your accounting firm

Best answer. An agent can only touch what it can reach. One dedicated folder limits any mistake to the files you chose to put there.

Reasonable, but not the best. Manual approval, where it pauses for a yes before each action, helps, but every client's files are still within reach. One working folder plus manual approval is safer.

Not the best move. Being careful does not narrow what the agent can reach. One wrong instruction could touch any client's files.

Best answer. Pausing before each action lets people see what it is about to do and stop it, while everyone is still learning how it behaves.

Not the best move. Seniority does not change what the agent might get wrong. Everyone starts on manual approval until the firm has months of supervised use.

Not the best move. Reviewing output catches a bad draft, but not an action already taken, such as a file changed or deleted. Manual approval stops it beforehand.

Not the best move. The provider's terms cover how it treats your data, not whether your use meets your own obligations to clients.

Best answer. The Privacy Act sets how you must handle personal information, and your engagement letters set what clients have agreed to. Your setup has to fit both.

Reasonable, but not the best. Team buy-in is worth having, but it does not tell you whether the setup meets the law or what you have promised clients.

Not the best move. Limiting it to one person does not reduce what the agent could reach or change through that connection.

Not the best move. These are the systems with the most client data and the power to change records, so they are the last to connect, not the first.

Best answer. Adding connectors one at a time, starting low-risk, lets you see how each behaves. Systems with client records, ledger access or the power to send messages come much later.

Not the best move. The account decides what the connector can reach, so choosing by convenience can hand over far more access than you meant to.

Not the best move. If the account can see everything, so can the connector. That is far more than a calendar connection needs.

Best answer. Because the connector inherits that account's permissions, a restricted account limits what Cowork can see and change.

Reasonable, but not the best. Her view is useful, but estimates tend to run high. Timing each task gives you numbers you can rely on when deciding what to roll out.

Best answer. Timed figures task by task show which jobs really worked and which needed so much fixing they saved nothing. That is your evidence for rolling it out.

Not the best move. Written up from memory at the end, the details are gone and the log turns back into guesses.

Not the best move. A couple of weeks of good summaries is not enough history. The firm stays on manual approval until it has months of supervised use.

Best answer. A skill captures what worked, so every session follows the same format, and it becomes what you train the rest of the team on.

Not the best move. Rolling out before the trial ends and a policy exists skips the safeguards. The skill you write now is what makes a later rollout consistent.

Reasonable, but not the best. It does no harm, but a policy written now rests on guesses. The data rules you already wrote down keep the trial safe in the meantime.

Best answer. The trial shows which data, connections and jobs actually come up, so the policy covers real situations. The early safeguards are already in writing.

Not the best move. The wider team should be trained under a policy that already exists, so the policy has to come first.

Best answer. AI drafts can look polished and still contain a wrong figure or name. Only a full read by a person catches that.

Not the best move. Spot-checking lets unread letters reach clients, and any one of them could be wrong.

Not the best move. Earlier approvals say nothing about this letter's figures and names. Each piece needs its own read.

Best answer. Clients should know AI is used on their work and that a person reviews it. The engagement letter is where that is agreed.

Reasonable, but not the best. Publishing it does no harm, but most clients will never read it. The engagement letter is where clients agree to how their work is done.

Not the best move. The policy is internal, but clients still need to be told AI is used on their work, and the engagement letter is where that happens.

Best answer. Proven jobs and written skills give her results that work from day one, with output her reviewer already knows how to check.

Reasonable, but not the best. Holding her back is safe but wastes her enthusiasm. People pick it up at different speeds, and keen ones can start on proven jobs now.

Not the best move. Free experimenting throws away what the trial learned and produces work nobody has a way to check.

Not the best move. Seniority does not change what the agent might get wrong. The whole firm stays on manual approval for now.

Not the best move. Reviewing output catches a bad draft, not an action already taken, such as a changed or deleted file. One does not replace the other.

Best answer. Three weeks is not enough supervised use to know how it behaves across your work. Manual approval stops a bad action before it happens.

Reasonable, but not the best. Worth hearing, but liking it does not tell you whether it is used on the right jobs or whether the output is being reviewed.

Best answer. That shows whether it is being used where it works, and whether every piece of output is still read by a person before it goes out.

Not the best move. A total says nothing about which jobs it is used for or whether its work is being checked.