Signalcroft AI

Ask Signalcroft

Anything about putting Claude to work in a firm: what a release means, what something costs, where the data goes, where to start. A person on the team answers in writing, usually within a working day.

Every answer is dated and sourced, published here for everyone, and the best of the week goes into the Monday brief. We do not name customers, quote their figures or claim results, so a question that needs one of those is answered without it.

Questions are read by a person and answered in public without your name. An email address is used only to tell you the answer is up, then nothing. Privacy Notice.

Answered

Is it safe to put client data into Claude?

Asked: Is it safe to put client data into Claude?

It depends which plan your staff are using, and most firms asking have not checked.

Anthropic's Privacy Center, in an article dated 18 August 2026, says that by default it does not use inputs or outputs from its commercial products, Claude for Work and the API among them, to train its models. Free, Pro and Max accounts are consumer products under separate terms. So the first step is an inventory: who in the firm uses Claude, and on which kind of account.

Inside a commercial plan there is one exception. A thumbs up or thumbs down stores the whole related conversation for up to five years, and Anthropic may use it to train its models. An Owner of a Team or Enterprise plan can switch that off under Organization settings, Data and Privacy, Rate chats. The same article says feedback does not include raw content from connectors unless it was copied into the conversation.

One date matters. Anthropic says Claude Data, which works through a data connector such as a warehouse or CRM, is on by default for Team plans and, from 20 November 2026, for Enterprise plans unless an admin changes it. The setting sits under Organization settings, Artifacts. Make it a decision rather than a default.

Training is only one of the questions. Whether a firm may share a client's data with any supplier turns on its engagement terms and UK data protection law, and that answer belongs to the firm, not to Anthropic.

Answered by the Signalcroft AI team (grace), 17 November 2026. Link to this answer

Our scheduled AI jobs run unattended. How do we know they finished rather than waited?

Asked: Our scheduled AI jobs run unattended. How do we know they finished rather than waited?

Count, do not assume. When we went back through nine days of our own unattended runs, 32 of 152 had blocked waiting for a person to click a permission prompt, seven were killed while still blocked, and five produced no output at all. None of them reported an error. The scheduler said each run started on time, which was true and told us nothing.

The pattern is that automated work rarely fails loudly. It fails by waiting, and waiting produces the same silence as success until somebody counts. So the two questions to put to anything you have automated are not "did it run" but "what proportion of runs finished, and where would I go to find out". If you cannot answer the second, you do not know the first.

Practically: keep three records and read them against each other. The scheduler knows a job fired. The session transcript knows what it did. The application log knows why it stopped. Only the third holds the answer to "was it waiting", and in our case none of the three had ever been compared until we built the thing that compares them. Then make the check a rule: our own suite now fails if any unattended run in the last two days stopped on a permission prompt, and names the run and the tool it wanted.

The fix for the waiting itself is usually a permission that should have been granted in advance, which is a configuration decision, not a model problem. A job that runs unattended by definition should never be able to ask a question.

Sources: our own estate, measured 16 to 25 August 2026, written up in the Insights piece "Our scheduled jobs were not failing. They were waiting" on signalcroft.ai/insights and in the Intelligence Platform's guide on running agents unattended.

Answered by the Signalcroft AI team (jack), 25 September 2026. Link to this answer

Can a firm with data-residency obligations use Claude Managed Agents?

Asked: Can a firm with data-residency obligations use Claude Managed Agents?

Read the retention position before the code. Claude Managed Agents is Anthropic's hosted harness for running Claude as an autonomous agent: sandboxes, sessions, credentials in a vault, scheduled runs. Its documentation states that it is in beta, that sessions store conversation history, sandbox state and outputs on Anthropic's side, and that it is not currently eligible for zero data retention or a HIPAA Business Associate Agreement. You retain the ability to delete sessions and uploaded files through the API.

So the answer depends on what your obligation actually says. If it requires that customer data is not retained by a processor beyond the request, Managed Agents does not meet it today and the Messages API with zero data retention does. If it requires processing in a named region, check the current documentation for the region options on the day, because that has been changing through 2026. If it requires that a named processor holds the data under a data processing agreement with defined retention, Anthropic's commercial terms and DPA are the documents to read, and Managed Agents sits inside them like any other API feature.

What we would do first in your position: write down which of the three obligations you are under, then run one read-only agent on data that carries no obligation (a scheduled triage of an error feed is the pattern Anthropic's own quickstarts use), so the firm learns the operating model before the data question is live.

Sources, read 22 September 2026: platform.claude.com/docs/en/managed-agents/overview (beta status, retention and BAA eligibility); the Managed Agents quickstarts on github.com/anthropics/claude-quickstarts; the Intelligence Platform's Security and Compliance section for the DPA and region routes.

Answered by the Signalcroft AI team (nadia), 25 September 2026. Link to this answer

Where should a firm start with the redesigned Claude Projects?

Asked: Where should a firm start with the redesigned Claude Projects?

Not with Projects, unless you are already running more than one Claude Code session on the same piece of work. The redesign of 17 September 2026 turns a Project from a folder of files and instructions into a conversation with a coordinator: you state a goal, Claude splits it into threads, each a full Claude Code cloud session on its own branch, and every thread lands as a pull request a person reviews. That is a product for a problem you only have once you are stitching sessions together by hand.

If you do have that problem and your plan has the beta (selected Pro and Max subscribers using cloud sessions in Claude Code, widening from there; Team and Enterprise later), start with one goal that has a testable finish and one connected repository. Anthropic's own first example is a latency target on a checkout, with a thread per endpoint. Plan your review capacity before the thread count: the developers who ran parallel sessions by hand for a year agree that validation, not orchestration, is the constraint, and a coordinator does not change how many pull requests a person can read in a day.

Three things to know before you commit to it. It is a beta. It runs in the cloud only, so code that may not leave your network waits for the local option Anthropic says is coming. And usage limits arrive faster, because every thread is a whole session.

Sources, read 22 September 2026: claude.com/blog/projects-redesigned; support.claude.com, "What are Projects"; devops.com, 18 September 2026 (The Futurum Group); our post of 23 September and the entry in the Intelligence Platform's releases record.

Answered by the Signalcroft AI team (jack), 25 September 2026. Link to this answer