On 30 January 2026, Anthropic expanded its Cowork tool with plug-in agent personas for sales, legal, finance, marketing and customer support. The move pushes Claude further into the domain-specific work that professional-services firms charge for.
The question it raises is not whether these personas can be useful. They can. The question is how much judgement they can safely replace.
What Cowork plug-ins promise
Cowork is Anthropic’s take on an AI workspace: a shared environment where teams can deploy agents that read documents, answer questions, draft content and perform tasks. The new plug-ins add pre-built personas tuned for specific functions.
A sales persona might research prospects, draft outreach and update a CRM. A legal persona might review contracts, summarise clauses and flag risks. A finance persona might extract numbers, reconcile statements and produce variance commentary. A marketing persona might generate campaign copy and briefs. A support persona might triage tickets and draft responses.
In each case, the value proposition is the same: a specialist assistant that understands the vocabulary, constraints and outputs of a particular discipline.
Where the promise is real
For mid-market firms, the appeal is strong. Specialist headcount is expensive. Junior staff spend much of their time on tasks that are repetitive but require enough domain context to be hard to automate with traditional software. Agent personas sit in that gap: they understand language, follow instructions and can work across unstructured documents.
Used well, they can accelerate first drafts, reduce turnaround time and free senior people for higher-value judgement. That is a genuine productivity gain.
Where the limits matter
The risk is that domain-specific styling gets mistaken for domain-specific judgement. A legal persona can write in the voice of a lawyer. It cannot hold a solicitor’s practising certificate. It does not owe a client a duty of care. It cannot be disciplined or sued for negligence.
This matters because professional-services work is defined by accountability. A clause-by-clause contract review is not just text processing; it is advice that carries risk. A marketing persona can generate copy, but it cannot reliably judge whether a claim is compliant with advertising regulation.
Firms that delegate too far will discover these limits through incidents. Firms that understand the limits can design useful boundaries.
How to use personas responsibly
Treat personas as assistants, not substitutes. Let them draft, summarise, research and check. Require human sign-off for anything that commits the firm or affects a client outcome.
Match the persona to the risk. A support persona answering frequently asked questions is lower risk than a legal persona interpreting a contract. Adjust oversight accordingly.
Keep the expert in the loop. The senior professional should review the agent’s work, not just the final output. Otherwise the firm loses the quality control that justifies its fees.
Document the scope. Be clear internally and, where relevant, externally about what the persona does and does not do. Ambiguity about AI involvement is a liability waiting to happen.
The bigger picture
Anthropic’s expansion is part of a broader trend: agentic AI packaged by function rather than by general capability. That makes the technology more accessible and more tempting to deploy. It also makes the governance challenge more urgent, because the gap between what the persona appears to do and what it can be trusted to do is not always visible to the user.
Professional-services firms that thrive will be the ones that use personas to amplify expertise, not replace it. The technology can handle the routine. The judgement still belongs to the professional.