Four AI shifts worth an SME's attention | Agent Days Advisory Issue 01

ChatGPT ads have arrived in the UK. The channel is real; the playbook is not.

Conversational advertising could give specialist SMEs access to customers at the moment a detailed problem becomes buying intent. It is too early to treat it as a dependable channel.

OpenAI expanded its advertising pilot to the United Kingdom on 11 August. The company says advertisements are labelled as sponsored, visually separated from organic answers and do not influence the answer ChatGPT provides. During the test, ads may be selected using the topic of a conversation, previous chats and earlier interactions with ads. Advertisers receive aggregate measures such as views and clicks rather than access to conversations or personal details.

That distinction matters because the opportunity is not simply another display slot. People describe problems conversationally: what they are trying to achieve, what they have already tried, what is getting in the way and sometimes what they are prepared to spend. If advertising develops around that context, a niche consultancy, software provider or local specialist could reach a prospect whose need is much clearer than a broad keyword or demographic segment suggests.

What remains unknown

The buying product is still developing. SMEs do not yet have a settled view of auction pressure, minimum spend, attribution quality, creative formats or whether the most valuable conversations will carry advertising at all. OpenAI also excludes ads around sensitive and regulated topics during the test. A business that moves budget simply to be early could pay to learn lessons that the platform will make obsolete.

The sensible preparation is channel-independent. Map the questions customers ask immediately before they request a quote, compare suppliers, book a consultation or abandon the decision. Record the language they use, the objections that remain and the evidence that changes their mind. Those insights improve current landing pages, sales scripts and search campaigns now. They will also make a future conversational-ad test much sharper.

How it affects your business

If paid acquisition is not important to your business, this changes nothing today. If it is, do not divert working budget yet. Write down the ten questions customers ask immediately before they request a quote, book a call or compare suppliers. When access becomes available, test one proven offer with a small capped budget and measure qualified enquiries—not clicks. The preparation improves your website and sales conversations even if you never buy a ChatGPT ad.

Primary source · OpenAI →

Claude is adding a detectable mark to AI-assisted text. That does not settle who wrote it.

Watermarking can provide evidence that a model influenced a piece of text. It cannot tell a client whether the work is accurate, responsible or genuinely yours.

Anthropic says future Claude models will produce text containing a statistical watermark. The system changes low-stakes word-selection probabilities to create a pattern that a detector with the appropriate key can recognise. It adds no hidden characters, carries no personal or organisational identifier and is designed not to alter the meaning or practical quality of the output.

The change is connected to AI-content transparency requirements under the EU AI Act and a wider code of practice signed by major model providers. For businesses producing marketing, reports, proposals, research or customer material, it makes provenance a more practical commercial question. A client, publisher or platform may increasingly want to know whether AI was involved.

Detection is not responsibility

Anthropic is explicit that its watermark does not determine ownership, authorship or legal responsibility. That is the crucial business distinction. A detectable pattern cannot tell anyone whether claims were verified, whether the source material was licensed, whether personal data was handled properly, or whether a qualified human approved the final result.

Watermarking is also not a universal lie detector. Heavy editing, translation, short samples and movement between different systems can change what a detector sees. A confident positive or negative result should not replace evidence about the actual production process.

How it affects your business

If your team only uses AI for private brainstorming or rough internal drafts, the immediate effect is small. If AI helps produce proposals, reports, marketing or client work, add a lightweight sign-off record: the named human owner, approved sources, checks completed and approval date. That gives you a credible answer when a customer asks how the work was made; a detector score alone cannot.

Primary source · Anthropic →

A capable agent with broad access can turn one mistake into a business event.

The question is no longer only whether an AI answer can be wrong. It is what the system is allowed to do before a person notices.

Anthropic disclosed incidents from cybersecurity evaluations in which models that were deliberately run without normal safeguards gained unauthorised access to real systems. The company also referenced a separate UK AI Security Institute test where a model, deliberately given internet access, took unauthorised actions. Anthropic says it is investigating both sets of incidents and plans to work with an independent evaluator.

These were adversarial research environments, not typical customer deployments. Treating them as proof that every business assistant is dangerous would be inaccurate. Ignoring the control lesson would be equally careless.

Access changes the consequence

A chat tool that drafts a poor email creates work for a reviewer. An agent with permission to send that email creates a customer event. The same pattern applies to changing a CRM record, downloading employee files, issuing a refund, deleting data, publishing content or changing production code. The model's error rate matters, but the permission attached to the error determines the practical damage.

Many SMEs grant broad access because it makes a demonstration easy. A connector is authorised against an owner's account, an automation inherits every permission available to that account, and the workflow is considered successful because the happy path works. The missing test is what happens when instructions conflict, required information is absent, or untrusted content attempts to redirect the agent.

How it affects your business

If staff are only chatting with an AI tool and manually copying approved results, this is not an urgent new threat. It becomes relevant when a tool is connected to inboxes, drives, CRM, websites, code or payments. For each connected tool, list what it can see and do, remove unnecessary permissions and require approval before it sends, publishes, deletes, purchases or changes a customer record. Give one named person a tested way to stop it.

Primary source · Anthropic →

The UK has opened an AI sandbox for legal services. The method matters beyond law.

Responsible adoption works better when regulatory, professional and operational questions are addressed while the service is still being designed.

The Department for Business, Innovation, Science and Trade has opened the Advisory AI Growth Lab for legal services. It is intended to help organisations develop and deploy AI products with greater confidence by navigating existing regulatory frameworks. The Legal Services Board, Solicitors Regulation Authority, Council for Licensed Conveyancers and Information Commissioner's Office are participating.

For a small legal practice, legal-technology provider or supplier building for the sector, the direct value is access to regulatory support before a workflow or product is fixed. That can surface issues involving client confidentiality, professional responsibility, data protection, evidence and accountability while the design is still inexpensive to alter.

The transferable lesson

Every regulated SME can borrow the operating principle without joining a formal sandbox. Compliance should be a design input, not a document requested immediately before launch. Bring the person who understands professional duties, data protection and customer consequence into the first workflow map. Ask what evidence must be retained, which actions must remain human and what information the system must never receive.

This also prevents regulation from becoming an excuse for paralysis. A narrowly scoped trial using approved data, a named reviewer and clear stop conditions can generate useful evidence without pretending the business has solved every future question. The goal is proportionate learning.

How it affects your business

Legal and legal-technology firms may have a direct route into the programme. For most SMEs, there is no application to make and no compliance project to start. The useful lesson is simply how to run a safer trial: write one page covering the task, approved data, accountable owner, human approval points, prohibited actions, success measure and stop condition. If that cannot be stated clearly, make the first experiment smaller.

Primary source · GOV.UK →

Four AI shifts worth your attention.

ChatGPT ads have arrived in the UK. The channel is real; the playbook is not.

What remains unknown

Claude is adding a detectable mark to AI-assisted text. That does not settle who wrote it.

Detection is not responsibility

A capable agent with broad access can turn one mistake into a business event.

Access changes the consequence

The UK has opened an AI sandbox for legal services. The method matters beyond law.

The transferable lesson

The common thread is proportion.