Why AI Training Doesn't Fail, But Follow-Through Does

Ask a team if they're confident using AI, and most will say yes. Ask a specific question about it, when to use one tool over another, whether they've checked a setting recently, and that confidence often doesn't hold. That gap between using AI a lot and using it well is the most common problem we see in businesses after AI training, and it's rarely anyone's fault. The tools move faster than most people's understanding of them.

September was a clear example. In the space of a few weeks, OpenAI released a model built to hand back finished work rather than a chat reply, Anthropic switched the default Claude model overnight with no announcement, and Microsoft added both of those models inside Copilot alongside a push towards more autonomous, agent-run tasks. None of that needed anyone's permission. It just changed underneath whatever a team learned at their last session.

What actually changed this month

OpenAI's GPT-6 Astra is built for delivered work, not conversation. It's designed to produce a finished, on-brand piece of work rather than a back-and-forth chat reply, and it works inside a business's existing software and access controls. It's also the first model to reach the "Critical" threshold on OpenAI's own cybersecurity capability framework (OpenAI), and OpenAI has flagged that it uses up a workspace's usage allowance faster than earlier models.

Claude's default model changed without warning. Anthropic made Opus 5.5 the standard model across Claude chat, Cowork and Claude Code, priced around 20% lower than Opus 5 with roughly 40% lower cost per task on typical workloads (Anthropic). Sonnet 5.5 and Haiku 5.5 are expected to follow in the coming weeks.

Microsoft 365 Copilot is now multi-model, and moving towards more autonomy. Claude Opus 5.5 and GPT-6 Sol are now selectable models inside Word, Excel, PowerPoint, Copilot Chat and Cowork, alongside Microsoft's "Work IQ" layer that grounds responses in an organisation's own files and data (Microsoft). Microsoft has also set out where Copilot is heading next: more agent-run tasks handled with less checking at each step, billed through usage-based Copilot Credits rather than a flat licence fee (Microsoft).

More autonomy is genuinely useful. It also means more decisions get made without a human checking each one, which is a governance question before it's a rollout question. If nobody in your business has looked at the admin controls since Copilot was set up, that's worth ten minutes before autonomy increases, not after.

Why this matters more than the average AI news roundup

None of these changes required a business to do anything. That's exactly the problem. A team can leave a training session confident and capable, and still fall behind within a few months, not because they stopped learning, but because the tools kept moving and nobody's job was to notice.

At the same time, how a business gets found is changing too. Three separate reports this month say the same thing: one business lost 140 million website visits in a year to AI search, click-throughs fall 60 to 70% on average when Google shows an AI Overview instead of a standard result, and just over half of B2B buyers now say they start research inside an AI chatbot rather than a search engine (G2, 2026 buyer survey, reported via Search Engine Land). If someone asks ChatGPT or Gemini who the best provider of a service is in their area, and a business doesn't come up, that's a lost customer before a website was ever visited.

What actually works

Training gets a business to a starting point. Keeping pace with it afterwards is a different job, and it's usually nobody's job by default. That's the reasoning behind fortysix ai's retainer service: ongoing, embedded AI support from £995 a month, delivered through virtual check-ins and on-site visits, essentially a fractional AI lead whose role is making sure a team's AI use keeps earning its keep as the tools underneath it change. It sits alongside our usual bespoke programmes, which remain the right starting point for a business beginning AI adoption or bringing a new team up to speed.

Frequently asked questions

• What is a fractional AI lead?

  Description text goes hereA fractional AI lead is an outside specialist who takes ongoing responsibility for a business's AI adoption, on a part-time or retainer basis, rather than a single training engagement. The role covers checking settings and permissions, keeping teams current as tools change, and being accountable for AI use staying effective over time, without the cost of a full-time internal hire.

• How often should a business retrain staff on AI tools?

  There's no fixed interval, because the tools themselves don't update on a fixed schedule. Major model and product changes happened at least three times in September 2026 alone. Rather than scheduling retraining around a calendar, the more reliable approach is ongoing, lighter-touch checks, whether through an internal owner or an external retainer, so gaps get caught as they appear rather than discovered months later.

• What's the difference between AI training and an AI retainer?

  Training is a defined session that builds capability at a point in time. A retainer is ongoing support that maintains that capability as the tools change, through regular check-ins, settings reviews and troubleshooting. Most businesses need both: training to start, and something to stop that training going stale.

• Is my team using AI well, or just using it a lot?

  Frequent use and effective use aren't the same thing. A useful test: ask a few regular AI users specific questions, for example when they'd choose one tool over another, whether they've used a Skill or a connector, or what a particular setting does. Confident, vague answers usually signal heavy use without real understanding.

• Do small businesses need AI governance, or is that only for large enterprises?

  Governance scales down, it doesn't disappear. A small business doesn't need the same formal framework as an FTSE 250 organisation, but it does need someone who knows what permissions are switched on, what data an AI tool can access, and what to do if something goes wrong. The scale of the governance changes. The need for it doesn't.

• What does fortysix ai's bespoke AI training involve?

  Full-day, in-person sessions built around a specific team, their tools and their actual work, rather than a generic slide deck. Programmes are tailored by sector and by role, drawing on governance-first discovery so the day addresses real risks and use cases rather than AI in the abstract. Get in touch for a scope suited to your business.

• What is AI governance, in practical terms?

  AI governance is the set of decisions a business makes about how AI tools are used: what data they can access, who can approve autonomous actions, what gets logged, and what happens if something goes wrong. It's the difference between a policy document telling staff what they may type, and actual control over what the software is allowed to do on its own.

• What is AEO (answer engine optimisation) and why does it matter for a business?

  AEO is the practice of structuring a website's content so AI tools like ChatGPT, Claude and Gemini can find, understand and cite it when someone asks a relevant question. It matters because over half of B2B buyers now start their research in an AI chatbot rather than a search engine. If a business isn't structured to be found there, it's effectively invisible to a growing share of its own prospective customers before they ever reach a website.

• How much does AI training cost for a business?

  Cost depends on team size, sector and scope, since programmes are built around a specific business rather than sold as a fixed package. The ongoing retainer service starts from £995 a month. Reach out for a quote tailored to your team.